full_info,tags
"real-time visual place recognition personal localization mobile device. paper presents approach indoor personal localization mobile device based visual place recognition. implemented smartphone two state-of-the-art algorithms representative two different approaches visual place recognition: fab-map recognizes places using individual images, able-m utilizes sequences images. algorithms evaluated environments different structure, focusing problems commonly encountered mobile device camera used. conclusions drawn evaluation guidelines design fastable system, based able-m algorithm, introduces major modifications concept image matching. improvements radically cut processing time improve scalability, making possible localize user long image sequences limited computing power mobile device. resulting place recognition system compares favorably able-m fab-map solutions context real-time personal localization.",3
"local optimality generalization guarantees langevin algorithm via empirical metastability. study detailed path-wise behavior discrete-time langevin algorithm non-convex empirical risk minimization (erm) lens metastability, adopting techniques berglund gentz. particular local optimum empirical risk, arbitrary initialization, show that, high probability, one two mutually exclusive events occur: either langevin trajectory ends somewhere outside $\varepsilon$-neighborhood particular optimum within short recurrence time; enters $\varepsilon$-neighborhood recurrence time stays exponentially long escape time. call phenomenon empirical metastability. two-timescale characterization aligns nicely existing literature following two senses. first, recurrence time dimension-independent, resembles convergence time deterministic gradient descent (gd). however unlike gd, langevin algorithm require strong conditions local initialization, possibility eventually visiting optima. second, scaling escape time consistent eyring-kramers law, states langevin scheme eventually visit local minima, take exponentially long time transit among them. apply path-wise concentration result context statistical learning examine local notions generalization optimality.",3
"disunited nations? multiplex network approach detecting preference affinity blocs using texts votes. paper contributes emerging literature models votes text tandem better understand polarization expressed preferences. introduces new approach estimate preference polarization multidimensional settings, international relations, based developments natural language processing network science literatures -- namely word embeddings, retain valuable syntactical qualities human language, community detection multilayer networks, locates densely connected actors across multiple, complex networks. find employment tools tandem helps better estimate states' foreign policy preferences expressed un votes speeches beyond permitted votes alone. utility located affinity blocs demonstrated application conflict onset international relations, though tools interest scholars faced measurement preferences polarization multidimensional settings.",3
"direct approach multi-class boosting extensions. boosting methods combine set moderately accurate weaklearners form highly accurate predictor. despite practical importance multi-class boosting, received far less attention binary counterpart. work, propose fully-corrective multi-class boosting formulation directly solves multi-class problem without dividing multiple binary classification problems. contrast, previous multi-class boosting algorithms decompose multi-boost problem multiple binary boosting problems. explicitly deriving lagrange dual primal optimization problem, able construct column generation-based fully-corrective approach boosting directly optimizes multi-class classification performance. new approach updates weak learners' coefficients every iteration, manner flexible enough accommodate various loss functions regularizations. example, enables us introduce structural sparsity mixed-norm regularization promote group sparsity feature sharing. boosting shared features particularly beneficial complex prediction problems features expensive compute. experiments various data sets demonstrate direct multi-class boosting generalizes well as, better than, range competing multi-class boosting methods. end result highly effective compact ensemble classifier trained distributed fashion.",3
"graph-based compression dynamic 3d point cloud sequences. paper addresses problem compression 3d point cloud sequences characterized moving 3d positions color attributes. temporally successive point cloud frames similar, motion estimation key effective compression sequences. however remains challenging problem point cloud frames varying numbers points without explicit correspondence information. represent time-varying geometry sequences set graphs, consider 3d positions color attributes points clouds signals vertices graphs. cast motion estimation feature matching problem successive graphs. motion estimated sparse set representative vertices using new spectral graph wavelet descriptors. dense motion field eventually interpolated solving graph-based regularization problem. estimated motion finally used removing temporal redundancy predictive coding 3d positions color characteristics point cloud sequences. experimental results demonstrate method able accurately estimate motion consecutive frames. moreover, motion estimation shown bring significant improvement terms overall compression performance sequence. best knowledge, first paper exploits spatial correlation inside frame (through graph) temporal correlation frames (through motion estimation) compress color geometry 3d point cloud sequences efficient way.",3
"towards end-to-end learning dialog state tracking management using deep reinforcement learning. paper presents end-to-end framework task-oriented dialog systems using variant deep recurrent q-networks (drqn). model able interface relational database jointly learn policies language understanding dialog strategy. moreover, propose hybrid algorithm combines strength reinforcement learning supervised learning achieve faster learning speed. evaluated proposed model 20 question game conversational game simulator. results show proposed method outperforms modular-based baseline learns distributed representation latent dialog state.",3
"information-theoretic optimality principle deep reinforcement learning. methodologically address problem q-value overestimation deep reinforcement learning handle high-dimensional state spaces efficiently. adapting concepts information theory, introduce intrinsic penalty signal encouraging reduced q-value estimates. resultant algorithm encompasses wide range learning outcomes containing deep q-networks special case. different learning outcomes demonstrated tuning lagrange multiplier accordingly. furthermore propose novel scheduling scheme lagrange multiplier ensure efficient robust learning. experiments atari games, algorithm outperforms algorithms (e.g. deep double deep q-networks) terms game-play performance sample complexity.",3
"taking primitive optimality theory beyond finite state. primitive optimality theory (otp) (eisner, 1997a; albro, 1998), computational model optimality theory (prince smolensky, 1993), employs finite state machine represent set active candidates stage optimality theoretic derivation, well weighted finite state machines represent constraints themselves. purposes, however, would convenient set candidates limited set criteria capable described higher-level grammar formalism, context free grammar, context sensitive grammar, multiple context free grammar (seki et al., 1991). examples include reduplication phrasal stress models. introduce mechanism otp-like optimality theory constraints remain weighted finite state machines, sets candidates represented higher-level grammars. particular, use multiple context-free grammars model reduplication manner correspondence theory (mccarthy prince, 1995), develop extended version earley algorithm (earley, 1970) apply constraints reduplicating candidate set.",3
"cahn--hilliard inpainting double obstacle potential. inpainting damaged images wide range applications many different mathematical methods proposed solve problem. inpainting witht help cahn--hilliard models particularly successful, turns cahn--hilliard inpainting double obstacle potential lead better results compared inpainting smooth double well potential. however, mathematical analysis approach missing far. paper give first analytical results cahn--hilliard double obstacle model particular show existence stationary solutions without constraints parameters involved. help numerical results show effectiveness approach binary grayscale images.",10
"characterization combined effects overlap imbalance svm classifier. paper demonstrate two common problems machine learning---imbalanced overlapping data distributions---do independent effects performance svm classifiers. result notable since shows model either factors must account presence other. study relationship problems lead discovery previously unreported form ""covert"" overfitting resilient commonly used empirical regularization techniques. demonstrate existance covert phenomenon several methods based around parametric regularization trained svms. findings area suggest possible approach quantifying overlap real world data sets.",3
"cooperative multi-agent reinforcement learning low-level wireless communication. traditional radio systems strictly co-designed lower levels osi stack compatibility efficiency. although enabled success radio communications, also introduced lengthy standardization processes imposed static allocation radio spectrum. various initiatives undertaken research community tackle problem artificial spectrum scarcity making frequency allocation dynamic building flexible radios replace static ones. reason believe computer vision control overhauled introduction machine learning, wireless communication also improved utilizing similar techniques increase flexibility wireless networks. work, pose problem discovering low-level wireless communication schemes ex-nihilo two agents fully decentralized fashion reinforcement learning problem. proposed approach uses policy gradients learn optimal bi-directional communication scheme shows surprisingly sophisticated intelligent learning behavior. present results extensive experiments analysis fidelity approach.",5
noisy expectation-maximization: applications generalizations. present noise-injected version expectation-maximization (em) algorithm: noisy expectation maximization (nem) algorithm. nem algorithm uses noise speed convergence em algorithm. nem theorem shows injected noise speeds average convergence em algorithm local maximum likelihood surface positivity condition holds. generalized form noisy expectation-maximization (nem) algorithm allow arbitrary modes noise injection including adding multiplying noise data. demonstrate noise benefits em algorithms gaussian mixture model (gmm) additive multiplicative nem noise injection. separate theorem (not presented here) shows noise benefit independent identically distributed additive noise decreases sample size mixture models. theorem implies noise benefit pronounced data sparse. injecting blind noise slowed convergence.,18
"bayesian nonparametric feature policy learning decision-making. learning demonstrations gained increasing interest recent past, enabling agent learn make decisions observing experienced teacher. many approaches proposed solve problem, little work focuses reasoning observed behavior. assume that, many practical problems, agent makes decision based latent features, indicating certain action. therefore, propose generative model states actions. inference reveals number features, features, policies, allowing us learn analyze underlying structure observed behavior. further, approach enables prediction actions new states. simulations used assess performance algorithm based upon model. moreover, problem learning driver's behavior investigated, demonstrating performance proposed model real-world scenario.",3
"trainable referring expression generation using overspecification preferences. referring expression generation (reg) models use speaker-dependent information require considerable amount training data produced every individual speaker, may otherwise perform poorly. work present simple reg experiment allows use larger training data sets grouping speakers according overspecification preferences. intrinsic evaluation shows method generally outperforms personalised method found previous work.",3
"fast scalable joint estimator learning multiple related sparse gaussian graphical models. estimating multiple sparse gaussian graphical models (sggms) jointly many related tasks (large $k$) high-dimensional (large $p$) situation important task. previous studies joint estimation multiple sggms rely penalized log-likelihood estimators involve expensive difficult non-smooth optimizations. propose novel approach, fasjem \underline{fa}st \underline{s}calable \underline{j}oint structure-\underline{e}stimation \underline{m}ultiple sggms large scale. first study joint sggm using elementary estimator framework, work three major contributions: (1) solve fasjem entry-wise manner parallelizable. (2) choose proximal algorithm optimize fasjem. improves computational efficiency $o(kp^3)$ $o(kp^2)$ reduces memory requirement $o(kp^2)$ $o(k)$. (3) theoretically prove fasjem achieves consistent estimation convergence rate $o(\log(kp)/n_{tot})$. several synthetic four real-world datasets, fasjem shows significant improvements baselines accuracy, computational complexity, memory costs.",18
"invertible orientation scores 3d images. enhancement detection elongated structures noisy image data relevant many biomedical applications. handle complex crossing structures 2d images, 2d orientation scores introduced, already showed use variety applications. extend work 3d orientation scores. first, construct orientation score given dataset, achieved invertible coherent state type transform. transformation introduce 3d versions 2d cake-wavelets, complex wavelets simultaneously detect oriented structures oriented edges. efficient implementation different steps wavelet creation use spherical harmonic transform. finally, show first results practical applications 3d orientation scores.",10
"learnability pauli noise. recently, several noise benchmarking algorithms developed characterize noisy quantum gates today's quantum devices. well-known issue benchmarking everything quantum noise learnable due existence gauge freedom, leaving open question information noise learnable not, unclear even single cnot gate. give precise characterization learnability pauli noise channels attached clifford gates, showing learnable information corresponds cycle space pattern transfer graph gate set, unlearnable information corresponds cut space. implies optimality cycle benchmarking, sense learn learnable information pauli noise. experimentally demonstrate noise characterization ibm's cnot gate 2 unlearnable degrees freedom, obtain bounds using physical constraints. addition, give attempt characterize unlearnable information assuming perfect initial state preparation. however, based experimental data, conclude assumption inaccurate yields unphysical estimates, obtain lower bound state preparation noise.",17
"fast robust methods singular state-space models. state-space models used wide range time series analysis formulations. kalman filtering smoothing work-horse algorithms settings. classic algorithms assume gaussian errors simplify estimation, recent advances use broader range optimization formulations allow outlier-robust estimation, well constraints capture prior information. develop methods state-space models either innovations error covariances may singular. models frequently arise navigation (e.g. `colored noise' models deterministic integrals) ubiquitous auto-correlated time series models arma. reformulate state-space models (singular well nonsinguar) constrained convex optimization problems, develop efficient algorithm reformulation. convergence rate {\it locally linear}, constants depend conditioning problem. numerical comparisons show new approach outperforms competing approaches {\it nonsingular} models, including state art interior point (ip) methods. ip methods converge superlinear rates; expect dominate. however, steep rate proposed approach (independent problem conditioning) combined cheap iterations wins ip run-time comparison. therefore suggest proposed approach {\it default choice} estimating state space models outside gaussian context, regardless whether error covariances singular not.",10
"material recognition cnns hierarchical planning biped robot locomotion slippery terrain. paper tackle problem visually predicting surface friction environments diverse surfaces, integrating knowledge biped robot locomotion planning. problem essential autonomous robot locomotion since diverse surfaces varying friction abound real world, wood ceramic tiles, grass ice, may cause difficulties huge energy costs robot locomotion considered. propose estimate friction uncertainty visual estimation material classes using convolutional neural networks, together probability distribution functions friction associated material. robustly integrate friction predictions hierarchical (footstep full-body) planning method using chance constraints, optimize trajectory costs levels planning method consistency. solution achieves fully autonomous perception locomotion slippery terrain, considers friction uncertainty, also collision, stability trajectory cost. show promising friction prediction results real pictures outdoor scenarios, planning experiments real robot facing surfaces different friction.",3
"computing functions random variables via reproducing kernel hilbert space representations. describe method perform functional operations probability distributions random variables. method uses reproducing kernel hilbert space representations probability distributions, applicable operations applied points drawn respective distributions. refer approach {\em kernel probabilistic programming}. illustrate synthetic data, show used nonparametric structural equation models, application causal inference.",18
"implementation back-propagation learning gf11, large simd parallel computer. current connectionist simulations require huge computational resources. describe neural network simulator ibm gf11, experimental simd machine 566 processors peak arithmetic performance 11 gigaflops. present parallel implementation backpropagation learning algorithm, techniques increasing efficiency, performance measurements nettalk text-to-speech benchmark, performance model simulator. simulator currently runs back-propagation learning algorithm 900 million connections per second, ""connection per second"" includes forward backward pass. figure obtained machine 356 processors working; 566 processors operational, simulation run one billion connections per second. conclude gf11 well-suited neural network simulation, analyze use machine determine features important high performance.",3
"lamarckism mechanism synthesis: approaching constrained optimization ideas biology. nonlinear constrained optimization problems encountered many scientific fields. utilize huge calculation power current computers, many mathematic models also rebuilt optimization problems. constrained conditions need handled. borrowing biological concepts, study accomplished dealing constraints synthesis four-bar mechanism. biologically regarding constrained condition form selection characteristics population, four new algorithms proposed, new explanation given penalty method. using algorithms, three cases tested differential-evolution based programs. better, comparable, results show presented algorithms methodology may become common means constraint handling optimization problems.",10
paddlespeech: easy-to-use all-in-one speech toolkit. paddlespeech open-source all-in-one speech toolkit. aims facilitating development research speech processing technologies providing easy-to-use command-line interface simple code structure. paper describes design philosophy core architecture paddlespeech support several essential speech-to-text text-to-speech tasks. paddlespeech achieves competitive state-of-the-art performance various speech datasets implements popular methods. also provides recipes pretrained models quickly reproduce experimental results paper. paddlespeech publicly avaiable https://github.com/paddlepaddle/paddlespeech.,5
"revisiting stochastic off-policy action-value gradients. off-policy stochastic actor-critic methods rely approximating stochastic policy gradient order derive optimal policy. one may also derive optimal policy approximating action-value gradient. use action-value gradients desirable policy improvement occurs along direction steepest ascent. studied extensively within context natural gradient actor-critic algorithms recently within context deterministic policy gradients. paper briefly discuss off-policy stochastic counterpart deterministic action-value gradients, well incremental approach following policy gradient lieu natural gradient.",18
"fault-tolerant, paradoxical path-finding physical conceptual systems. report initial investigations reliability path-finding based models propose future areas interest. inspired broken sidewalks on-campus construction projects, develop two models navigating ""unreliable network."" based concept ""accumulating risk"" backward destination, operate directed acyclic graphs probability failure associated edge. first serves introduce faults addressed second, conservative model. next, show paradox models used construct polynomials conceptual networks, design processes software development life cycles. risk network increases uniformly, reliable path changes wider longer shorter narrower. let professional inexperience--such entry level cooks software developers--represent probability edge failure, change path imply novice follow instructions fewer ""back-up"" plans, yet alternative routes followed expert?",3
"biodynamo project: platform computer simulations biological dynamics. paper brief update developments biodynamo project, new platform computer simulations biological research. discuss new capabilities simulator, important new concepts simulation methodology well numerous applications computational biology nanoscience communities.",3
"pathological oct retinal layer segmentation using branch residual u-shape networks. automatic segmentation retinal layer structures enables clinically-relevant quantification monitoring eye disorders time oct imaging. eyes late-stage diseases particularly challenging segment, shape highly warped due pathological biomarkers. context, propose novel fully convolutional neural network (cnn) architecture combines dilated residual blocks asymmetric u-shape configuration, segment multiple layers highly pathological eyes one shot. validate approach dataset late-stage amd patients demonstrate lower computational costs higher performance compared state-of-the-art methods.",3
"calibration changing checking rules application short-term trading. provide natural learning process financial trader without risk receives gain case stock market inefficient. process, trader rationally choose gambles using prediction made randomized calibrated algorithm. strategy based dawid's notion calibration general changing checking rules modification kakade foster's randomized algorithm computing calibrated forecasts.",3
"autonomous agents modelling agents: comprehensive survey open problems. much research artificial intelligence concerned development autonomous agents interact effectively agents. important aspect agents ability reason behaviours agents, constructing models make predictions various properties interest (such actions, goals, beliefs) modelled agents. variety modelling approaches exist vary widely methodology underlying assumptions, catering needs different sub-communities within developed reflecting different practical uses intended. purpose present article provide comprehensive survey salient modelling methods found literature. article concludes discussion open problems may form basis fruitful future research.",3
"graph quantization. vector quantization(vq) lossy data compression technique signal processing, restricted feature vectors therefore inapplicable combinatorial structures. contribution presents theoretical foundation graph quantization (gq) extends vq domain attributed graphs. present necessary lloyd-max conditions optimality graph quantizer consistency results optimal gq design based empirical distortion measures stochastic optimization. results statistically justify existing clustering algorithms domain graphs. proposed approach provides template link structural pattern recognition methods gq statistical pattern recognition.",3
"deep visual attention prediction. work, aim predict human eye fixation view-free scenes based end-to-end deep learning architecture. although convolutional neural networks (cnns) made substantial improvement human attention prediction, still needed improve cnn based attention models efficiently leveraging multi-scale features. visual attention network proposed capture hierarchical saliency information deep, coarse layers global saliency information shallow, fine layers local saliency response. model based skip-layer network structure, predicts human attention multiple convolutional layers various reception fields. final saliency prediction achieved via cooperation global local predictions. model learned deep supervision manner, supervision directly fed multi-level layers, instead previous approaches providing supervision output layer propagating supervision back earlier layers. model thus incorporates multi-level saliency predictions within single network, significantly decreases redundancy previous approaches learning multiple network streams different input scales. extensive experimental analysis various challenging benchmark datasets demonstrate method yields state-of-the-art performance competitive inference time.",3
"filamentary structures ionized gas cygnus x. ionized gas probes influence massive stars environment. cygnus x region (d~1.5 kpc) one massive star forming complexes galaxy, cyg ob2 association (age 3-5 myr stellar mass $2 \times 10^{4}$ m$_{\odot}$) dominant influence. observe cygnus x region 148 mhz using low frequency array (lofar) take account short-spacing information image deconvolution. together data canadian galactic plane survey, investigate morphology, distribution, physical conditions low-density ionized gas $4^{\circ} \times 4^{\circ}$ (100 pc $\times$ 100 pc) region resolution 2' (0.9 pc). galactic radio emission region analyzed almost entirely thermal (free-free) 148 mhz, emission measures $10^3 < em~{\rm[pc~cm^{-6}]} < 10^6$. filamentary structure prominent feature emission, use disperse filchap identify filamentary ridges characterize radial ($em$) profiles. distribution radial profiles characteristic width 4.3 pc power-law distribution ($\beta = -1.8 \pm 0.1$) peak $em$ completeness limit 4200 pc cm$^{-6}$. electron densities filamentary structure range $10 < n_e~{\rm[cm^{-3}]} < 400$ median value 35 cm$^{-3}$, remarkably similar [n ii] surveys ionized gas. cyg ob2 may ionize two-thirds total ionized gas ionized gas filaments. half filamentary structures likely photoevaporating surfaces flowing surrounding diffuse (~5 cm$^{-3}$) medium. however, likely case ionized gas ridges. characteristic width distribution ionized gas points stellar winds cyg ob2 creating fraction ionized filaments swept-up ionized gas dissipated turbulence.",0
"nonconvex penalties analytical solutions one-bit compressive sensing. one-bit measurements widely exist real world, used recover sparse signals. task known problem learning halfspaces learning theory one-bit compressive sensing (1bit-cs) signal processing. paper, propose novel algorithms based convex nonconvex sparsity-inducing penalties robust 1bit-cs. provide sufficient condition verify whether solution globally optimal not. show globally optimal solution positive homogeneous penalties obtained two steps: proximal operator normalization step. several nonconvex penalties, including minimax concave penalty (mcp), $\ell_0$ norm, sorted $\ell_1$ penalty, provide fast algorithms finding analytical solutions solving dual problem. specifically, algorithm $200$ times faster existing algorithm mcp. efficiency comparable algorithm $\ell_1$ penalty time, performance much better. among penalties, sorted $\ell_1$ penalty robust noise different settings.",3
"entropy-based closure probabilistic learning manifolds. recent paper, authors proposed general methodology probabilistic learning manifolds. method used generate numerical samples statistically consistent existing dataset construed realization non-gaussian random vector. manifold structure learned using diffusion manifolds statistical sample generation accomplished using projected ito stochastic differential equation. probabilistic learning approach extended polynomial chaos representation databases manifolds probabilistic nonconvex constrained optimization fixed budget function evaluations. methodology introduces isotropic-diffusion kernel hyperparameter {\epsilon}. currently, {\epsilon} less arbitrarily chosen. paper, propose selection criterion identifying optimal value {\epsilon}, based maximum entropy argument. result comprehensive, closed, probabilistic model characterizing data sets hidden constraints. entropy argument ensures possible models, one uncertain beyond specified constraints, selected. applications presented several databases.",10
"threshold solutions intercritical inhomogeneous nls. consider focusing inhomogeneous nonlinear schr\""odinger equation $h^1(\mathbb{r}^3)$, \begin{equation} i\partial_t u + \delta u + |x|^{-b}|u|^{2}u=0,{equation} $0 < b <\tfrac{1}{2}$. previous works established blowup/scattering dichotomy mass-energy threshold determined ground state solution $q$. work, study solutions exactly mass-energy threshold. addition ground state solution, prove existence solutions $q^\pm$, approach standing wave positive time direction, either blow scatter negative time direction. using particular solutions, classify possible behaviors threshold solutions. particular, solution either behaves sub-threshold case, agrees $e^{it}q$, $q^+$, $q^-$ symmetries equation.",10
"online multiple kernel learning structured prediction. despite recent progress towards efficient multiple kernel learning (mkl), structured output case remains open research front. current approaches involve repeatedly solving batch learning problem, makes inadequate large scale scenarios. propose new family online proximal algorithms mkl (as well group-lasso variants thereof), overcomes drawback. show regret, convergence, generalization bounds proposed method. experiments handwriting recognition dependency parsing testify successfulness approach.",18
"effects data size frequency range distributional semantic models. paper investigates effects data size frequency range distributional semantic models. compare performance number representative models several test settings data varying sizes, test items various frequency. results show neural network-based models underperform data small, reliable model data varying sizes frequency ranges inverted factorized model.",3
"preadapted universal switch distribution testing hilberg's conjecture. hilberg's conjecture natural language states mutual information two adjacent long blocks text grows like power block length. exponent statement upper bounded using pointwise mutual information estimate computed carefully chosen code. bound better, lower compression rate requirement code universal. improve received upper bound hilberg's exponent, paper, introduce two novel universal codes, called plain switch distribution preadapted switch distribution. generally speaking, switch distributions certain mixtures adaptive markov chains varying orders additional communication avoid called catch-up phenomenon. advantage distributions achieve low compression rate guaranteed universal. using switch distributions obtain sample text english non-markovian hilberg's exponent $\le 0.83$, improves previous bound $\le 0.94$ obtained using lempel-ziv code.",3
"machine learning meets network science: dimensionality reduction fast efficient embedding networks hyperbolic space. complex network topologies hyperbolic geometry seem specularly connected, one fascinating challenging problems recent complex network theory map given network hyperbolic space. popularity similarity optimization (pso) model represents - moment - climax theory. suggests trade-off node popularity similarity mechanism explain complex network topologies emerge - discrete samples - continuous world hyperbolic geometry. hyperbolic space seems appropriate represent real complex networks. fact, preserves many fundamental topological properties, exploited real applications as, among others, link prediction community detection. here, observe first time topological-based machine learning class algorithms - nonlinear unsupervised dimensionality reduction - directly approximate network's node angular coordinates hyperbolic model two-dimensional space, according similar topological organization named angular coalescence. basis phenomenon, propose new class algorithms offers fast accurate coalescent embedding networks hyperbolic space even graphs thousands nodes.",2
"price anarchy auctions. survey outlines general modular theory proving approximation guarantees equilibria auctions complex settings. theory complements traditional economic techniques, generally focus exact optimal solutions accordingly limited relatively stylized settings. highlight three user-friendly analytical tools: smoothness-type inequalities, immediately yield approximation guarantees many auction formats interest special case complete information deterministic strategies; extension theorems, extend guarantees randomized strategies, no-regret learning outcomes, incomplete-information settings; composition theorems, extend guarantees simpler complex auctions. combining tools yields tight worst-case approximation guarantees equilibria many widely-used auction formats.",3
"hybrid gradient boosting trees neural networks forecasting operating room data. time series data constitutes distinct growing problem machine learning. corpus time series data grows larger, deep models simultaneously learn features classify features intractable suboptimal. paper, present feature learning via long short term memory (lstm) networks prediction via gradient boosting trees (xgb). focusing consequential setting electronic health record data, predict occurrence hypoxemia five minutes future based past features. make two observations: 1) long short term memory networks effective capturing long term dependencies based single feature 2) gradient boosting trees capable tractably combining large number features including static features like height weight. observations mind, generate features performing ""supervised"" representation learning lstm networks. augmenting original xgb model features gives significantly better performance either individual method.",3
"combinatorial solution non-rigid 3d shape-to-image matching. propose combinatorial solution problem non-rigidly matching 3d shape 3d image data. end, model shape triangular mesh allow triangle mesh rigidly transformed achieve suitable matching image. penalising distance relative rotation neighbouring triangles matching compromises image shape information. paper, resolve two major challenges: firstly, address resulting large np-hard combinatorial problem suitable graph-theoretic approach. secondly, propose efficient discretisation unbounded 6-dimensional lie group se(3). knowledge first combinatorial formulation non-rigid 3d shape-to-image matching. contrast existing local (gradient descent) optimisation methods, obtain solutions require good initialisation within bound optimal solution. evaluate proposed method two problems non-rigid 3d shape-to-shape non-rigid 3d shape-to-image registration demonstrate provides promising results.",3
"semantic social network analysis. social network analysis (sna) tries understand exploit key features social networks order manage life cycle predict evolution. increasingly popular web 2.0 sites forming huge social network. classical methods social network analysis (sna) applied online networks. paper, propose leveraging semantic web technologies merge exploit best features domain. present facilitate enhance analysis online social networks, exploiting power semantic social network analysis.",3
"fully-charm -bottom pentaquarks lattice-qcd inspired quark model. fully-charm -bottom pentaquarks, \emph{i.e.} $cccc\bar{c}$ $bbbb\bar{b}$, spin-parity quantum numbers $j^p=\frac{1}{2}^-$, $\frac{3}{2}^-$ $\frac{5}{2}^-$, investigated within lattice-qcd inspired quark model, already successfully described recently announced fully-charm tetraquark candidate $x(6900)$, also predicted several fully-heavy tetraquarks. powerful computational technique, based gaussian expansion method combined complex-scaling range approach, employed predict, distinguish, bound, resonance scattering states mentioned five-body system. baryon-meson diquark-diquark-antiquark configurations, along possible color channels comprehensively considered. narrow resonances obtained spin-parity channel fully-charm -bottom systems. moreover, seems compact multiquarks whose wave-functions dominated either hidden-color baryon-meson diquark-diquark-antiquark structure, coupling them.",8
"efficient vvc intra prediction based deep feature fusion probability estimation. ever-growing multimedia traffic underscored importance effective multimedia codecs. among them, up-to-date lossy video coding standard, versatile video coding (vvc), attracting attentions video coding community. however, gain vvc achieved cost significant encoding complexity, brings need realize fast encoder comparable rate distortion (rd) performance. paper, propose optimize vvc complexity intra-frame prediction, two-stage framework deep feature fusion probability estimation. first stage, employ deep convolutional network extract spatialtemporal neighboring coding features. fuse reference features obtained different convolutional kernels determine optimal intra coding depth. second stage, employ probability-based model spatial-temporal coherence select candidate partition modes within optimal coding depth. finally, selected depths partitions executed whilst unnecessary computations excluded. experimental results standard database demonstrate superiority proposed method, especially high definition (hd) ultra-hd (uhd) video sequences.",5
"learnable frequency filters speech feature extraction speaker verification. mel-scale spectrum features used various recognition classification tasks speech signals. reason expect features optimal different tasks, including speaker verification (sv). paper describes learnable front-end feature extraction model. model comprises group filters transform fourier spectrum. model parameters define filters trained end-to-end optimized specifically task speaker verification. compared standard mel-scale filter-bank, filters' bandwidths center frequencies adjustable. experimental results show applying learnable acoustic front-end improves speaker verification performance conventional mel-scale spectrum features. analysis learned filter parameters suggests narrow-band information benefits sv system performance. proposed model achieves good balance performance computation cost. resource-constrained computation settings, model significantly outperforms cnn-based learnable front-ends. generalization ability proposed model also demonstrated different embedding extraction models datasets.",5
"fastmap algorithm shortest path computations. present new preprocessing algorithm embedding nodes given edge-weighted undirected graph euclidean space. euclidean distance two nodes space approximates length shortest path given graph. later, runtime, shortest path two nodes computed a* search using euclidean distances heuristic. preprocessing algorithm, called fastmap, inspired data mining algorithm name runs near-linear time. hence, fastmap orders magnitude faster competing approaches produce euclidean embedding using semidefinite programming. fastmap also produces admissible consistent heuristics therefore guarantees generation shortest paths. moreover, fastmap applies general undirected graphs many traditional heuristics, manhattan distance heuristic, well defined. empirically, demonstrate a* search using fastmap heuristic competitive a* search using state-of-the-art heuristics, differential heuristic.",3
"evolutionary artificial neural network based chemical reaction optimization. evolutionary algorithms (eas) popular tools design evolve artificial neural networks (anns), especially train them. methods advantages conventional backpropagation (bp) method low computational requirement searching large solution space. paper, employ chemical reaction optimization (cro), newly developed global optimization method, replace bp training neural networks. cro population-based metaheuristics mimicking transition molecules interactions chemical reaction. simulation results show cro outperforms many ea strategies commonly used train neural networks.",3
"impact gravitational lensing black hole mass function inference third-generation gravitational wave detectors. recent rapid growth black hole (bh) catalog gravitational waves (gws), allowed us study substructure black hole mass function (bhmf) beyond simplest power-law distribution. however, bh masses inferred binary bh merger events, may systematically 'brightened' 'dimmed' gravitational lensing effect. work, investigate impact gravitational lensing bhmf inference considering detection third-generation gw detector -- einstein telescope (et). focus high redshift, $z=10$, order obtain upper-limits effect. use monte carlo(mc) method simulate data adopting 3 original bhmfs un-lensed lensed scenarios, recovery parameters bhmfs mock data, compare difference results, respectively. found parameters well recovered within one standard deviation(std., 1$\sigma$), 3 bhmf models reconstructed within 68\% credible interval, suggesting lensing would change main structure drastically, even high redshifts high precision et. modest influence beyond $50m_{\odot}$, depends modeling high mass tail substructure bhmf. conclude impact lensing bhmf inference et safely ignored foreseeable future. careful handling lensing effects required focus accurate estimation high mass end bhmf high redshifts.",0
"observation same-sign ww production double parton scattering proton-proton collisions $\sqrt{s}$ = 13 tev. first observation production w$^\pm$w$^\pm$ bosons double parton scattering processes using same-sign electron-muon dimuon events proton-proton collisions reported. data sample corresponds integrated luminosity 138 fb$^{-1}$ recorded center-of-mass energy 13 tev using cms detector cern lhc. multivariate discriminants used distinguish signal process main backgrounds. binned maximum likelihood fit performed extract signal cross section. measured cross section production same-sign w bosons decaying leptonically 80.7 $\pm$ 11.2 (stat) $^{+9.5}_{-8.6}$ (syst) $\pm$ 12.1 (model) fb, whereas measured fiducial cross section 6.28 $\pm$ 0.81 (stat) $\pm$ 0.69 (syst) $\pm$ 0.37 (model) fb. observed significance signal 6.2 standard deviations background-only hypothesis.",7
calibration one-class svm mv set estimation. general approach anomaly detection novelty detection consists estimating high density regions minimum volume (mv) sets. one-class support vector machine (ocsvm) state-of-the-art algorithm estimating regions high dimensional data. yet suffers practical limitations. applied limited number samples lead poor performance even picking best hyperparameters. moreover solution ocsvm sensitive selection hyperparameters makes hard optimize unsupervised setting. present new approach estimate mv sets using ocsvm different choice parameter controlling proportion outliers. solution function ocsvm learnt training set desired probability mass obtained adjusting offset test set prevent overfitting. models learnt different train/test splits aggregated reduce variance induced random splits. approach makes possible tune hyperparameters automatically obtain nested set estimates. experimental results show approach outperforms standard ocsvm formulation suffering less curse dimensionality kernel density estimates. results actual data sets also presented.,18
"vqa-machine: learning use existing vision algorithms answer new questions. one intriguing features visual question answering (vqa) challenge unpredictability questions. extracting information required answer demands variety image operations detection counting, segmentation reconstruction. train method perform even one operations accurately {image,question,answer} tuples would challenging, aim achieve limited set training data seems ambitious best. propose instead general scalable approach exploits fact good methods achieve operations already exist, thus need trained. method thus learns exploit set external off-the-shelf algorithms achieve goal, approach something common neural turing machine. core proposed method new co-attention model. addition, proposed approach generates human-readable reasons decision, still trained end-to-end without ground truth reasons given. demonstrate effectiveness two publicly available datasets, visual genome vqa, show produces state-of-the-art results cases.",3
"majority vote diverse classifiers late fusion. past years, lot attention devoted multimedia indexing fusing multimodal informations. two kinds fusion schemes generally considered: early fusion late fusion. focus late classifier fusion, one combines scores modality decision level. tackle problem, investigate recent elegant well-founded quadratic program named mincq coming machine learning pac-bayesian theory. mincq looks weighted combination, set real-valued functions seen voters, leading lowest misclassification rate, maximizing voters' diversity. propose extension mincq tailored multimedia indexing. method based order-preserving pairwise loss adapted ranking allows us improve mean averaged precision measure taking account diversity voters want fuse. provide evidence method naturally adapted late fusion procedures confirm good behavior approach challenging pascal voc'07 benchmark.",18
"algorithm quasi-associative quasi-markovian rules combination information fusion. paper one proposes simple algorithm combining fusion rules, rules first use conjunctive rule transfer conflicting mass non-empty sets, way gain property associativity fulfill markovian requirement dynamic fusion. also, new rule, sdl-improved, presented.",3
"dimensionality-reduced subspace clustering. subspace clustering refers problem clustering unlabeled high-dimensional data points union low-dimensional linear subspaces, whose number, orientations, dimensions unknown. practice one may access dimensionality-reduced observations data only, resulting, e.g., undersampling due complexity speed constraints acquisition device mechanism. pertinently, even high-dimensional data set available often desirable first project data points lower-dimensional space perform clustering there; reduces storage requirements computational cost. purpose paper quantify impact dimensionality reduction random projection performance three subspace clustering algorithms, based principles sparse signal recovery. specifically, analyze thresholding based subspace clustering (tsc) algorithm, sparse subspace clustering (ssc) algorithm, orthogonal matching pursuit variant thereof (ssc-omp). find, three algorithms, dimensionality reduction order subspace dimensions possible without incurring significant performance degradation. moreover, results order-wise optimal sense reducing dimensionality leads fundamentally ill-posed clustering problem. findings carry noisy case illustrated analytical results tsc simulations ssc ssc-omp. extensive experiments synthetic real data complement theoretical findings.",18
"deep architecture unified aesthetic prediction. image aesthetics become important criterion visual content curation social media sites media content repositories. previous work aesthetic prediction models computer vision community focused aesthetic score prediction binary image labeling. however, raw aesthetic annotations form score histograms provide richer precise information binary labels mean scores. consequently, work focus rarely-studied problem predicting aesthetic score distributions propose novel architecture training procedure model. model achieves state-of-the-art results standard ava large-scale benchmark dataset three tasks: (i) aesthetic quality classification; (ii) aesthetic score regression; (iii) aesthetic score distribution prediction, using one model trained distribution prediction task. also introduce method modify image predicted aesthetics changes, use modification gain insight model.",3
"nearly tight bounds $\ell_1$ approximation self-bounding functions. study complexity learning approximation self-bounding functions uniform distribution boolean hypercube ${0,1}^n$. informally, function $f:{0,1}^n \rightarrow \mathbb{r}$ self-bounding every $x \in {0,1}^n$, $f(x)$ upper bounds sum $n$ marginal decreases value function $x$. self-bounding functions include well-known classes functions submodular fractionally-subadditive (xos) functions. introduced boucheron et al. context concentration measure inequalities. main result nearly tight $\ell_1$-approximation self-bounding functions low-degree juntas. specifically, self-bounding functions $\epsilon$-approximated $\ell_1$ polynomial degree $\tilde{o}(1/\epsilon)$ $2^{\tilde{o}(1/\epsilon)}$ variables. show degree junta-size optimal logarithmic terms. previous techniques considered stronger $\ell_2$ approximation proved nearly tight bounds $\theta(1/\epsilon^{2})$ degree $2^{\theta(1/\epsilon^2)}$ number variables. bounds rely analysis noise stability self-bounding functions together stronger connection noise stability $\ell_1$ approximation low-degree polynomials. technique also used get tighter bounds $\ell_1$ approximation low-degree polynomials faster learning algorithm halfspaces. results lead improved several cases almost tight bounds pac agnostic learning self-bounding functions relative uniform distribution. particular, assuming hardness learning juntas, show pac agnostic learning self-bounding functions complexity $n^{\tilde{\theta}(1/\epsilon)}$.",3
"left co-kothe generalized co-kothe rings solving old problem kothe non-commutative rings. solve classical kothes problem, concerning structure non-commutative rings property that: every left module direct sum cyclic modules. ring $r$ left (resp., right) kothe every left (resp., right) r-module direct sum cyclic $r$-modules. kothe [math. z. 39 (1934), 31-44] showed artinian principal ideal rings left kothe rings. cohen kaplansky [math.z 54 (1951), 97-101] proved commutative kothe rings artinian principal ideal rings. faith [math. ann. 164 (1966), 207-212] characterized commutative rings whose proper factor rings kothe rings. however, nakayama [proc. imp. acad. japan 16 (1940), 285-289] gave example left kothe ring principal ideal ring. kawada [sci. rep. tokyo kyoiku daigaku, sect. 7 (1962), 154-230; sect. 8 (1965),165-250] completely solved kothes problem basic finite dimensional algebras. far kothe's problem still open non-commutative setting. paper, among related results, solve kothes problem ring. also determine structure left co-kothe rings (rings whose left modules direct sums co-cyclics modules). finally, application, present several characterizations left kawada rings, generalizes well-known result ringel finite dimensional kawada algebras (a ring r called left kawada ring morita equivalent r left kothe ring).",10
"multiagent approach representation information decision support system. emergency situation, actors need assistance allowing react swiftly efficiently. prospect, present paper decision support system aims prepare actors crisis situation thanks decision-making support. global architecture system presented first part. focus part system designed represent information current situation. part composed multiagent system made factual agents. agent carries semantic feature aims represent partial part situation. agents develop thanks interactions comparing semantic features using proximity measures according specific ontologies.",3
"cortical prediction markets. investigate cortical learning perspective mechanism design. first, show discretizing standard models neurons synaptic plasticity leads rational agents maximizing simple scoring rules. second, main result scoring rules proper, implying neurons faithfully encode expected utilities synaptic weights encode high-scoring outcomes spikes. third, foundation hand, propose biologically plausible mechanism whereby neurons backpropagate incentives allows optimize usefulness rest cortex. finally, experiments show networks backpropagate incentives learn simple tasks.",3
"transductive adversarial networks (tan). transductive adversarial networks (tan) novel domain-adaptation machine learning framework designed learning conditional probability distribution unlabelled input data target domain, also access to: (1) easily obtained labelled data related source domain, may different conditional probability distribution target domain, (2) marginalised prior distribution labels target domain. tan leverages fully adversarial training procedure unique generator/encoder architecture approximates transductive combination available source- target-domain data. benefit tan allows distance source- target-domain label-vector marginal probability distributions greater 0 (i.e. different tasks across source target domains) whereas domain-adaptation algorithms require distance equal 0 (i.e. single task across source target domains). tan can, however, still handle latter case generalised approach case. another benefit tan due fully adversarial algorithm, potential accurately approximate highly complex distributions. theoretical analysis demonstrates viability tan framework.",18
"rgb-d salient object detection based discriminative cross-modal transfer learning. work, propose utilize convolutional neural networks boost performance depth-induced salient object detection capturing high-level representative features depth modality. formulate depth-induced saliency detection cnn-based cross-modal transfer problem bridge gap ""data-hungry"" nature cnns unavailability sufficient labeled training data depth modality. proposed approach, leverage auxiliary data source modality effectively training rgb saliency detection network obtain task-specific pre-understanding layers target modality. meanwhile, exploit depth-specific information pre-training modality classification network encourages modal-specific representations optimizing course. thus, could make feature representations rgb depth modalities discriminative possible. two modules pre-trained independently stitched initialize optimize eventual depth-induced saliency detection model. experiments demonstrate effectiveness proposed novel pre-training strategy well significant consistent improvements proposed approach state-of-the-art methods.",3
"unsupervised feature learning writer identification writer retrieval. deep convolutional neural networks (cnn) shown great success supervised classification tasks character classification dating. deep learning methods typically need lot annotated training data, available many scenarios. cases, traditional methods often better equivalent deep learning methods. paper, propose simple, yet effective, way learn cnn activation features unsupervised manner. therefore, train deep residual network using surrogate classes. surrogate classes created clustering training dataset, cluster index represents one surrogate class. activations penultimate cnn layer serve features subsequent classification tasks. evaluate feature representations two publicly available datasets. focus lies icdar17 competition dataset historical document writer identification (historical-wi). show activation features trained without supervision superior descriptors state-of-the-art writer identification methods. additionally, achieve comparable results case handwriting classification using icfhr16 competition dataset historical latin script types (clamm16).",3
"directed reduction algorithms decomposable graphs. recent years, intense research efforts develop efficient methods probabilistic inference probabilistic influence diagrams belief networks. many people concluded best methods based undirected graph structures, methods inherently superior based node reduction operations influence diagram. show two approaches essentially same, since explicitly implicity building operating underlying graphical structures. paper examine graphical structures show insight lead improved class directed reduction methods.",3
"scale coding bag deep features human attribute action recognition. approaches human attribute action recognition still images based image representation multi-scale local features pooled across scale single, scale-invariant encoding. bag-of-words recently popular representations based convolutional neural networks, local features computed multiple scales. however, multi-scale convolutional features pooled single scale-invariant representation. argue entirely scale-invariant image representations sub-optimal investigate approaches scale coding within bag deep features framework. approach encodes multi-scale information explicitly image encoding stage. propose two strategies encode multi-scale information explicitly final image representation. validate two scale coding techniques five datasets: willow, pascal voc 2010, pascal voc 2012, stanford-40 human attributes (hat-27). datasets, proposed scale coding approaches outperform scale-invariant method standard deep features network. further, combining scale coding approaches standard deep features leads consistent improvement state-of-the-art.",3
"role commutativity constraint propagation algorithms. constraint propagation algorithms form important part constraint programming systems. provide simple, yet general framework allows us explain several constraint propagation algorithms systematic way. framework proceed two steps. first, introduce generic iteration algorithm partial orderings prove correctness abstract setting. instantiate algorithm specific partial orderings functions obtain specific constraint propagation algorithms. particular, using notions commutativity semi-commutativity, show {\tt ac-3}, {\tt pc-2}, {\tt dac} {\tt dpc} algorithms achieving (directional) arc consistency (directional) path consistency instances single generic algorithm. work reported extends simplifies apt \citeyear{apt99b}.",3
"fact sheet semantic web. report gives overview activities topic semantic web. released technical report project ""ktweb -- connecting knowledge technologies communities"" 2003.",3
"$\mathcal{o}(n\log n)$ projection operator weighted $\ell_1$-norm regularization sum constraint. provide simple efficient algorithm projection operator weighted $\ell_1$-norm regularization subject sum constraint, together elementary proof. implementation proposed algorithm downloaded author's homepage.",3
"image denoising restoration cnn-lstm encoder decoder direct attention. image denoising always challenging task field computer vision image processing. paper, proposed encoder-decoder model direct attention, capable denoising reconstruct highly corrupted images. model consists encoder decoder, encoder convolutional neural network decoder multilayer long short-term memory network. proposed model, encoder reads image catches abstraction image vector, decoder takes vector well corrupted image reconstruct clean image. trained model mnist handwritten digit database making lower half every image black well adding noise top that. massive destruction images hard human understand content images, model retrieve image minimal error. proposed model compared convolutional encoder-decoder, model performed better generating missing part images convolutional autoencoder.",18
"structure-aware temporally coherent 3d human pose estimation. deep learning methods 3d human pose estimation rgb images require huge amount domain-specific labeled data good in-the-wild performance. however, obtaining annotated 3d pose data requires complex motion capture setup generally limited controlled settings. propose semi-supervised learning method using structure-aware loss function able utilize abundant 2d data learn 3d information. furthermore, present simple temporal network uses additional context present pose sequences improve temporally harmonize pose estimates. complete pipeline improves upon state-of-the-art 11.8% works 30 fps commodity graphics card.",3
"exponentially consistent kernel two-sample tests. given two sets independent samples unknown distributions $p$ $q$, two-sample test decides whether reject null hypothesis $p=q$. recent attention focused kernel two-sample tests test statistics easy compute, converge fast, low bias finite sample estimates. however, still lacks exact characterization asymptotic performance tests, particular, rate type-ii error probability decays zero large sample limit. work, show class kernel two-sample tests exponentially consistent polish, locally compact hausdorff space, e.g., $\mathbb r^d$. obtained exponential decay rate shown optimal among two-sample tests meeting given level constraint, independent particular kernels provided bounded continuous characteristic. key approach extended version sanov's theorem recent result identifies maximum mean discrepancy metric weak convergence probability measures.",18
"fitting simplicial complex using variation k-means. give simple effective two stage algorithm approximating point cloud $\mathcal{s}\subset\mathbb{r}^m$ simplicial complex $k$. first stage iterative fitting procedure generalizes k-means clustering, second stage involves deleting redundant simplices. form dimension reduction $\mathcal{s}$ obtained consequence.",3
"dynamics based features graph classification. numerous social, medical, engineering biological challenges framed graph-based learning tasks. here, propose new feature based approach network classification. show dynamics network useful reveal patterns organization components underlying graph process takes place. define generalized assortativities networks use generalized features across multiple time scales. features turn suitable signatures discriminating different classes networks. method evaluated empirically established network benchmarks. also introduce new dataset human brain networks (connectomes) use evaluate method. results reveal dynamics based features competitive often outperform state art accuracies.",18
"somz: photometric redshift pdfs self organizing maps random atlas. paper explore applicability unsupervised machine learning technique self organizing maps (som) estimate galaxy photometric redshift probability density functions (pdfs). technique takes spectroscopic training set, maps photometric attributes, redshifts, two dimensional surface using process competitive learning neurons compete closely resemble training data multidimensional space. key feature som retains topology input set, revealing correlations attributes easily identified. test three different 2d topological mapping: rectangular, hexagonal, spherical, using data deep2 survey. also explore different implementations boundary conditions map also introduce idea random atlas large number different maps created individual predictions aggregated produce robust photometric redshift pdf. also introduced new metric, $i$-score, efficiently incorporates different metrics, making easier compare different results (from different parameters different photometric redshift codes). find using spherical topology mapping obtain better representation underlying multidimensional topology, provides accurate results comparable other, state-of-the-art machine learning algorithms. results illustrate unsupervised approaches great potential many astronomical problems, particular computation photometric redshifts.",0
"japanese word sense disambiguation based examples synonyms. (this abstract): language japanese. printer fonts japases characters, characters figures printed correctly. dissertation bachelor's degree kyoto university(nagao lab.),march 1994.",1
"predicting face recognition performance using image quality. paper proposes data driven model predict performance face recognition system based image quality features. model relationship image quality features (e.g. pose, illumination, etc.) recognition performance measures using probability density function. address issue limited nature practical training data inherent data driven models, developed bayesian approach model distribution recognition performance measures small regions quality space. since model based solely image quality features, predict performance even actual recognition taken place. evaluate performance predictive capabilities proposed model six face recognition systems (two commercial four open source) operating three independent data sets: multipie, frgc cas-peal. results show proposed model accurately predict performance using accurate unbiased image quality assessor (iqa). furthermore, experiments highlight impact unaccounted quality space -- image quality features considered iqa -- contributing performance prediction errors.",3
"qed many-body theory worldlines: i. general formalism infrared structure. discuss reformulation qed matter gauge fields integrated explicitly, resulting many-body lorentz covariant theory 0+1 dimensional worldlines describing super-pairs spinning charges interacting lorentz forces. provides powerful, string inspired definition amplitudes loop orders. particular, one obtains general formulation wilson loops lines, exponentiated dynamical fields spin precession contributions, worldline contour averages exactly defined first quantized path integrals. discuss detail attractive features formalism high order perturbative computations. show worldline s-matrix elements, loop orders perturbation theory, constructed manifestly free soft singularities, infrared (ir) divergences captured removed endpoint photon exchanges infinity equivalent soft coherent dressings dyson s-matrix proposed faddeev kulish. discuss ir structures make connections soft theorems, abelian exponentiation ir divergences cusp anomalous dimensions. follow-up papers discuss efficient computation cusp anomalous dimensions universal features soft theorems.",9
"obstacle avoidance deep networks based intermediate perception. obstacle avoidance monocular images challenging problem robots. though multi-view structure-from-motion could build 3d maps, robust textureless environments. learning based methods exploit human demonstration predict steering command directly single image. however, method usually biased towards certain tasks demonstration scenarios also biased human understanding. paper, propose new method predict trajectory images. train system diverse nyuv2 dataset. ground truth trajectory computed designed cost functions automatically. convolutional neural network perception divided two stages: first, predict depth map surface normal rgb images, two important geometric properties related 3d obstacle representation. second, predict trajectory depth normal. results show intermediate perception increases accuracy 20% direct prediction. model generalizes well public indoor datasets also demonstrated robot flights simulation experiments.",3
"acquiring knowledge evaluation teachers performance higher education using questionnaire. paper, present step step knowledge acquisition process choosing structured method using questionnaire knowledge acquisition tool. want depict problem domain as, evaluate teachers performance higher education use expert system technology. problem acquire specific knowledge selected problem efficiently effectively human experts encode suitable computer format. acquiring knowledge human experts process expert systems development one common problems cited till yet. questionnaire sent 87 domain experts within public private universities pakistani. among 25 domain experts sent valuable opinions. domain experts highly qualified, well experienced highly responsible persons. whole questionnaire divided 15 main groups factors, divided 99 individual questions. facts analyzed give final shape questionnaire. knowledge acquisition technique may used learning tool research work.",3
"fast accurate reconstruction compressed color light field. light field photography studied thoroughly recent years. one drawbacks need multilens imaging. compensate that, compressed light field photography proposed tackle tradeoffs spatial angular resolutions. obtains using one lens, compressed version regular multi-lens system. acquisition system consists dedicated hardware followed decompression algorithm, usually suffers high computational time. work, suggest compress color channels well propose computationally efficient neural network achieves high-quality color light field reconstruction single coded image. approach outperforms existing solutions terms recovery quality computational complexity. also present neural network depth map extraction decompressed light field, trained unsupervised way without ground truth depth map.",3
"multi-wavelength study blazar 4c +01.02 long-term flaring activity 2014-2017. conducted detailed long-term spectral temporal study flat spectrum radio quasar 4c +01.02, using multi-wavelength observations fermi-lat, swift-xrt, swift-uvot. $2$-day bin $\gamma$-ray lightcurve 2014-2017 active state displays $14$ peak structures maximum integral flux $(\rm e > 100 \ mev)$ $\rm (2.5 \pm 0.2) \times 10^{-6}\ ph\ cm^{-2}\ s^{-1}$ mjd 57579.1, approximately $61$ times higher base flux $\rm (4.1 \pm 0.3) \times 10^{-8}\ ph\ cm^{-2}\ s^{-1}$, calculated averaging flux points source quiescent state. shortest $\gamma$-ray variability $0.66 \pm 0.08$ days observed source. correlation study $\gamma$-ray spectral index flux suggests source deviates usual trend harder brighter feature shown blazars. understand likely physical scenario responsible flux variation, performed detailed broadband spectral analysis source selecting different flux states multi-wavelength lightcurve. single zone leptonic model able reproduce broadband spectral energy distribution (sed) state. parameters model flux state determined using $\chi^2$ fit. observed synchrotron, synchrotron-self-compton (ssc), external-compton (ec) processes produce broadband sed varied flux states. adjoining contribution seed photons broad-line region (blr) ir torus ec process required provide adequate fits gev spectrum chosen states.",0
"learning 2d gabor filters infinite kernel learning regression. gabor functions wide-spread applications image processing computer vision. paper, prove 2d gabor functions translation-invariant positive-definite kernels propose novel formulation problem image representation gabor functions based infinite kernel learning regression. using formulation, obtain support vector expansion image based mixture gabor functions. problem representation gabor functions present support vector pixels. applying lasso support vector expansion, obtain sparse representation gabor function positioned small set pixels. application, introduce method learning dataset-specific set gabor filters used subsequently feature extraction. experiments show use learned gabor filters improves recognition accuracy recently introduced face recognition algorithm.",3
"evolutionary design photometric systems application gaia. designing photometric system best fulfil set scientific goals complex task, demanding compromise conflicting requirements subject various constraints. specific example determination stellar astrophysical parameters (aps) - effective temperature, metallicity etc. - across wide range stellar types. present novel approach problem makes minimal assumptions required filter system. considering filter system set free parameters may designed optimizing figure-of-merit (fom) respect parameters. example considered, fom measure well filter system `separate' stars different aps. separation vectorial nature, sense local directions ap variance preferably mutually orthogonal avoid ap degeneracy. optimization carried evolutionary algorithm, uses principles evolutionary biology search parameter space. model, hfd (heuristic filter design), applied design photometric systems gaia space astrometry mission. optimized systems show number interesting features, least persistence broad, overlapping filters. hfd systems perform least well proposed systems gaia, although inadequacies remain all. principles underlying hfd quite generic may applied filter design numerous projects, search specific types objects photometric redshift determination.",0
"practical reasoning expressive description logics. description logics (dls) family knowledge representation formalisms mainly characterised constructors build complex concepts roles atomic ones. expressive role constructors important many applications, computationally problematical. present algorithm decides satisfiability dl alc extended transitive inverse roles, role hierarchies, qualifying number restrictions. early experiments indicate algorithm well-suited implementation. additionally, show alc extended transitive inverse roles still pspace. finally, investigate limits decidability family dls.",3
"automatic labeling molecular biomarkers whole slide immunohistochemistry images using fully convolutional networks. paper addresses problem quantifying biomarkers multi-stained tissues, based color spatial information. deep learning based method automatically localize quantify cells expressing biomarker(s) whole slide image proposed. deep learning network fully convolutional network (fcn) whose input true rgb color image tissue output map different biomarkers. fcn relies convolutional neural network (cnn) classifies cell separately according biomarker expresses. study, images immunohistochemistry (ihc) stained slides collected used. 4,500 rgb images cells manually labeled based expressing biomarkers. labeled cell images used train cnn (obtaining accuracy 92% test set). trained cnn extended fcn generates map biomarkers whole slide image acquired scanner (instead classifying every cell image). evaluate method, manually labeled nuclei expressing different biomarkers two whole slide images used theses ground truth. proposed method immunohistochemical analysis compares well manual labeling humans (average f-score 0.96).",15
"collective prediction individual mobility traces exponential weights. present test sequential learning algorithm short-term prediction human mobility. novel approach pairs exponential weights forecaster large ensemble experts. experts individual sequence prediction algorithms constructed mobility traces 10 million roaming mobile phone users european country. average prediction accuracy significantly higher individual sequence prediction algorithms, namely constant order markov models derived user's data, shown achieve high accuracy previous studies human mobility prediction. algorithm uses time stamped location data, accuracy depends completeness expert ensemble, contain redundant records typical mobility patterns. proposed algorithm applicable prediction sufficiently large dataset sequences.",14
"data-driven approach semantic role labeling induced grammar structures language. semantic roles play important role extracting knowledge text. current unsupervised approaches utilize features grammar structures, induce semantic roles. dependence grammars, however, makes difficult adapt noisy new languages. paper develop data-driven approach identifying semantic roles, approach entirely unsupervised point rules need learned identify position semantic role occurs. specifically develop modified-adios algorithm based adios solan et al. (2005) learn grammar structures, use grammar structures learn rules identifying semantic roles based context grammar structures appeared. results obtained comparable current state-of-art models inherently dependent human annotated data.",3
"deep convolutional neural networks microscopy-based point care diagnostics. point care diagnostics using microscopy computer vision methods applied number practical problems, particularly relevant low-income, high disease burden areas. however, subject limitations sensitivity specificity computer vision methods used. general, deep learning recently revolutionised field computer vision, cases surpassing human performance object recognition tasks. paper, evaluate performance deep convolutional neural networks three different microscopy tasks: diagnosis malaria thick blood smears, tuberculosis sputum samples, intestinal parasite eggs stool samples. cases accuracy high substantially better alternative approach representative traditional medical imaging techniques.",3
"concurrent learning based adaptive control euler lagrange systems guaranteed parameter convergence. work presents solution adaptive tracking control euler lagrange systems guaranteed tracking parameter estimation error convergence. specifically concurrent learning based update rule fused filtered version desired system dynamics conjunction desired state based regression matrix utilized ensure position tracking error parameter estimation error terms converge origin exponentially. regression matrix used proposed controller makes use desired versions system states, initial, sufficiently exciting memory stack formed knowledge desired system trajectory priori, thus removing initial excitation condition required previously proposed concurrent learning based controllers literature. output feedback versions proposed method position measurements available controller design, (for gradient composite type adaptions) also presented order illustrate modularity proposed method. stability boundedness closed loop signals proposed controllers ensured via lyapunov based analysis. %trajectory tracking control class fully actuated euler lagrange systems considered work. system dynamics considered subject parametric uncertainties on--line identification uncertain model parameters also aimed. compared relevant past research, via novel approach, desired states proposed used forming regression matrix desired compensation based concurrent learning type adaptive update rule designed. via utilizing novel lyapunov analysis, semi--global exponential convergence tracking parameter identification error origin ensured.",5
"automatic method finding topic boundaries. article outlines new method locating discourse boundaries based lexical cohesion graphical technique called dotplotting. application dotplotting discourse segmentation performed either manually, examining graph, automatically, using optimization algorithm. results two experiments involving automatically locating boundaries series concatenated documents presented. areas application future directions work also outlined.",1
"benefits output sparsity multi-label classification. multi-label classification framework, observation associated set labels, generated tremendous amount attention recent years. modern multi-label problems typically large-scale terms number observations, features labels, amount labels even comparable amount observations. context, different remedies proposed overcome curse dimensionality. work, aim exploiting output sparsity introducing new loss, called sparse weighted hamming loss. proposed loss seen weighted version classical ones, active inactive labels weighted separately. leveraging influence sparsity loss function, provide improved generalization bounds empirical risk minimizer, suitable property large-scale problems. new loss, derive rates convergence linear underlying output-sparsity rather linear number labels. practice, minimizing associated risk performed efficiently using convex surrogates modern convex optimization algorithms. provide experiments various real-world datasets demonstrating pertinence approach compared non-weighted techniques.",10
"approximate inference amortised mcmc. propose novel approximate inference algorithm approximates target distribution amortising dynamics user-selected mcmc sampler. idea initialise mcmc using samples approximation network, apply mcmc operator improve samples, finally use samples update approximation network thereby improving quality. provides new generic framework approximate inference, allowing us deploy highly complex, implicitly defined approximation families intractable densities, including approximations produced warping source randomness deep neural network. experiments consider image modelling deep generative models challenging test method. deep models trained using amortised mcmc shown generate realistic looking samples well producing diverse imputations images regions missing pixels.",18
"quantum thermodynamic uncertainties nonequilibrium systems robertson-schrödinger relations. thermodynamic uncertainty principles make one rare anchors largely uncharted waters nonequilibrium systems, fluctuation theorems familiar. work aim trace uncertainties thermodynamic quantities nonequilibrium systems quantum origins, namely, quantum uncertainty principles. results enable us make categorical statement: gaussian systems, thermodynamic functions functionals robertson-schrodinger uncertainty function, always non-negative quantum systems, necessarily classical systems. here, quantum refers noncommutativity canonical operator pairs. nonequilibrium free energy[1], succeeded deriving several inequalities certain thermodynamic quantities. assume forms conventional thermodynamics, nonequilibrium nature hold times strong coupling. addition show fluctuation-dissipation inequality exists times nonequilibrium dynamics system. nonequilibrium systems relax equilibrium state late times, fluctuation-dissipation inequality leads robertson-schrodinger uncertainty principle help cauchy-schwarz inequality. work provides microscopic quantum basis certain important thermodynamic properties macroscopic nonequilibrium systems.",2
"accelerating deep learning memcomputing. restricted boltzmann machines (rbms) extensions, called 'deep-belief networks', powerful neural networks found applications fields machine learning big data. standard way training models resorts iterative unsupervised procedure based gibbs sampling, called 'contrastive divergence' (cd), additional supervised tuning via back-propagation. however, procedure shown follow gradient lead suboptimal solutions. paper, show efficient alternative cd means simulations digital memcomputing machines (dmms). test approach pattern recognition using modified version mnist data set. dmms sample effectively vast phase space given model distribution rbm, provide good approximation close optimum. efficient search significantly reduces number pretraining iterations necessary achieve given level accuracy, well total performance gain cd. fact, acceleration pretraining achieved simulating dmms comparable to, number iterations, recently reported hardware application quantum annealing method network data set. notably, however, dmms perform far better reported quantum annealing results terms quality training. also compare method advances supervised training, like batch-normalization rectifiers, work reduce advantage pretraining. find memcomputing method still maintains quality advantage ($>1\%$ accuracy, $20\%$ reduction error rate) approaches. furthermore, method agnostic connectivity network. therefore, extended train full boltzmann machines, even deep networks once.",3
"contextual regression: accurate conveniently interpretable nonlinear model mining discovery scientific data. machine learning algorithms linear regression, svm neural network played increasingly important role process scientific discovery. however, none interpretable accurate nonlinear datasets. present contextual regression, method joins two desirable properties together using hybrid architecture neural network embedding dot product layer. demonstrate high prediction accuracy sensitivity task predictive feature selection simulated dataset application predicting open chromatin sites human genome. simulated data, method achieved high fidelity recovery feature contributions random noise levels 200%. open chromatin dataset, application method outperformed state art method terms accuracy, also unveiled two previously unfound open chromatin related histone marks. method fill blank accurate interpretable nonlinear modeling scientific data mining tasks.",15
"retrodictive quantum computing. quantum models computation widely believed powerful classical ones. efforts center proving that, given problem, quantum algorithms resource efficient classical one. this, however, assumes standard predictive paradigm reasoning where, given initial conditions, future holds answer. bringing information future present exploit one's advantage? radical new approach reasoning, so-called retrodictive computation, benefits specific form computed functions. demonstrate use tools symbolic computation realize retrodictive quantum computing scale exploit efficiently, classically, solve instances quantum deutsch-jozsa, bernstein-vazirani, simon, grover, shor's algorithms.",17
"simulator: understanding adaptive sampling moderate-confidence regime. propose novel technique analyzing adaptive sampling called {\em simulator}. approach differs existing methods considering much information could gathered fixed sampling strategy, difficult distinguish good sampling strategy bad one given limited amount data collected given time. change perspective allows us match strength fano change-of-measure techniques, without succumbing limitations either method. concreteness, apply techniques structured multi-arm bandit problem fixed-confidence pure exploration setting, show constraints means imply substantial gap moderate-confidence sample complexity, asymptotic sample complexity $\delta \to 0$ found literature. also prove first instance-based lower bounds top-k problem incorporate appropriate log-factors. moreover, lower bounds zero-in number times \emph{individual} arm needs pulled, uncovering new phenomena drowned aggregate sample complexity. new analysis inspires simple near-optimal algorithm best-arm top-k identification, first {\em practical} algorithm kind latter problem removes extraneous log factors, outperforms state-of-the-art experiments.",3
"spectral community detection heterogeneous large networks. article, study spectral methods community detection based $ \alpha$-parametrized normalized modularity matrix hereafter called $ {\bf l}_\alpha $ heterogeneous graph models. show, regime community detection asymptotically trivial, $ {\bf l}_\alpha $ well approximated tractable random matrix falls family spiked random matrices. analysis equivalent spiked random matrix allows us improve spectral methods community detection assess performances regime study. particular, prove existence optimal value $ \alpha_{\rm opt} $ parameter $ \alpha $ detection communities best ensured provide on-line estimation $ \alpha_{\rm opt} $ based knowledge graph adjacency matrix. unlike classical spectral methods community detection clustering performed eigenvectors associated extreme eigenvalues, show theoretical analysis regularization instead performed eigenvectors prior clustering heterogeneous graphs. finally, deeper study regularized eigenvectors used clustering, assess performances new algorithm community detection. numerical simulations course article show methods outperform state-of-the-art spectral methods dense heterogeneous graphs.",18
"initialization multilayer forecasting artifical neural networks. paper, new method developed initialising artificial neural networks predicting dynamics time series. initial weighting coefficients determined neurons analogously case linear prediction filter. moreover, improve accuracy initialization method multilayer neural network, variants decomposition transformation matrix corresponding linear prediction filter suggested. efficiency proposed neural network prediction method forecasting solutions lorentz chaotic system shown paper.",3
"cgan-based manga colorization using single training image. japanese comic format known manga popular world. traditionally produced black white, colorization time consuming costly. automatic colorization methods generally rely greyscale values, present manga. furthermore, due copyright protection, colorized manga available training scarce. propose manga colorization method based conditional generative adversarial networks (cgan). unlike previous cgan approaches use many hundreds thousands training images, method requires single colorized reference image training, avoiding need large dataset. colorizing manga using cgans produce blurry results artifacts, resolution limited. therefore also propose method segmentation color-correction mitigate issues. final results sharp, clear, high resolution, stay true character's original color scheme.",3
"snake: stochastic proximal gradient algorithm regularized problems large graphs. regularized optimization problem large unstructured graph studied, regularization term tied graph geometry. typical regularization examples include total variation laplacian regularizations graph. applying proximal gradient algorithm solve problem, exist quite affordable methods implement proximity operator (backward step) special case graph simple path without loops. paper, algorithm, referred ""snake"", proposed solve regularized problems general graphs, taking benefit fast methods. algorithm consists properly selecting random simple paths graph performing proximal gradient algorithm simple paths. algorithm instance new general stochastic proximal gradient algorithm, whose convergence proven. applications trend filtering graph inpainting provided among others. numerical experiments conducted large graphs.",10
"language distribution prediction based batch markov monte carlo simulation migration. language spreading complex mechanism involves issues like culture, economics, migration, population etc. paper, propose set methods model dynamics spreading system. model randomness language spreading, propose batch markov monte carlo simulation migration(bmmcsm) algorithm, agent treated language stack. agent learns languages migrates based proposed batch markov property according transition matrix migration matrix m. since population plays crucial role language spreading, also introduce mortality fertility mechanism, controls birth death simulated agents, bmmcsm algorithm. simulation results bmmcsm show numerical geographic distribution languages varies across time. change distribution fits world cultural economic development trend. next, construct matrix t, entries directly calculated historical statistics entries unknown. thus, key success bmmcsm lies accurate estimation transition matrix estimating unknown entries supervision known entries. achieve this, first construct 20 20 5 factor tensor x characterize entry t. train random forest regressor known entries use trained regressor predict unknown entries. reason choose random forest(rf) that, compared single decision tree, conquers problem fitting shapiro test also suggests residual rf subjects normal distribution.",3
"identifying best interventions online importance sampling. motivated applications computational advertising systems biology, consider problem identifying best several possible soft interventions source node $v$ acyclic causal directed graph, maximize expected value target node $y$ (located downstream $v$). setting imposes fixed total budget sampling various interventions, along cost constraints different types interventions. pose best arm identification bandit problem $k$ arms arm soft intervention $v,$ leverage information leakage among arms provide first gap dependent error simple regret bounds problem. results significant improvement traditional best arm identification results. empirically show algorithms outperform state art flow cytometry data-set, also apply algorithm model interpretation inception-v3 deep net classifies images.",18
"symmetric tensor completion multilinear entries learning product mixtures hypercube. give algorithm completing order-$m$ symmetric low-rank tensor multilinear entries time roughly proportional number tensor entries. apply tensor completion algorithm problem learning mixtures product distributions hypercube, obtaining new algorithmic results. centers product distribution linearly independent, recover distributions many $\omega(n)$ centers polynomial time sample complexity. general case, recover distributions many $\tilde\omega(n)$ centers quasi-polynomial time, answering open problem feldman et al. (siam j. comp.) special case distributions incoherent bias vectors. main algorithmic tool iterated application low-rank matrix completion algorithm matrices adversarially missing entries.",3
"efficient pseudo-likelihood method sparse binary pairwise markov network estimation. pseudo-likelihood method one popular algorithms learning sparse binary pairwise markov networks. paper, formulate $l_1$ regularized pseudo-likelihood problem sparse multiple logistic regression problem. way, many insights optimization procedures sparse logistic regression applied learning discrete markov networks. specifically, use coordinate descent algorithm generalized linear models convex penalties, combined strong screening rules, solve pseudo-likelihood problem $l_1$ regularization. therefore substantial speedup without losing accuracy achieved. furthermore, method stable node-wise logistic regression approach unbalanced high-dimensional data penalized small regularization parameters. thorough numerical experiments simulated data real world data demonstrate advantages proposed method.",18
interpreting extracted rules ensemble trees: application computer-aided diagnosis breast mri. high predictive performance ease use interpretability important requirements applicability computer-aided diagnosis (cad) human reading studies. propose cad system specifically designed comprehensible radiologist reviewing screening breast mri studies. multiparametric imaging features combined produce cad system differentiating cancerous non-cancerous lesions. complete system uses rule-extraction algorithm present lesion classification results easy understand graph visualization.,18
"search higgs boson decay charm quark-antiquark pair proton-proton collisions $\sqrt{s}$ = 13 tev. search standard model higgs boson decaying charm quark-antiquark pair, h $\to \mathrm{c\bar{c}}$, produced association leptonically decaying v (w z) boson presented. search performed proton-proton collisions $\sqrt{s}$ = 13 tev collected cms experiment, corresponding integrated luminosity 138 fb$^{-1}$. novel charm jet identification analysis methods using machine learning techniques employed. analysis validated searching z $\to \mathrm{c\bar{c}}$ vz events, leading first observation hadron collider significance 5.7 standard deviations. observed (expected) upper limit $\sigma$(vh)$\mathcal{b}$(h$\to \mathrm{c\bar{c}}$) 0.94 (0.50$^{+0.22}_{-0.15}$) pb 95% confidence level (cl), corresponding 14 (7.6$^{+3.4}_{-2.3}$) times standard model prediction. higgs-charm yukawa coupling modifier, $\kappa_\mathrm{c}$, observed (expected) 95% cl interval 1.1 $\lt\vert\kappa_\mathrm{c}\vert \lt$ 5.5 ($\vert\kappa_\mathrm{c}\vert\lt$ 3.4), stringent constraint date.",7
"onsets frames: dual-objective piano transcription. consider problem transcribing polyphonic piano music emphasis generalizing unseen instruments. use deep neural networks propose novel approach predicts onsets frames using cnns lstms. model predicts pitch onset events uses predictions condition framewise pitch predictions. inference, restrict predictions framewise detector allowing new note start unless onset detector also agrees onset pitch present frame. focus improving onsets offsets together instead either isolation believe correlates better human musical perception. technique results 100% relative improvement note offset score maps dataset.",3
"time-varying study chinese investor sentiment, stock market liquidity volatility: based deep learning bert model tvp-var model. based commentary data shenzhen stock index bar eastmoney website january 1, 2018 december 31, 2019. paper extracts embedded investor sentiment using deep learning bert model investigates time-varying linkage investment sentiment, stock market liquidity volatility using tvp-var model. results show impact investor sentiment stock market liquidity volatility stronger. although inverse effect relatively small, pronounced state stock market. cases, response pronounced short term medium long term, impact asymmetric, shocks stronger market downward spiral.",16
free form based active contours image segmentation free space perception. paper present novel approach representing evolving deformable active contours. method combines piecewise regular b{\'e}zier models curve evolution defined local free form deformation. contour deformation locally constrained allows contour convergence almost linear complexity adapting various shape settings handling topology changes active contour. demonstrate effectiveness new active contour scheme visual free space perception segmentation using omnidirectional images acquired robot exploring unknown indoor outdoor environments. several experiments validate approach comparison state-of-the art parametric geometric active contours provide fast real-time robot free space segmentation navigation.,3
"deep gradient compression: reducing communication bandwidth distributed training. large-scale distributed training requires significant communication bandwidth gradient exchange limits scalability multi-node training, requires expensive high-bandwidth network infrastructure. situation gets even worse distributed training mobile devices (federated learning), suffers higher latency, lower throughput, intermittent poor connections. paper, find 99.9% gradient exchange distributed sgd redundant, propose deep gradient compression (dgc) greatly reduce communication bandwidth. preserve accuracy compression, dgc employs four methods: momentum correction, local gradient clipping, momentum factor masking, warm-up training. applied deep gradient compression image classification, speech recognition, language modeling multiple datasets including cifar10, imagenet, penn treebank, librispeech corpus. scenarios, deep gradient compression achieves gradient compression ratio 270x 600x without losing accuracy, cutting gradient size resnet-50 97mb 0.35mb, deepspeech 488mb 0.74mb. deep gradient compression enables large-scale distributed training inexpensive commodity 1gbps ethernet facilitates distributed training mobile.",3
"category space approach supervised dimensionality reduction. supervised dimensionality reduction emerged important theme last decade. despite plethora models formulations, lack simple model aims project set patterns space defined classes (or categories). end, set model class represented 1d subspace vector space formed features. assuming set classes exceed cardinality features, model results multi-class supervised learning features class projected class subspace. class discrimination automatically guaranteed via imposition orthogonality 1d class sub-spaces. resulting optimization problem - formulated minimization sum quadratic functions stiefel manifold - non-convex (due constraints), nevertheless structure identify reached global minimum. formulating version standard inner products, extend formulation reproducing kernel hilbert spaces straightforward manner. optimization approach also extends similar fashion kernel version. results comparisons multi-class fisher linear (and kernel) discriminants principal component analysis (linear kernel) showcase relative merits approach dimensionality reduction.",18
"parallel distributed thompson sampling large-scale accelerated exploration chemical space. chemical space large brute force searches new interesting molecules infeasible. high-throughput virtual screening via computer cluster simulations speed discovery process collecting large amounts data parallel, e.g., hundreds thousands parallel measurements. bayesian optimization (bo) produce additional acceleration sequentially identifying useful simulations experiments performed next. however, current bo methods cannot scale large numbers parallel measurements massive libraries molecules currently used high-throughput screening. here, propose scalable solution based parallel distributed implementation thompson sampling (pdts). show that, small scale problems, pdts performs similarly parallel expected improvement (ei), batch version widely used bo heuristic. additionally, settings parallel ei scale, pdts outperforms scalable baselines greedy search, $\epsilon$-greedy approaches random search method. results show pdts successful solution large-scale parallel bo.",18
"leverage, influence, jackknife clustered regression models: reliable inference using summclust. cluster-robust inference widely used modern empirical work economics many disciplines. data clustered, key unit observation cluster. propose measures ""high-leverage"" clusters ""influential"" clusters linear regression models. measures leverage partial leverage, functions them, used diagnostic tools identify datasets regression designs cluster-robust inference likely challenging. measures influence provide valuable information results depend data various clusters. also show calculate two jackknife variance matrix estimators, cv3 cv3j, byproduct computations. quantities, including jackknife variance estimators, computed new stata package called summclust summarizes cluster structure dataset.",4
"deep supervised learning using local errors. error backpropagation highly effective mechanism learning high-quality hierarchical features deep networks. updating features weights one layer, however, requires waiting propagation error signals higher layers. learning using delayed non-local errors makes hard reconcile backpropagation learning mechanisms observed biological neural networks requires neurons maintain memory input long enough higher-layer errors arrive. paper, propose alternative learning mechanism errors generated locally layer using fixed, random auxiliary classifiers. lower layers could thus trained independently higher layers training could either proceed layer layer, simultaneously layers using local error information. address biological plausibility concerns weight symmetry requirements show proposed learning mechanism based fixed, broad, random tuning neuron classification categories outperforms biologically-motivated feedback alignment learning technique mnist, cifar10, svhn datasets, approaching performance standard backpropagation. approach highlights potential biological mechanism supervised, task-dependent, learning feature hierarchies. addition, show well suited learning deep networks custom hardware drastically reduce memory traffic data communication overheads.",3
"hybrid word-character model abstractive summarization. abstractive summarization popular research topic nowadays. due difference language property, chinese summarization also gains lots attention. studies use character-based representation instead word-based keep error introduced word segmentation oov problem. however, believe word-based representation capture semantics articles accurately. proposed hybrid word-character model preserves advantage word-based character-based representations. method also enables us use larger word vocabulary size anyone else. call new method hwc (hybrid word-character). conduct experiments lcsts chinese summarization dataset, out-perform current state-of-the-art least 8 rouge points.",3
"part-of-speech tagging two sequential transducers. present method constructing using cascade consisting left- right-sequential finite-state transducer (fst), t1 t2, part-of-speech (pos) disambiguation. compared hmm, fst cascade advantage significantly higher processing speed, cost slightly lower accuracy. applications information retrieval, speed important accuracy, could benefit approach. process tagging, first assign every word unique ambiguity class c_i looked lexicon encoded sequential fst. every c_i denoted single symbol, e.g. [adj_noun], although represents set alternative tags given word occur with. sequence c_i words one sentence input fst cascade. mapped t1, left right, sequence reduced ambiguity classes r_i. every r_i denoted single symbol, although represents set alternative tags. intuitively, t1 eliminates less likely tags c_i, thus creating r_i. finally, t2 maps sequence r_i, right left, sequence single pos tags t_i. intuitively, t2 selects likely t_i every r_i. probabilities t_i, r_i, c_i used compile time, run time. (directly) occur fsts, ""implicitly contained"" structure.",3
"online learning predictable sequences. present methods online linear optimization take advantage benign (as opposed worst-case) sequences. specifically sequence encountered learner described well known ""predictable process"", algorithms presented enjoy tighter bounds compared typical worst case bounds. additionally, methods achieve usual worst-case regret bounds sequence benign. approach seen way adding prior knowledge sequence within paradigm online learning. setting shown encompass partial side information. variance path-length bounds seen particular examples online learning simple predictable sequences. extend methods results include competing set possible predictable processes (models), ""learning"" predictable process concurrently using obtain better regret guarantees. show model selection possible various assumptions available feedback. results suggest promising direction research potential applications stock market time series prediction.",18
"brundlefly semeval-2016 task 12: recurrent neural networks vs. joint inference clinical temporal information extraction. submitted two systems semeval-2016 task 12: clinical tempeval challenge, participating phase 1, identified text spans time event expressions clinical notes phase 2, predicted relation event parent document creation time. temporal entity extraction, find joint inference-based approach using structured prediction outperforms vanilla recurrent neural network incorporates word embeddings trained variety large clinical document sets. document creation time relations, find combination date canonicalization distant supervision rules predicting relations events time expressions improves classification, though gains limited, likely due small scale training data.",3
"excitation-damping quantum channels. study class quantum channels describing quantum system, split direct sum excited ground sector, undergoing one-way transfer population former latter; construction, provides generalization amplitude-damping qubit channel, regarded way upgrade trace non-increasing quantum operation, defined excited sector, possibly trace preserving operation larger hilbert space. provide necessary sufficient conditions complete positivity channels, also show complete positivity equivalent simple positivity whenever ground sector one-dimensional. finally, examine time-dependent scenario characterize cp-divisible channels markovian semigroups belonging class.",17
"proceedings twenty-eighth conference uncertainty artificial intelligence (2012). proceedings twenty-eighth conference uncertainty artificial intelligence, held catalina island, ca august 14-18 2012.",3
"neural representation sketch drawings. present sketch-rnn, recurrent neural network (rnn) able construct stroke-based drawings common objects. model trained thousands crude human-drawn images representing hundreds classes. outline framework conditional unconditional sketch generation, describe new robust training methods generating coherent sketch drawings vector format.",3
"void hubble tension? end line hubble bubble. universe may feature large-scale inhomogeneities beyond standard paradigm, implying statistical homogeneity isotropy may reached much larger scales usually assumed $\sim$100 mpc. means necessarily typical observers copernican principle could recovered super-hubble scales. here, assume validity copernican principle let cosmic microwave background, baryon acoustic oscillations, type ia supernovae, local $h_0$, cosmic chronometers, compton y-distortion kinetic sunyaev-zeldovich observations constrain geometrical degrees freedom local structure, parametrize via $\lambda$ltb model -- basically non-linear radial perturbation flrw metric. order quantify non-copernican structure could explain away hubble tension, pay careful attention computing hubble constant inhomogeneous universe, adopt model selection via bayes factor akaike information criterion. results show that, $\lambda$ltb model successfully explain away $h_0$ tension, favored respect $\lambda$cdm model one solely considers supernovae redshift range used fit hubble constant, is, $0.023<z<0.15$. one considers supernova sample, $h_0$ tension solved support $\lambda$ltb model vanishes. combined data sets, solution hubble tension barely helps. finally, reconstructed local spacetime. found data best fit shallow void $\delta_l \approx -0.04$ $r^{\mathrm{out}}_l \approx 300$ mpc, which, interestingly, lies border 95\% credible region relative standard model expectation.",0
"comprehensive performance evaluation deformable face tracking ""in-the-wild"". recently, technologies face detection, facial landmark localisation face recognition verification matured enough provide effective efficient solutions imagery captured arbitrary conditions (referred ""in-the-wild""). partially attributed fact comprehensive ""in-the-wild"" benchmarks developed face detection, landmark localisation recognition/verification. important technology thoroughly evaluated yet deformable face tracking ""in-the-wild"". now, performance mainly assessed qualitatively visually assessing result deformable face tracking technology short videos. paper, perform first, best knowledge, thorough evaluation state-of-the-art deformable face tracking pipelines using recently introduced 300vw benchmark. evaluate many different architectures focusing mainly task on-line deformable face tracking. particular, compare following general strategies: (a) generic face detection plus generic facial landmark localisation, (b) generic model free tracking plus generic facial landmark localisation, well (c) hybrid approaches using state-of-the-art face detection, model free tracking facial landmark localisation technologies. evaluation reveals future avenues research topic.",3
morse theory bestvina-brady type posets matchings. introduce morse theory posets bestvina-brady type combining matchings height functions. theory generalizes forman's discrete morse theory regular cw-complexes extends previous results morse theory $h$-regular posets finite posets. also develop relative version morse theory allows us compare topology poset given subposet.,10
"distilling knowledge using parallel data far-field speech recognition. order improve performance far-field speech recognition, paper proposes distill knowledge close-talking model far-field model using parallel data. close-talking model called teacher model. far-field model called student model. student model trained imitate output distributions teacher model. constraint realized minimizing kullback-leibler (kl) divergence output distribution student model teacher model. experimental results ami corpus show best student model achieves 4.7% absolute word error rate (wer) reduction compared conventionally-trained baseline models.",3
"border algorithms computing hasse diagrams arbitrary lattices. border algorithm ipred algorithm find hasse diagrams fca lattices. show generalized arbitrary lattices. case ipred, requires identification join-semilattice homomorphism distributive lattice.",3
"simple connectome inference partial correlation statistics calcium imaging. work, propose simple yet effective solution problem connectome inference calcium imaging data. proposed algorithm consists two steps. first, processing raw signals detect neural peak activities. second, inferring degree association neurons partial correlation statistics. paper summarises methodology led us win connectomics challenge, proposes simplified version method, finally compares results respect inference methods.",18
"automated analysis continuum fields atomistic simulations using statistical machine learning. atomistic simulations molecular dynamics/statics kind regularly used study small scale plasticity. contemporary simulations performed tens hundreds millions atoms, snapshots configurations written regular intervals analysis. continuum scale constitutive models material behavior benefit information atomic scale, particular terms deformation mechanisms, accommodation total strain partitioning stress strain fields individual grains. work develop methodology using statistical data mining machine learning algorithms automate analysis continuum field variables atomistic simulations. focus three important field variables: total strain, elastic strain microrotation. results show elastic strain individual grains exhibits unimodal log-normal distribution, whilst total strain microrotation fields evidence multimodal distribution. peaks distribution total strain identified gaussian mixture model methods circumvent overfitting problems presented. subsequently, evaluate identified peaks terms deformation mechanisms grain, e.g., helps quantify strain individual deformation mechanisms responsible. overall statistics distributions grains important input higher scale models, ultimately also helps able quantitatively discuss implications information transfer phenomenological models.",2
"emfet: e-mail features extraction tool. emfet open source flexible tool used extract large number features email corpus emails saved eml format. extracted features categorized three main groups: header features, payload (body) features, attachment features. purpose tool help practitioners researchers build datasets used training machine learning models spam detection. far, 140 features extracted using emfet. emfet extensible easy use. source code emfet publicly available github (https://github.com/wadeahijjawi/emailfeaturesextraction)",3
"toward practical n2 monte carlo: marginal particle filter. sequential monte carlo techniques useful state estimation non-linear, non-gaussian dynamic models. methods allow us approximate joint posterior distribution using sequential importance sampling. framework, dimension target distribution grows time step, thus necessary introduce resampling steps ensure estimates provided algorithm reasonable variance. many applications, interested marginal filtering distribution defined space fixed dimension. present sequential monte carlo algorithm called marginal particle filter operates directly marginal distribution, hence avoiding perform importance sampling space growing dimension. using idea, also derive improved version auxiliary particle filter. show theoretic empirical results demonstrate reduction variance conventional particle filtering, present techniques reducing cost marginal particle filter n particles o(n2) o(n logn).",18
"calcium removal cardiac ct images using deep convolutional neural network. coronary calcium causes beam hardening blooming artifacts cardiac computed tomography angiography (cta) images, lead overestimation lumen stenosis reduction diagnostic specificity. properly remove coronary calcification restore arterial lumen precisely, propose machine learning-based method multi-step inpainting process. developed new network configuration, dense-unet, achieve optimal performance low computational cost. results calcium removal process validated comparing gold-standard x-ray angiography. results demonstrated removing coronary calcification images proposed approach feasible, may potentially improve diagnostic accuracy cta.",3
"bayesian calibration forensic evidence reporting. introduce bayesian solution problem forensic speaker recognition, may little background material estimating score calibration parameters. work within bayesian paradigm evidence reporting develop principled probabilistic treatment problem, results bayesian likelihood-ratio vehicle reporting weight evidence. show contrast, reporting likelihood-ratio distribution solve problem. solution experimentally exercised simulated forensic scenario, using nist sre'12 scores, demonstrates clear advantage proposed method compared traditional plugin calibration recipe.",18
"scale large minibatch sgd: residual network training imagenet-1k improved accuracy reduced time train. past 5 years, ilsvrc competition imagenet dataset attracted lot interest computer vision community, allowing state-of-the-art accuracy grow tremendously. credited use deep artificial neural network designs. became complex, storage, bandwidth, compute requirements increased. means non-distributed approach, even using high-density server available, training process may take weeks, making prohibitive. furthermore, datasets grow, representation learning potential deep networks grows well using complex models. synchronicity triggers sharp increase computational requirements motivates us explore scaling behaviour petaflop scale supercomputers. paper describe challenges novel solutions needed order train resnet-50 large scale environment. demonstrate 90\% scaling efficiency training time 28 minutes using 104k x86 cores. supported software tools intel's ecosystem. moreover, show regular 90 - 120 epoch train runs achieve top-1 accuracy high 77\% unmodified resnet-50 topology. also introduce novel collapsed ensemble (ce) technique allows us obtain 77.5\% top-1 accuracy, similar resnet-152, training unmodified resnet-50 topology fixed training budget. resnet-50 models well scripts needed replicate posted shortly.",18
overview deep learning based methods unsupervised semi-supervised anomaly detection videos. videos represent primary source information surveillance applications available large amounts cases contain little annotation supervised learning. article reviews state-of-the-art deep learning based methods video anomaly detection categorizes based type model criteria detection. also perform simple studies understand different approaches provide criteria evaluation spatio-temporal anomaly detection.,3
"inference networks sequential monte carlo graphical models. introduce new approach amortizing inference directed graphical models learning heuristic approximations stochastic inverses, designed specifically use proposal distributions sequential monte carlo methods. describe procedure constructing learning structured neural network represents inverse factorization graphical model, resulting conditional density estimator takes input particular values observed random variables, returns approximation distribution latent variables. recognition model learned offline, independent particular dataset, prior performing inference. output networks used automatically-learned high-quality proposal distributions accelerate sequential monte carlo across diverse range problem settings.",18
"channel equalization using multilayer perceptron networks. digital communication systems, bandwidth limited channel along multipath propagation causes isi (inter symbol interference) occur. phenomenon causes distortion given transmitted symbol due transmitted symbols. help equalization isi reduced. paper presents solution isi problem performing blind equalization using ann (artificial neural networks). simulated network multilayer feedforward perceptron ann, trained utilizing error back-propagation algorithm. weights network updated accordance training network. paper presents effective method blind channel equalization, efficient pre-existing algorithms. obtained results show visible reduction noise content.",3
"perception games, image understanding interpretational geometry. interactive game theoretical approach description perception processes proposed. subject treated formally terms new class verbalizable interactive games called perception games. application previously elaborated formalism dialogues verbalizable interactive games visual perception allows combine linguistic (such formal grammars), psycholinguistic (interactive) game theoretical methods analysis image understanding human may also useful elaboration computer vision systems. way interactive game theoretical aspects interpretational geometries clarified.",10
"agent-based model origins scaling human language. background/introduction: zipf's law establishes words (large) text ordered decreasing frequency, frequency versus rank decreases power law exponent close -1. previous work stressed pattern arises conflict interests participants communication: speakers hearers. methods: challenge define computational language game population agents, playing games mainly based parameter measures relative participant's interests. results: numerical simulations suggest critical values parameter human-like vocabulary, exhibiting scaling properties, seems appear. conclusions: appearance intermediate distribution frequencies critical values parameter suggests population artificial agents emergence scaling partly arises self-organized process local interactions agents.",14
"exploiting smoothness statistical learning, sequential prediction, stochastic optimization. last several years, intimate connection convex optimization learning problems, statistical sequential frameworks, shifted focus algorithmic machine learning examine interplay. particular, one hand, intertwinement brings forward new challenges reassessment performance learning algorithms including generalization regret bounds assumptions imposed convexity analytical properties loss functions (e.g., lipschitzness, strong convexity, smoothness). hand, emergence datasets unprecedented size, demands development novel efficient optimization algorithms tackle large-scale learning problems. overarching goal thesis reassess smoothness loss functions statistical learning, sequential prediction/online learning, stochastic optimization explicate consequences. particular examine smoothness loss function could beneficial detrimental settings terms sample complexity, statistical consistency, regret analysis, convergence rate, investigate smoothness leveraged devise efficient learning algorithms.",3
"adaptive classification prediction budget. propose novel adaptive approximation approach test-time resource-constrained prediction. given input instance test-time, gating function identifies prediction model input among collection models. objective minimize overall average cost without sacrificing accuracy. learn gating prediction models fully labeled training data means bottom-up strategy. novel bottom-up method first trains high-accuracy complex model. low-complexity gating prediction model subsequently learned adaptively approximate high-accuracy model regions low-cost models capable making highly accurate predictions. pose empirical loss minimization problem cost constraints jointly train gating prediction models. number benchmark datasets method outperforms state-of-the-art achieving higher accuracy cost.",18
"combinatorial penalties: structures preserved convex relaxations?. consider homogeneous non-homogeneous convex relaxations combinatorial penalty functions defined support sets. study identifies key differences tightness resulting relaxations notion lower combinatorial envelope set-function along new necessary conditions support identification. propose general adaptive estimator convex monotone regularizers, derive new sufficient conditions support recovery asymptotic setting.",3
"over-representation extreme events decision-making: rational metacognitive account. availability bias, manifested over-representation extreme eventualities decision-making, well-known cognitive bias, generally taken evidence human irrationality. work, present first rational, metacognitive account availability bias, formally articulated marr's algorithmic level analysis. concretely, present normative, metacognitive model cognitive system over-represent extreme eventualities, depending amount time available disposal decision-making. model also accounts two well-known framing effects human decision-making risk---the fourfold pattern risk preferences outcome probability (tversky & kahneman, 1992) outcome magnitude (markovitz, 1952)---thereby providing first metacognitively-rational basis effects. empirical evidence, furthermore, confirms important prediction model. surprisingly, model unimaginably robust respect focal parameter. discuss implications work studies human decision-making, conclude presenting counterintuitive prediction model, which, confirmed, would intriguing implications human decision-making risk. knowledge, model first metacognitive, resource-rational process model cognitive biases decision-making.",15
"welldefined decision scenarios. influence diagrams serve powerful tool modelling symmetric decision problems. solving influence diagram determine set strategies decisions involved. strategy decision variable principle function past. however, past may irrelevant decision, computational reasons important deal redundant variables strategies. show current methods (e.g. ""decision bayes-ball"" algorithm shachter uai98) determine relevant past, present complete algorithm. actually, paper takes general outset: formulating decision scenario influence diagram, linear temporal ordering decisions variables required. constraint ensures decision scenario welldefined. however, structure decision scenario often yields certain decisions conditionally independent, therefore unnecessary impose linear temporal ordering decisions. paper deal partial influence diagrams i.e. influence diagrams partial temporal ordering specified. present set conditions necessary sufficient ensure partial influence diagram welldefined. conditions used basis construction algorithm determining whether partial influence diagram welldefined.",3
"reusable iterative optimization software library solve combinatorial problems approximate reasoning. real world combinatorial optimization problems scheduling typically complex solve exact methods. additionally, problems often observe vaguely specified constraints different importance, available data may uncertain, compromises antagonistic criteria may necessary. present combination approximate reasoning based constraints iterative optimization based heuristics help model solve problems framework c++ software libraries called starflip++. initially developed schedule continuous caster units steel plants, present paper results reusing library components shift scheduling system workforce industrial production plant.",3
"dynamical quantum ergodicity energy level statistics. ergodic theory provides rigorous mathematical description classical dynamical systems, including formal definition ergodic hierarchy consisting merely ergodic, weakly-, strongly-, k-mixing systems. closely related hierarchy less-known notion cyclic approximate periodic transformations [see, e.g., i. cornfield, s. fomin, y. sinai, ergodic theory (springer-verlag new york, 1982)], maps ""ergodic"" dynamical system cyclic permutation circle arguably represents elementary notion ergodicity. paper shows cyclic ergodicity generalizes quantum dynamical systems, proposed basic rigorous definition quantum ergodicity. implies ability construct orthonormal basis, quantum dynamics transports initial basis vector basis vectors one one, minimizing error overlap time-evolved initial state given basis state certain precision. proven basis, optimizing error cyclic permutations, obtained via discrete fourier transform energy eigenstates. relates quantum cyclic ergodicity level statistics. show wigner-dyson level statistics implies quantum cyclic ergodicity, reverse necessarily true. latter, study irrational flow 2d torus argue classical quantum flows cyclic ergodic, level statistics non-universal. use cyclic construction motivate quantum ergodic hierarchy operators argue additional assumption poincare recurrences, cyclic ergodicity necessary condition operators satisfy eigenstate thermalization. work provides general framework transplanting rigorous results ergodic theory quantum dynamical systems.",17
"slow stale gradients win race: error-runtime trade-offs distributed sgd. distributed stochastic gradient descent (sgd) run synchronous manner, suffers delays waiting slowest learners (stragglers). asynchronous methods alleviate stragglers, cause gradient staleness adversely affect convergence. work present first theoretical characterization speed-up offered asynchronous methods analyzing trade-off error trained model actual training runtime (wallclock time). novelty work runtime analysis considers random straggler delays, helps us design compare distributed sgd algorithms strike balance stragglers staleness. also present new convergence analysis asynchronous sgd variants without bounded exponential delay assumptions.",18
"tunably rugged landscapes known maximum minimum. propose nm landscapes new class tunably rugged benchmark problems. nm landscapes well-defined alphabets arity, including discrete real-valued alphabets, include epistasis natural transparent manner, proven known value location global maximum and, additional constraints, proven also known global minimum. empirical studies used illustrate that, coefficients selected recommended distribution, ruggedness nm landscapes smoothly tunable correlates several measures search difficulty. discuss properties make nm landscapes preferable nk landscapes walsh polynomials benchmark landscape models tunable epistasis.",3
"compressive adaptive computational ghost imaging. compressive sensing considered huge breakthrough signal acquisition. allows recording image consisting $n^2$ pixels using much fewer $n^2$ measurements transformed basis pixels take negligibly small values. standard compressive sensing techniques suffer computational overhead needed reconstruct image typical computation times hours days thus optimal applications physics spectroscopy. demonstrate adaptive compressive sampling technique performs measurements directly sparse basis. needs much fewer $n^2$ measurements without computational overhead, result available instantly.",14
"new viable mass region dark matter dirac neutrino mass generation scotogenic extension sm. propose scotogenic extension standard model provide scalar dark matter candidate new, theoretically previously unaddressed, intermediate region ($200\leq m_{dm}\leq 550$ gev) also generate light dirac neutrino masses. framework, standard model extended three gauge singlet fermions, two singlet scalar fields, one additional scalar doublet, odd $z_{2} \times z_{4}$ discrete symmetry. additional symmetries prevent singlet fermions obtaining majorana mass terms along providing stability dark matter candidate. known case scalar singlet dm model, region yet excluded narrow region close higgs resonance $m_{s} \simeq\frac{m_{h}}{2}$ - others ruled different experimental theoretical bounds. case inert doublet model, mass region ($\sim 60$-$80$ gev) high mass region (heavier $ 550 $ gev) allowed. motivates us explore parameter range intermediate-mass region $m_{w}\leq m_{dm}\leq 550 $ gev, scotogenic extension sm scalar doublet scalar singlets. dark matter model mixture singlet doublet scalars. constrain allowed parameter space model using planck bound present dark matter relic abundance, neutrino mass, latest bound spin-independent dm-nucleon scattering cross-section xenon1t experiment. model may provide viable dm candidate new, previously unexplored mass range ($200\leq m_{dm}\leq 550$ gev), new window dm candidate mass detected future experiments, along explanation dirac mass neutrinos, since far strong evidence support majorana nature neutrino mass.",8
"high-frequency tunable resistorless memcapacitor emulator application. paper, new design proposed realization high-frequency memcapacitor emulators built three otas. paper also proposes application memcapacitor amplitude modulator. furthermore, applications memcapacitor filter, oscillator point attractor, periodic doubler also shown. proposed circuits configured incremental decremental topology. proposed circuits application claim circuit much simpler design utilized topologies. performance proposed circuits verified cadence virtuoso spectre using standard cmos 180nm. furthermore, post-layout simulations comparison carried out.",5
"gaussian process regression student-t likelihood. paper considers robust efficient implementation gaussian process regression student-t observation model. challenge student-t model analytically intractable inference several approximative methods proposed. expectation propagation (ep) found accurate method many empirical studies convergence ep known problematic models containing non-log-concave site functions student-t distribution. paper illustrate situations standard ep fails converge review different modifications alternative algorithms improving convergence. demonstrate convergence problems may occur type-ii maximum posteriori (map) estimation hyperparameters show standard ep may converge map values difficult cases. present robust implementation relies primarily parallel ep updates utilizes moment-matching-based double-loop algorithm adaptively selected step size difficult cases. predictive performance ep compared laplace, variational bayes, markov chain monte carlo approximations.",18
"theory application penalized methods reproducing kernel hilbert spaces made easy. popular cubic smoothing spline estimate regression function arises minimizer penalized sum squares $\sum_j(y_j - {\mu}(t_j))^2 + {\lambda}\int_a^b [{\mu}""(t)]^2 dt$, data $t_j,y_j$, $j=1,..., n$. minimization taken infinite-dimensional function space, space functions square integrable second derivatives. calculations carried finite-dimensional space. reduction minimizing infinite dimensional space minimizing finite dimensional space occurs general objective functions: data may related function ${\mu}$ another way, sum squares may replaced suitable expression, penalty, $\int_a^b [{\mu}""(t)]^2 dt$, might take different form. paper reviews reproducing kernel hilbert space structure provides finite-dimensional solution general minimization problem. particular attention paid penalties based linear differential operators. case, one sometimes easily calculate minimizer explicitly, using green's functions.",18
"search nonresonant pair production highly energetic higgs bosons decaying bottom quarks. search nonresonant higgs boson (h) pair production via gluon vector boson (v) fusion performed four-bottom-quark final state, using proton-proton collision data 13 tev corresponding 138 fb$^{-1}$ collected cms experiment lhc. analysis targets lorentz-boosted h pairs identified using graph neural network. constrains strengths relative standard model h self-coupling quartic vvhh couplings, $\kappa_{2v}$, excluding $\kappa_{2v}$ = 0 first time, significance 6.3 standard deviations h couplings fixed standard model values.",7
"fusionnet: deep fully residual convolutional neural network image segmentation connectomics. electron microscopic connectomics ambitious research direction goal studying comprehensive brain connectivity maps using high-throughput, nano-scale microscopy. one main challenges connectomics research developing scalable image analysis algorithms require minimal user intervention. recently, deep learning drawn much attention computer vision exceptional performance image classification tasks. reason, application connectomic analyses holds great promise, well. paper, introduce novel deep neural network architecture, fusionnet, automatic segmentation neuronal structures connectomics data. fusionnet leverages latest advances machine learning, semantic segmentation residual neural networks, novel introduction summation-based skip connections allow much deeper network architecture accurate segmentation. demonstrate performance proposed method comparing state-of-the-art electron microscopy (em) segmentation methods isbi em segmentation challenge. also show segmentation results two different tasks including cell membrane cell body segmentation statistical analysis cell morphology.",3
"learning low-dimensional metrics. paper investigates theoretical foundations metric learning, focused three key questions fully addressed prior work: 1) consider learning general low-dimensional (low-rank) metrics well sparse metrics; 2) develop upper lower (minimax)bounds generalization error; 3) quantify sample complexity metric learning terms dimension feature space dimension/rank underlying metric;4) also bound accuracy learned metric relative underlying true generative metric. results involve novel mathematical approaches metric learning problem, lso shed new light special case ordinal embedding (aka non-metric multidimensional scaling).",18
"cost-sensitive tree classifiers. recently, machine learning algorithms successfully entered large-scale real-world industrial applications (e.g. search engines email spam filters). here, cpu cost test time must budgeted accounted for. paper, address challenge balancing test-time cost classifier accuracy principled fashion. test-time cost classifier often dominated computation required feature extraction-which vary drastically across eatures. decrease extraction time constructing tree classifiers, test inputs traverse along individual paths. path extracts different features optimized specific sub-partition input space. computing features inputs benefit most, cost sensitive tree classifiers match high accuracies current state-of-the-art small fraction computational cost.",18
"exponentially vanishing sub-optimal local minima multilayer neural networks. background: statistical mechanics results (dauphin et al. (2014); choromanska et al. (2015)) suggest local minima high error exponentially rare high dimensions. however, prove low error guarantees multilayer neural networks (mnns), previous works far required either heavily modified mnn model training method, strong assumptions labels (e.g., ""near"" linear separability), unrealistic hidden layer $\omega\left(n\right)$ units. results: examine mnn one hidden layer piecewise linear units, single output, quadratic loss. prove that, high probability limit $n\rightarrow\infty$ datapoints, volume differentiable regions empiric loss containing sub-optimal differentiable local minima exponentially vanishing comparison volume global minima, given standard normal input dimension $d_{0}=\tilde{\omega}\left(\sqrt{n}\right)$, realistic number $d_{1}=\tilde{\omega}\left(n/d_{0}\right)$ hidden units. demonstrate results numerically: example, $0\%$ binary classification training error cifar $n/d_{0}\approx 16$ hidden neurons.",18
"bayesian deep convolutional encoder-decoder networks surrogate modeling uncertainty quantification. interested development surrogate models uncertainty quantification propagation problems governed stochastic pdes using deep convolutional encoder-decoder network similar fashion approaches considered deep learning image-to-image regression tasks. since normal neural networks data intensive cannot provide predictive uncertainty, propose bayesian approach convolutional neural nets. recently introduced variational gradient descent algorithm based stein's method scaled deep convolutional networks perform approximate bayesian inference millions uncertain network parameters. approach achieves state art performance terms predictive accuracy uncertainty quantification comparison approaches bayesian neural networks well techniques include gaussian processes ensemble methods even training data size relatively small. evaluate performance approach, consider standard uncertainty quantification benchmark problems including flow heterogeneous media defined terms limited data-driven permeability realizations. performance surrogate model developed good even though underlying structure shared input (permeability) output (flow/pressure) fields often case image-to-image regression models used computer vision problems. studies performed underlying stochastic input dimensionality $4,225$ uncertainty quantification methods fail. uncertainty propagation tasks considered predictive output bayesian statistics compared obtained monte carlo estimates.",14
"model-based catheter segmentation mri-images. accurate reliable segmentation catheters mr-gui- ded interventions remains challenge, step critical importance clinical workflows. work, reasonable assumptions, me- chanical model based heuristics guide segmentation process allows correct catheter identification rates greater 98% (error 2.88 mm), reduction outliers one-fourth compared state art. given distal tips, searching towards proximal ends catheters guided mechanical models estimated per-catheter basis. bending characteristics used constrain image fea- ture based candidate points. final catheter trajectories hybrid sequences individual points, derived model image fea- tures. evaluate method database 10 patient mri scans including 101 manually segmented catheters. mean errors 1.40 mm median errors 1.05 mm. number outliers devi- ating 2 mm gold standard 7, number outliers deviating 3 mm gold standard 2.",3
"adaptive down-sampling dimension reduction time elastic kernel machines efficient recognition isolated gestures. scope gestural action recognition, size feature vector representing movements general quite large especially full body movements considered. furthermore, feature vector evolves movement performance complete movement fully represented matrix size dxt , whose element i, j represents value feature timestamps j. many studies addressed dimensionality reduction considering size feature vector lying r reduce variability gestural sequences expressed reduced space, computational complexity processing. return, methods explicitly addressed dimensionality reduction along time axis. yet major issue considering use elastic distances characterized quadratic complexity along time axis. present paper evaluation straightforward approaches aiming reducing dimensionality matrix movement, leading consider dimensionality reduction feature vector well reduction along time axis. dimensionality reduction feature vector achieved selecting remarkable joints skeleton performing movement, basically extremities articulatory chains composing skeleton. temporal dimen-sionality reduction achieved using either regular adaptive down-sampling seeks minimize reconstruction error movements. elastic euclidean kernels compared support vector machine learning. two data sets 1 widely referenced domain human gesture recognition, quite distinctive terms quality motion capture, used experimental assessment proposed approaches. data sets experimentally show feasible, possibly desirable, significantly reduce simultaneously size feature vector number skeleton frames represent body movements maintaining good recognition rate. method proves give satisfactory results level currently reached state-of-the-art methods data sets. experimentally show computational complexity reduction obtained makes approach eligible real-time applications.",3
"energy landscape deep networks. introduce ""annealsgd"", regularized stochastic gradient descent algorithm motivated analysis energy landscape particular class deep networks sparse random weights. loss function networks approximated hamiltonian spherical spin glass gaussian coupling. different currently-popular architectures convolutional ones, spin glasses amenable analysis, provides insights topology loss function motivates algorithms minimize it. specifically, show regularization term akin magnetic field modulated single scalar parameter transition loss function complex, non-convex landscape exponentially many local minima, phase polynomial number minima, way trivial landscape unique minimum. annealsgd starts training relaxed polynomial regime gradually tightens regularization parameter steer energy towards original exponential regime. even convolutional neural networks, quite unlike sparse random networks, empirically show annealsgd improves generalization error using competitive baselines mnist cifar-10.",3
"detecting adversarial perturbations. machine learning deep learning particular advanced tremendously perceptual tasks recent years. however, remains vulnerable adversarial perturbations input crafted specifically fool system quasi-imperceptible human. work, propose augment deep neural networks small ""detector"" subnetwork trained binary classification task distinguishing genuine data data containing adversarial perturbations. method orthogonal prior work addressing adversarial perturbations, mostly focused making classification network robust. show empirically adversarial perturbations detected surprisingly well even though quasi-imperceptible humans. moreover, detectors trained detect specific adversary, generalize similar weaker adversaries. addition, propose adversarial attack fools classifier detector novel training procedure detector counteracts attack.",18
"spatially resolved spectroscopy alkali metal vapour diffusing inside hollow-core photonic crystal fibres. present new type compact all-glass based vapour cell integrating hollow-core photonic crystal fibres. absence metals, traditional vacuum chamber much compact geometry allows fast homogeneous heating. consequence fill fibres much faster timescales, ranging minutes hours. additionally all-glass design ensures optical access along fibre. allows live monitoring diffusion rubidium atoms inside hollow-core measuring frequency-dependent fluorescence atoms. atomic density numerically retrieved using 5-level system bloch-equations.",14
"quasars proper motions link double multiple agns. gaia used large sample photometrically selected active galactic nuclei (agns) quasars remove residual spin global proper motion system order achieve maximally inertial reference frame. small fraction reference objects statistically significant astrometric proper motions gaia edr3. compile source sample $105,593$ high-fidelity agns accurate spectroscopically determined redshifts 0.5 sdss normalized proper motions 4. rate genuinely perturbed proper motions least 0.17\%. smaller high completeness sample 152 quasars excess proper motions confidence level 0.9995 examined detail. pan-starrs images gaia-resolved pairs reveal 29\% sample either double sources gravitationally lensed quasars. anderson--darling test parameters smaller high-reliability sample statistical controls reveals 17 significant factors favor multiplicity multi-source structure main cause perturbed astrometry. using nearest neighbor distance statistical analysis counts close companions gaia much larger initial sample agns, excess closely separated sources gaia detected. least 0.33\% optical quasars genuinely double multiply imaged. provide list 44 candidate double multiple agns four previously known gravitational lenses. many proper motion quasars may closely separated, unresolved doubles exhibiting variability imposed motion (vim) effect, smaller fraction may chance alignments foreground stars causing weak gravitational lensing.",0
learning similarity character recognition 3d object recognition. describe approach similarity motivated bayesian methods. yields similarity function learnable using standard bayesian methods. relationship approach variable kernel variable metric methods discussed. approach related variable kernel experimental results character recognition 3d object recognition presented..,3
"prediction daytime hypoglycemic events using continuous glucose monitoring data classification technique. daytime hypoglycemia accurately predicted achieve normoglycemia avoid disastrous situations. hypoglycemia, abnormally low blood glucose level, divided daytime hypoglycemia nocturnal hypoglycemia. many studies hypoglycemia prevention deal nocturnal hypoglycemia. paper, propose new predictor variables predict daytime hypoglycemia using continuous glucose monitoring (cgm) data. apply classification regression tree (cart) prediction method. independent variables prediction model rate decrease peak absolute level bg decision point. evaluation results showed model able detect almost 80% hypoglycemic events 15 min advance, higher existing methods similar conditions. proposed method might achieve real-time prediction well embedded bg monitoring device.",18
"cuckoo search: recent advances applications. cuckoo search (cs) relatively new algorithm, developed yang deb 2009, cs efficient solving global optimization problems. paper, review fundamental ideas cuckoo search latest developments well applications. analyze algorithm gain insight search mechanisms find efficient. also discuss essence algorithms link self-organizing systems, finally propose important topics research.",10
"train longer, generalize better: closing generalization gap large batch training neural networks. background: deep learning models typically trained using stochastic gradient descent one variants. methods update weights using gradient, estimated small fraction training data. observed using large batch sizes persistent degradation generalization performance - known ""generalization gap"" phenomena. identifying origin gap closing remained open problem. contributions: examine initial high learning rate training phase. find weight distance initialization grows logarithmically number weight updates. therefore propose ""random walk random landscape"" statistical model known exhibit similar ""ultra-slow"" diffusion behavior. following hypothesis conducted experiments show empirically ""generalization gap"" stems relatively small number updates rather batch size, completely eliminated adapting training regime used. investigate different techniques train models large-batch regime present novel algorithm named ""ghost batch normalization"" enables significant decrease generalization gap without increasing number updates. validate findings conduct several additional experiments mnist, cifar-10, cifar-100 imagenet. finally, reassess common practices beliefs concerning training deep models suggest may optimal achieve good generalization.",18
"fuzzy-based navigation control non-holonomic mobile robot. recent years, use non-analytical methods computing fuzzy logic, evolutionary computation, neural networks demonstrated utility potential paradigms intelligent control mobile robot navigation. paper, theoretical model fuzzy based controller autonomous mobile robot developed. paper begins mathematical model robot involves kinematic model. then, fuzzy logic controller developed discussed detail. proposed method successfully tested simulations, compares effectiveness three different set membership functions. shown fuzzy logic controller input membership three provides better performance compared five seven membership functions.",3
"neural coordination enhanced occasional interruption normal firing patterns: self-optimizing spiking neural network model. state space conventional hopfield network typically exhibits many different attractors small subset satisfy constraints neurons globally optimal fashion. recently demonstrated combining hebbian learning occasional alterations normal neural states avoids problem means self-organized enlargement best basins attraction. however, far clear extent process self-optimization also operative real brains. demonstrate transferred biologically plausible neural networks implementing self-optimizing spiking neural network model. addition, using spiking neural network emulate hopfield network hebbian learning, attempt make connection rate-based temporal coding based neural systems. although work required make model realistic, already suggests efficacy self-optimizing process independent simplifying assumptions conventional hopfield network. also discuss natural cultural processes could responsible occasional alteration neural firing patterns actual brains",12
"fraternal dropout. recurrent neural networks (rnns) important class architectures among neural networks useful language modeling sequential prediction. however, optimizing rnns known harder compared feed-forward neural networks. number techniques proposed literature address problem. paper propose simple technique called fraternal dropout takes advantage dropout achieve goal. specifically, propose train two identical copies rnn (that share parameters) different dropout masks minimizing difference (pre-softmax) predictions. way regularization encourages representations rnns invariant dropout mask, thus robust. show regularization term upper bounded expectation-linear dropout objective shown address gap due difference train inference phases dropout. evaluate model achieve state-of-the-art results sequence modeling tasks two benchmark datasets - penn treebank wikitext-2. also show approach leads performance improvement significant margin image captioning (microsoft coco) semi-supervised (cifar-10) tasks.",18
"adaptive amoeba algorithm shortest path tree computation dynamic graphs. paper presents adaptive amoeba algorithm address shortest path tree (spt) problem dynamic graphs. dynamic graphs, edge weight updates consists three categories: edge weight increases, edge weight decreases, mixture them. existing work problem solve issue analyzing nodes influenced edge weight updates recompute affected vertices. however, network becomes big, process become complex. proposed method overcome disadvantages existing approaches. important feature algorithm adaptivity. edge weight changes, proposed algorithm recognize affected vertices reconstruct spontaneously. evaluate proposed adaptive amoeba algorithm, compare label setting algorithm bellman-ford algorithm. comparison results demonstrate effectiveness proposed method.",3
"gravitational radiation higher order non-local gravity. paper examine gravitational radiation higher order non-local gravity described non-local gravitational lagrangian density $\mathcal{l}_{g}=r+\sum_{h=1}^{n}a_{h}r\box^{-h}r$. non-local theory gravitation always exhibits tensor transverse gravitational radiation $k_{1}^{2}=0$, corresponding angular frequency $\omega_{1}$, composed two standard $(+)$ $(\times)$ polarization modes, massless helicity 2. furthermore, shows, suitable constraint $n\geq 2$, additional massive transverse scalar gravitational radiation helicity 0. composed $n-1$ modes associated $n-1$ angular frequencies $\omega_{2},\ldots,\omega_{n}$, breathing polarization $(b)$ lowest order $\gamma$, parameter takes account difference speed slightly massive wave massless one. thanks np formalism, find $e(2)$ class non-local gravitational waves $n_{3}$, according petrov classification, presence absence modes observer independent. also, scalar radiation forbidden $n=1$ $n=2$ cases, conditions satisfied. finally, $\box^{-1}$ gravity $n=1$, possible degenerate case continuous infinity transverse massive scalar breathing modes appears particular constraint, reproduces two-dimensional spacetime polyakov effective action.",6
"reiter's default logic logic autoepistemic reasoning good one,. fact apparently observed earlier literature nonmonotonic reasoning reiter, default logic paper, directly formalize informal defaults. instead, translated default certain natural language proposition provided formalization latter. years later, moore noted propositions like one used reiter fundamentally different defaults exhibit certain autoepistemic nature. thus, reiter developed default logic formalization autoepistemic propositions rather defaults. first goal paper show problems reiter's default logic formal way reason informal defaults directly attributable autoepistemic nature default logic mismatch informal defaults reiter's formal defaults, latter formal expression autoepistemic propositions reiter used representation informal defaults. second goal paper compare work reiter moore. attempted formalize autoepistemic propositions, modes reasoning respective logics different. revisit moore's reiter's intuitions present perspective autotheoremhood, theories include propositions referring theory's theorems. discuss formalization perspective logics moore reiter, respectively, using unifying semantic framework default autoepistemic logics developed earlier. argue reiter's default logic better formalization moore's intuitions autoepistemic propositions moore's autoepistemic logic.",3
"non-elementary categoricity projective locally o-minimal classes. given cover $\mathbb{u}$ family smooth complex algebraic varieties, associate class $\mathcal{u},$ containing $\mathbb{u}$, structures locally definable o-minimal expansion reals. prove class $\aleph_0$-homogenous submodels stable. follows $\mathcal{u}$ categorical cardinality $\aleph_1.$ one-dimensional case prove slight modification $\mathcal{u}$ abstract elementary class categorical uncountable cardinals.",10
"riemann-theta boltzmann machine. general boltzmann machine continuous visible discrete integer valued hidden states introduced. mild assumptions connection matrices, probability density function visible units solved analytically, yielding novel parametric density function involving ratio riemann-theta functions. conditional expectation hidden state given visible states also calculated analytically, yielding derivative logarithmic riemann-theta function. conditional expectation used activation function feedforward neural network, thereby increasing modelling capacity network. boltzmann machine derived feedforward neural network successfully trained via standard gradient- non-gradient-based optimization techniques.",18
"exact solution $φ_{2}^{3}$ finite matrix model. find exact solutions $\phi_{2}^{3}$ finite matrix model (grosse-wulkenhaar model). $\phi_{2}^{3}$ finite matrix model, multipoint correlation functions expressed $g_{|a_{1}^{1}\ldots a_{n_{1}}^{1}|\ldots|a_{1}^{b}\ldots a_{n_{b}}^{b}|}$. $\displaystyle \sum_{i=1}^{b}n_{i}$-point function denoted $g_{|a_{1}^{1}\ldots a_{n_{1}}^{1}|\ldots|a_{1}^{b}\ldots a_{n_{b}}^{b}|}$ given sum feynman diagrams (ribbon graphs) riemann surfaces $b$-boundaries, $|a^{i}_{1}\cdots a^{i}_{n_{i}}|$ corresponds feynman diagrams $n_{i}$-external lines $i$-th boundary. known $g_{|a_{1}^{1}\ldots a_{n_{1}}^{1}|\ldots|a_{1}^{b}\ldots a_{n_{b}}^{b}|}$ expressed using $g_{|a^{1}|\ldots|a^{n}|}$ type $n$-point functions. thus focus rigorous calculations $g_{|a^{1}|\ldots|a^{n}|}$. formula $g_{|a^{1}|\ldots|a^{n}|}$ obtained, achieved using partition function $\mathcal{z}[j]$ calculated harish-chandra-itzykson-zuber integral. give $g_{|a|}$, $g_{|ab|}$, $g_{|a|b|}$, $g_{|a|b|c|}$ specific simple examples. described using airy functions.",9
"tight logic programs. note relationship two theories negation failure -- one based program completion, based stable models, answer sets. francois fages showed logic program satisfies certain syntactic condition, called ``tightness,'' stable models characterized models completion. extend definition tightness fages' theorem programs nested expressions bodies rules, study tight logic programs containing definition transitive closure predicate.",3
"keeping bad guys out: protecting vaccinating deep learning jpeg compression. deep neural networks (dnns) achieved great success solving variety machine learning (ml) problems, especially domain image recognition. however, recent research showed dnns highly vulnerable adversarially generated instances, look seemingly normal human observers, completely confuse dnns. adversarial samples crafted adding small perturbations normal, benign images. perturbations, imperceptible human eye, picked dnns cause misclassify manipulated instances high confidence. work, explore demonstrate systematic jpeg compression work effective pre-processing step classification pipeline counter adversarial attacks dramatically reduce effects (e.g., fast gradient sign method, deepfool). important component jpeg compression ability remove high frequency signal components, inside square blocks image. operation equivalent selective blurring image, helping remove additive perturbations. further, propose ensemble-based technique constructed quickly given well-performing dnn, empirically show ensemble leverages jpeg compression protect model multiple types adversarial attacks, without requiring knowledge model.",3
"trex: tomography reconstruction proximal framework robust sparse view x-ray applications. present trex, flexible robust tomographic reconstruction framework using proximal algorithms. provide overview perform experimental comparison famous iterative reconstruction methods terms reconstruction quality sparse view situations. derive proximal operators four best methods. show flexibility framework deriving solvers two noise models: gaussian poisson; plugging three powerful regularizers. compare framework state art methods, show superior quality synthetic real datasets.",10
"convexified modularity maximization degree-corrected stochastic block models. stochastic block model (sbm) popular framework studying community detection networks. model limited assumption nodes community statistically equivalent equal expected degrees. degree-corrected stochastic block model (dcsbm) natural extension sbm allows degree heterogeneity within communities. paper proposes convexified modularity maximization approach estimating hidden communities dcsbm. approach based convex programming relaxation classical (generalized) modularity maximization formulation, followed novel doubly-weighted $ \ell_1 $-norm $ k $-median procedure. establish non-asymptotic theoretical guarantees approximate clustering perfect clustering. approximate clustering results insensitive minimum degree, hold even sparse regime bounded average degrees. special case sbm, theoretical results match best-known performance guarantees computationally feasible algorithms. numerically, provide efficient implementation algorithm, applied synthetic real-world networks. experiment results show method enjoys competitive performance compared state art literature.",10
"deep convolutional neural network approach single-particle recognition cryo-electron microscopy. background: single-particle cryo-electron microscopy (cryo-em) become popular tool structural determination biological macromolecular complexes. high-resolution cryo-em reconstruction often requires hundreds thousands single-particle images. particle extraction experimental micrographs thus laborious presents major practical bottleneck cryo-em structural determination. existing computational methods particle picking often use low-resolution templates inputs particle matching, making possible cause reference-dependent bias. critical develop highly efficient template-free method automatically recognize particle images cryo-em micrographs. results: developed deep learning-based algorithmic framework, deepem, single-particle recognition noisy cryo-em micrographs, enabling automated particle picking, selection verification integrated fashion. kernel deepem built upon convolutional neural network (cnn) eight layers, recursively trained highly ""knowledgeable"". approach exhibits improved performance high precision tested standard klh dataset. application deepem several challenging experimental cryo-em datasets demonstrates capability avoiding selection un-wanted particles non-particles even true particles contain fewer features. conclusions: deepem method derived deep cnn allows automated particle extraction raw cryo-em micrographs absence templates, demonstrated improved performance, objectivity accuracy. application novel approach expected free labor involved single-particle verification, thus promoting efficiency cryo-em data processing.",14
"compliant conditions polynomial time approximation operator counts. paper, develop computationally simpler version operator count heuristic particular class domains. contribution abstract threefold, (1) propose efficient closed form approximation operator count heuristic using lagrangian dual; (2) leverage compressed sensing techniques obtain integer approximation operator counts polynomial time; (3) discuss relationship proposed formulation existing heuristics investigate properties domains approaches appear useful.",3
"noise-adaptive margin-based active learning lower bounds tsybakov noise condition. present simple noise-robust margin-based active learning algorithm find homogeneous (passing origin) linear separators analyze error convergence labels corrupted noise. show imposed noise satisfies tsybakov low noise condition (mammen, tsybakov, others 1999; tsybakov 2004) algorithm able adapt unknown level noise achieves optimal statistical rate poly-logarithmic factors. also derive lower bounds margin based active learning algorithms tsybakov noise conditions (tnc) membership query synthesis scenario (angluin 1988). result implies lower bounds stream based selective sampling scenario (cohn 1990) tnc fairly simple data distributions. quite surprisingly, show sample complexity cannot improved even underlying data distribution simple uniform distribution unit ball. proof involves construction well separated hypothesis set d-dimensional unit ball along carefully designed label distributions tsybakov noise condition. analysis might provide insights forms lower bounds well.",18
"influence learning method performance feedforward neural networks activity neurons modified. method allows us give different treatment neuron inside feedforward neural networks presented. algorithm implemented two different learning methods: standard back-propagation (bp) procedure evolutionary algorithm. first, demonstrated ea training method converges faster gives accurate results bp. made full analysis effects turning different combinations neurons training phase. demonstrate ea much robust bp cases study. even case two hidden neurons lost, ea training still able give good average results. difference implies must careful pruning redundancy effects studied since network performance losing neurons strongly depends training method. moreover, influence individual inputs also depend training algorithm. since ea keeps good classification performance units lost, method could good way simulate biological learning systems since must robust deficient neuron performance. although biological systems much complex simulations shown article, propose smart training strategy one shown could considered first protection losing certain number neurons.",3
"towards closing energy gap hog cnn features embedded vision. computer vision enables wide range applications robotics/drones, self-driving cars, smart internet things, portable/wearable electronics. many applications, local embedded processing preferred due privacy and/or latency concerns. accordingly, energy-efficient embedded vision hardware delivering real-time robust performance crucial. deep learning gaining popularity several computer vision algorithms, significant energy consumption difference exists compared traditional hand-crafted approaches. paper, provide in-depth analysis computation, energy accuracy trade-offs learned features deep convolutional neural networks (cnn) hand-crafted features histogram oriented gradients (hog). analysis supported measurements two chips implement algorithms. goal understand source energy discrepancy two approaches provide insight potential areas cnns improved eventually approach energy-efficiency hog maintaining outstanding performance accuracy.",3
"embarrassingly parallel variational inference nonconjugate models. develop parallel variational inference (vi) procedure use data-distributed settings, machine access subset data runs vi independently, without communicating machines. type ""embarrassingly parallel"" procedure recently developed mcmc inference algorithms; however, many cases possible directly extend procedure vi methods without requiring certain restrictive exponential family conditions form model. furthermore, existing (nonparallel) vi methods restricted use conditionally conjugate models, limits applicability. combat issues, make use recently proposed nonparametric vi facilitate embarrassingly parallel vi procedure applied wider scope models, including nonconjugate models. derive embarrassingly parallel vi algorithm, analyze method theoretically, demonstrate method empirically nonconjugate models.",18
"stochastic recursive gradient algorithm nonconvex optimization. paper, study analyze mini-batch version stochastic recursive gradient algorithm (sarah), method employing stochastic recursive gradient, solving empirical loss minimization case nonconvex losses. provide sublinear convergence rate (to stationary points) general nonconvex functions linear convergence rate gradient dominated functions, advantages compared modern stochastic gradient algorithms nonconvex losses.",18
"lithium depletion boundary, stellar associations, gaia. stellar ages key improving understanding different astrophysical phenomena. however, many techniques estimate stellar ages highly model-dependent. lithium depletion boundary (ldb), based presence absence lithium low-mass stars, used derive ages stellar associations 20 500~ma. purpose work revise former ldb ages stellar associations consistent way, taking advantage homogeneous \textit{gaia} parallaxes well bolometric luminosity estimations rely monochromatic bolometric corrections. studied nine open clusters three moving groups characterised previous determination ldb age. gathered available information data literature: membership, distances, photometric data, reddening, metallicity, surface gravity. re-assigned membership calculated bolometric luminosities effective temperatures using distances derived gaia dr2 multi-wavelength photometry individual objects around former ldb. located ldb using homogeneous method stellar associations. finally, estimated age comparing different evolutionary models. located ldb twelve stellar associations derived ages using several theoretical evolutionary models. compared ldb ages among them, along data obtained techniques, isochrone fitting, ultimately finding discrepancies among various approaches. finally, remark 32 ori mg likely composed least two populations different ages.",0
"short grb host galaxies ii: legacy sample redshifts, stellar population properties, implications neutron star merger origins. present stellar population properties 68 short gamma-ray burst (grb) host galaxies, representing largest uniformly-modeled sample date. using prospector stellar population inference code, jointly fit photometry and/or spectroscopy host galaxy. find population median redshift $z=0.60^{+0.90}_{-0.34}$ (68% confidence), including 10 new revised photometric redshifts $z\gtrsim1$. find median mass-weighted age $t_m=0.85^{+2.49}_{-0.57}$ gyr, stellar mass $\log(m_*/m_\odot)=9.69^{+0.95}_{-0.66}$, star formation rate sfr=$1.46^{+10.55}_{-1.38}m_\odot$/yr, stellar metallicity $\log(z_*/z_\odot)=-0.4^{+0.46}_{-0.42}$, dust attenuation $a_v=0.51^{+0.94}_{-0.43}$ mag (68% confidence). overall, majority short grb hosts star-forming ($\approx85$%), small fractions either transitioning ($\approx6$%) quiescent ($\approx$9%); however, observe much larger fraction ($\approx60$%) quiescent transitioning hosts $z\lesssim0.25$, commensurate galaxy evolution. find short grb hosts populate star-forming main sequence normal field galaxies, include many high-mass galaxies, implying binary neutron star (bns) merger progenitors dependent combination host star formation stellar mass. distribution ages redshifts implies broad delay-time distribution, fast-merging channel $z>1$ decreased bns formation efficiency lower redshifts. moreover, find short grbs originate environments wide range stellar metallicities. short grb hosts representative bns merger hosts within horizon current gravitational wave detectors, results inform future searches electromagnetic counterparts. data modeling products available bright website.",0
"duality certain multi-frobenius nonclassical curves higher dimensional spaces. show type multi-frobenius nonclassicality curve defined finite field $\mathbb{f}_q$ characteristic $p$ reflects geometry strict dual curve. particular, cases may describe possible intersection multiplicities strict dual curve linear system hyperplanes. among consequence, using result homma, able construct nonreflexive space curves tangent surfaces nonreflexive well, image generic point frobenius map osculating hyperplane. also obtain generalizations improvements known results literature.",10
"learning interactive segmentation system. many successful applications computer vision image video manipulation interactive nature. however, parameters systems often trained neglecting user. traditionally, interactive systems treated manner fully automatic counterparts. performance evaluated computing accuracy solutions fixed set user interactions. paper proposes new evaluation learning method brings user loop. based use active robot user - simulated model human user. show approach used evaluate learn parameters state-of-the-art interactive segmentation systems. also show simulated user models integrated popular max-margin method parameter learning propose algorithm solve resulting optimisation problem.",18
"mitigating evasion attacks deep neural networks via region-based classification. deep neural networks (dnns) transformed several artificial intelligence research areas including computer vision, speech recognition, natural language processing. however, recent studies demonstrated dnns vulnerable adversarial manipulations testing time. specifically, suppose testing example, whose label correctly predicted dnn classifier. attacker add small carefully crafted noise testing example dnn classifier predicts incorrect label, crafted testing example called adversarial example. attacks called evasion attacks. evasion attacks one biggest challenges deploying dnns safety security critical applications self-driving cars. work, develop new methods defend evasion attacks. key observation adversarial examples close classification boundary. therefore, propose region-based classification robust adversarial examples. benign/adversarial testing example, ensemble information hypercube centered example predict label. contrast, traditional classifiers point-based classification, i.e., given testing example, classifier predicts label based testing example alone. evaluation results mnist cifar-10 datasets demonstrate region-based classification significantly mitigate evasion attacks without sacrificing classification accuracy benign examples. specifically, region-based classification achieves classification accuracy testing benign examples point-based classification, region-based classification significantly robust point-based classification various evasion attacks.",3
"near-lossless l-infinity constrained multi-rate image decompression via deep neural network. recently number cnn-based techniques proposed remove image compression artifacts. restoration applications, techniques learn mapping decompressed patches original counterparts ubiquitous l2 metric. however, approach incapable restoring distinctive image details may statistical outliers high semantic importance (e.g., tiny lesions medical images). overcome weakness, propose incorporate l-infinity fidelity criterion design neural network small, distinctive structures original image dropped distorted. moreover, anti-artifacts neural network designed work range compression bit rates, rather fixed one past. experimental results demonstrate proposed method outperforms state-of-the-art methods l-infinity error metric perceptual quality, competitive l2 error metric well. restore subtle image details otherwise destroyed missed algorithms. research suggests new machine learning paradigm ultra high fidelity image compression ideally suited applications medicine, space, sciences.",3
"alberti's letter counts. four centuries modern statistical linguistics born, leon battista alberti (1404--1472) compared frequency vowels latin poems orations, making first quantified observation stylistic difference ever. using corpus 20 latin texts (over 5 million letters), alberti's observations statistically assessed. letter counts prove poets used significantly a's, e's, y's, whereas orators used vowels. sample sizes needed justify assertions studied, proved within reach alberti's scholarship.",10
"annotation graphs framework multidimensional linguistic data analysis. recent work presented formal framework linguistic annotation based labeled acyclic digraphs. `annotation graphs' offer simple yet powerful method representing complex annotation structures incorporating hierarchy overlap. here, motivate illustrate approach using discourse-level annotations text speech data drawn callhome, coconut, muc-7, damsl trains annotation schemes. help domain specialists, constructed hybrid multi-level annotation fragment boston university radio speech corpus includes following levels: segment, word, breath, tobi, tilt, treebank, coreference named entity. show annotation graphs represent hybrid multi-level structures derive diverse set file formats. also show approach facilitates substantive comparison multiple annotations single signal based different theoretical models. discussion shows annotation graphs open door wide-ranging integration tools, formats corpora.",3
"information content coevolutionary game landscapes. coevolutionary game dynamics result players may change strategies network interaction. games, based interpreting strategies configurations, strategy-to-payoff maps defined every interaction network, opens derive game landscapes. paper presents analysis game landscapes information content. analysis, particularly study effect rescaled payoff matrix generalizing social dilemmas differences well-mixed structured populations.",15
"information transfer swarms leaders. swarm dynamics study collections agents interact one another without central control. natural systems, insects, birds, fish large mammals function larger units increase overall fitness individuals. behavior coordinated local interactions enhance mate selection, predator detection, migratory route identification forth [andersson wallander 2003; buhl et al. 2006; nagy et al. 2010; partridge 1982; sumpter et al. 2008]. artificial systems, swarms autonomous agents augment human activities search rescue, environmental monitoring covering large areas multiple nodes [alami et al. 2007; caruso et al. 2008; ogren et al. 2004; paley et al. 2007; sibley et al. 2002]. paper, explore interplay swarm dynamics, covert leadership theoretical information transfer. leader member swarm acts upon information addition provided local interactions. depending upon leadership model, leaders use external information either time response local conditions [couzin et al. 2005; sun et al. 2013]. covert leader leader treated differently others swarm, leaders followers participate equally whatever interaction model used [rossi et al. 2007]. study, use theoretical information transfer means analyzing swarm interactions explore whether possible distinguish followers leaders based interactions within swarm. find covert leaders distinguished followers swarm receive less transfer entropy followers.",3
"linear contour learning: method supervised dimension reduction. propose novel approach sufficient dimension reduction regression, based estimating contour directions negligible variation response surface. directions span orthogonal complement minimal space relevant regression, extracted according measure variation response, leading general contour regression(gcr). comparison exiisting sufficient dimension reduction techniques, sontour-based mothology guarantees exhaustive estimation central space ellipticity predictoor distribution mild additional assumptions, maintaining vn-consisytency somputational ease. moreover, proves robust departures ellipticity. also establish useful population properties gcr. simulations compare performance standard techniques ordinary least squares, sliced inverse regression, principal hessian directions, sliced average variance estimation confirm advntages anticipated theoretical analyses. also demonstrate use contour-based methods data set concerning grades students massachusetts colleges.",3
"extracting keyword disambiguating name based overlap principle. name disambiguation become one main themes semantic web agenda. semantic web extension current web information given well-defined meaning, also many purposes contain ambiguous naturally lot thing came overlap, mainly deals persons name. therefore, develop approach extract keywords web snippet utilizing overlap principle, concept understand things ambiguous, whereby features person generated dealing variety web, web steadily gaining ground semantic research.",3
"weakly supervised image annotation segmentation objects attributes. propose model complex visual scenes using non-parametric bayesian model learned weakly labelled images abundant media sharing sites flickr. given weak image-level annotations objects attributes without locations associations them, model aims learn appearance object attribute classes well association object instance. learned, given image, model deployed tackle number vision problems joint coherent manner, including recognising objects scene (automatic object annotation), describing objects using attributes (attribute prediction association), localising delineating objects (object detection semantic segmentation). achieved developing novel weakly supervised markov random field stacked indian buffet process (ws-mrf-sibp) models objects attributes latent factors explicitly captures correlations within across superpixels. extensive experiments benchmark datasets demonstrate weakly supervised model significantly outperforms weakly supervised alternatives often comparable existing strongly supervised models variety tasks including semantic segmentation, automatic image annotation retrieval based object-attribute associations.",3
"time-sensitive customer churn prediction based pu learning. fast development internet companies throughout world, customer churn become serious concern. better help companies retain customers, important build customer churn prediction model identify customers likely churn ahead time. paper, propose time-sensitive customer churn prediction (tccp) framework based positive unlabeled (pu) learning technique. specifically, obtain recent data shortening observation period, start train model long enough positive samples collected, ignoring absence negative examples. conduct thoroughly experiments real industry data alipay.com. experimental results demonstrate tccp outperforms rule-based models traditional supervised learning models.",3
"generating precondition expressions instructional text. study employs knowledge intensive corpus analysis identify elements communicative context used determine appropriate lexical grammatical form instructional texts. \ig, instructional text generation system based analysis, presented, particularly reference expression precondition relations.",1
"algorithm noisy k-means. note, introduce new algorithm deal finite dimensional clustering errors variables. design algorithm based recent theoretical advances (see loustau (2013a,b)) statistical learning errors variables. previous mentioned papers, algorithm mixes different tools inverse problem literature machine learning community. coarsely, based two-step procedure: (1) deconvolution step deal noisy inputs (2) newton's iterations popular k-means.",18
"robotic multi-object focal plane system dark energy spectroscopic instrument (desi). system 5,020 robotic fiber positioners installed 2019 mayall telescope, kitt peak national observatory. robots automatically re-target optical fibers every 10 - 20 minutes, precision several microns, reconfiguration time less 2 minutes. next five years, enable newly-constructed dark energy spectroscopic instrument (desi) measure spectra 35 million galaxies quasars. desi produce largest 3d map universe date measure expansion history cosmos. addition 5,020 robotic positioners optical fibers, desi's focal plane system includes 6 guide cameras, 4 wavefront cameras, 123 fiducial point sources, metrology camera mounted primary mirror. system also includes associated structural, thermal, electrical systems. all, contains 675,000 individual parts. discuss design, construction, quality control, integration components. include summary key requirements, review acceptance process, on-sky validations requirements, lessons learned future multi-object, fiber-fed spectrographs.",0
"efficient unsupervised temporal segmentation motion data. introduce method automated temporal segmentation human motion data distinct actions compositing motion primitives based self-similar structures motion sequence. use neighbourhood graphs partitioning similarity information graph exploited cluster motion primitives larger entities semantic significance. method requires assumptions motion sequences hand user interaction required segmentation clustering. addition, introduce feature bundling preprocessing technique make segmentation robust noise, well notion motion symmetry refined primitive detection. test method several sensor modalities, including markered markerless motion capture well electromyograph accelerometer recordings. results highlight system's capabilities segmentation analysis finer structures motion data, completely unsupervised manner.",3
"locally scale-invariant convolutional neural networks. convolutional neural networks (convnets) shown excellent results many visual classification tasks. exception imagenet, datasets carefully crafted objects well-aligned similar scales. naturally, feature learning problem gets challenging amount variation data increases, models learn invariant certain changes appearance. recent results imagenet dataset show given enough data, convnets learn invariances producing discriminative features [1]. could more: use less parameters, less data, learn discriminative features, certain invariances built learning process? paper present simple model allows convnets learn features locally scale-invariant manner without increasing number model parameters. show modified mnist dataset faced scale variation, building scale-invariance allows convnets learn discriminative features reduced chances over-fitting.",3
"first short grb millimeter afterglow: wide-angled jet extremely energetic sgrb 211106a. present discovery first millimeter afterglow short-duration $\gamma$-ray burst (sgrb) first confirmed afterglow sgrb localized guano system swift. atacama large millimeter/sub-millimeter array (alma) detection sgrb 211106a solidifies origin faint host galaxy detected hubble space telescope (hst) imaging projected separation $\approx0.8$ kpc. millimeter-band light curve captures passage synchrotron peak afterglow forward shock, constraining afterglow kinetic energy, $\log(e_{\rm k,iso}/{\rm erg})=53.2\pm0.3$ density, $\log(n_0/{\rm cm}^{-3})=-0.6\pm0.2$ presumed redshift $z=1$. identify jet break $t_{\rm jet}=29.2^{+4.5}_{-4.0}$ days millimeter-band data infer opening angle $\theta_{\rm jet}=(15.5\pm1.4)$ degrees beaming-corrected kinetic energy $\log(e_{\rm k}/{\rm erg})=54.3\pm0.3$, widest highest ever measured sgrb, respectively. lack detectable optical afterglow, coupled bright millimeter counterpart, infer high extinction, $a_{\rm v}\gtrsim2.6$ mag along line sight, making one highly dust-extincted sgrbs known date. combining published millimeter-band upper limits conjunction energetics large sample sgrbs, find energetic, wide-angled outflows high density environments likely detectable millimeter counterparts. concerted afterglow searches alma yield detection fractions 24--40\% timescales $>2$ days rates $\approx0.8$--1.6 per year, outpacing historical discovery rate sgrb centimeter-band afterglows.",0
"lightweight neural networks. weights lightweight neural network value zero, remaining ones either +1 -1. universal approximators require approximately 1.1 bits/weight storage, posses quick forward pass achieve classification accuracies similar conventional continuous-weight networks. training regimen focuses error reduction initially, later emphasizes discretization weights. ignore insignificant inputs, remove unnecessary weights, drop unneeded hidden neurons. successfully tested mnist, credit card fraud, credit card defaults data sets using networks 2 16 hidden layers 4.4 million weights.",3
"thinking adaptive: towards behaviours virtual laboratory. paper name advantages virtual laboratories; propose behaviours virtual laboratory useful biologists ai researchers, offering new perspective understanding adaptive behaviour. present development behaviours virtual laboratory, stage focused action selection, show experiments illustrate properties proposal, accessed via internet.",3
"multi-view learning nonparametric nonlinear inter-battery factor analysis. factor analysis aims determine latent factors, traits, summarize given data set. inter-battery factor analysis extends notion multiple views data. paper show nonlinear, nonparametric version models recovered gaussian process latent variable model. gives us flexible formalism multi-view learning latent variables used exploratory purposes learning representations enable efficient inference ambiguous estimation tasks. learning performed bayesian manner formulation variational compression scheme gives rigorous lower bound log likelihood. bayesian framework provides strong regularization training, allowing structure latent space determined efficiently automatically. demonstrate producing first (to knowledge) published results learning dozens views, even data scarce. show experimental results several different types multi-view data sets different kinds tasks, including exploratory data analysis, generation, ambiguity modelling latent priors classification.",18
"towards learning-theoretic analysis spike-timing dependent plasticity. paper suggests learning-theoretic perspective synaptic plasticity benefits global brain functioning. introduce model, selectron, (i) arises fast time constant limit leaky integrate-and-fire neurons equipped spiking timing dependent plasticity (stdp) (ii) amenable theoretical analysis. show selectron encodes reward estimates spikes error bound spikes controlled spiking margin sum synaptic weights. moreover, efficacy spikes (their usefulness reward maximizing selectrons) also depends total synaptic strength. finally, based analysis, propose regularized version stdp, show regularization improves robustness neuronal learning faced multiple stimuli.",15
"deviation optimal learning using greedy q-aggregation. given finite family functions, goal model selection aggregation construct procedure mimics function family closest unknown regression function. precisely, consider general regression model fixed design measure distance functions mean squared error design points. procedures based exponential weights known solve problem model selection aggregation expectation, are, surprisingly, sub-optimal deviation. propose new formulation called q-aggregation addresses limitation; namely, solution leads sharp oracle inequalities optimal minimax sense. moreover, based new formulation, design greedy q-aggregation procedures produce sparse aggregation models achieving optimal rate. convergence performance greedy procedures illustrated compared standard methods simulated examples.",10
"shared representation learning heterogeneous face recognition. intensive research, heterogenous face recognition still challenging problem. main difficulties owing complex relationship heterogenous face image spaces. heterogeneity always tightly coupled variations, makes relationship heterogenous face images highly nonlinear. many excellent methods proposed model nonlinear relationship, apt overfit training set, due limited samples. inspired unsupervised algorithms deep learning, paper proposes novel framework heterogeneous face recognition. first extract gabor features localized facial points, use restricted boltzmann machines (rbms) learn shared representation locally remove heterogeneity around facial point. finally, shared representations local rbms connected together processed pca. two problems (sketch-photo nir-vis) three databases selected evaluate proposed method. sketch-photo problem, obtain perfect results cufs database. nir-vis problem, produce new state-of-the-art performance casia hfb nir-vis 2.0 databases.",3
"recovery simultaneous low rank two-way sparse coefficient matrices, nonconvex approach. study problem recovery matrices simultaneously low rank row and/or column sparse. matrices appear recent applications cognitive neuroscience, imaging, computer vision, macroeconomics, genetics. propose gdt (gradient descent hard thresholding) algorithm efficiently recover matrices structure, minimizing bi-convex function nonconvex set constraints. show linear convergence iterates obtained gdt region within statistical error optimal solution. application method, consider multi-task learning problems show statistical error rate obtained gdt near optimal compared minimax rate. experiments demonstrate competitive performance much faster running speed compared existing methods, simulations real data sets.",18
"aognets: deep and-or grammar networks visual recognition. paper presents method learning deep and-or grammar (aog) networks visual recognition, term aognets. aognet consists number stages composed number aog building blocks. aog building block designed based principled and-or grammar represented hierarchical compositional and-or graph. node applies basic operation (e.g., conv-batchnorm-relu) input. three types nodes: and-node explores composition, whose input computed concatenating features child nodes; or-node represents alternative ways composition spirit exploitation, whose input element-wise sum features child nodes; terminal-node takes input channel-wise slice input feature map aog building block. aognets aim harness best two worlds (grammar models deep neural networks) representation learning end-to-end training. experiments, aognets tested three highly competitive image classification benchmarks: cifar-10, cifar-100 imagenet-1k. aognets obtain better performance widely used residual net variants, tightly comparable dense net. aognets also tested object detection pascal voc 2007 2012 using vanilla faster rcnn system obtain better performance residual net.",3
"double-quantitative $γ^{\ast}-$fuzzy coverings approximation operators. digital-based information boom, fuzzy covering rough set model important mathematical tool artificial intelligence, build bridge fuzzy covering rough set theory pawlak's model becoming hot research topic. paper, first present $\gamma-$fuzzy covering based probabilistic grade approximation operators double-quantitative approximation operators. also study relationships among three types $\gamma-$fuzzy covering based approximation operators. second, propose $\gamma^{\ast}-$fuzzy coverings based multi-granulation probabilistic grade lower upper approximation operators multi-granulation double-quantitative lower upper approximation operators. also investigate relationships among types $\gamma-$fuzzy coverings based approximation operators. finally, employ several examples illustrate construct lower upper approximations fuzzy sets absolute relative quantitative information.",3
"seismology solar system. nasa insight mission helped measure deep interior mars using observations seismic waves excited marsquakes. currently, installation seismometers moon foreseen. review case seismic experiments major planetary bodies solar system. discuss scientific goals accordance decadal survey planetary science astrobiology esa voyage 2050 program well technical challenges potential mission concepts, answer question: could seismology planets it?",0
"structured convex optimization submodular constraints. number discrete continuous optimization problems machine learning related convex minimization problems submodular constraints. paper, deal submodular function directed graph structure, show wide range convex optimization problems submodular constraints solved much efficiently general submodular optimization methods reduction maximum flow problem. furthermore, give applications, including sparse optimization methods, proposed methods effective. additionally, evaluate performance proposed method computational experiments.",3
"fast stochastic methods nonsmooth nonconvex optimization. analyze stochastic algorithms optimizing nonconvex, nonsmooth finite-sum problems, nonconvex part smooth nonsmooth part convex. surprisingly, unlike smooth case, knowledge fundamental problem limited. example, known whether proximal stochastic gradient method constant minibatch converges stationary point. tackle issue, develop fast stochastic algorithms provably converge stationary point constant minibatches. furthermore, using variant algorithms, show provably faster convergence batch proximal gradient descent. finally, prove global linear convergence rate interesting subclass nonsmooth nonconvex functions, subsumes several recent works. paper builds upon recent series papers fast stochastic methods smooth nonconvex optimization [22, 23], novel analysis nonconvex nonsmooth functions.",10
"hybrid linear modeling via local best-fit flats. present simple fast geometric method modeling data union affine subspaces. method begins forming collection local best-fit affine subspaces, i.e., subspaces approximating data local neighborhoods. correct sizes local neighborhoods determined automatically jones' $\beta_2$ numbers (we prove certain geometric conditions method finds optimal local neighborhoods). collection subspaces processed greedy selection procedure spectral method generate final model. discuss applications tracking-based motion segmentation clustering faces different illuminating conditions. give extensive experimental evidence demonstrating state art accuracy speed suggested algorithms problems also synthetic hybrid linear data well mnist handwritten digits data; demonstrate use algorithms fast determination number affine subspaces.",3
"upgrading pulse detection time shift properties using wavelets support vector machines. current approaches pulse detection use domain transformations concentrate frequency related information distinguishable noise. real cases know pulse begin, need time search process time windows scheduled analysed. window contain pulsed signal (either complete incomplete) / noise. paper simple search process introduced, allowing algorithm process information, upgrading capabilities terms probability detection (pd) probability false alarm (pfa).",3
"injury risk prediction traffic accidents porto alegre/rs, brazil. study describes experimental application machine learning techniques build prediction models assess injury risk associated traffic accidents. work uses freely available data set traffic accident records took place city porto alegre/rs (brazil) year 2013. study also provides analysis important attributes traffic accident could produce outcome injury people involved accident.",3
"asymptotic comparison identifying constraints bradley-terry models. bradley-terry model widely used pairwise comparison data analysis. paper, analyze asymptotic behavior maximum likelihood estimator bradley-terry model logistic parameterization, general class linear identifiability constraints. show constraint requiring bradley-terry scores compared objects sum zero minimizes sum variances estimated scores, recommend using constraint practice.",10
"deep narrow boltzmann machines universal approximators. show deep narrow boltzmann machines universal approximators probability distributions activities visible units, provided sufficiently many hidden layers, containing number units visible layer. show that, within certain parameter domains, deep boltzmann machines studied feedforward networks. provide upper lower bounds sufficient depth width universal approximators. results settle various intuitions regarding undirected networks and, particular, show deep narrow boltzmann machines least compact universal approximators narrow sigmoid belief networks restricted boltzmann machines, respect currently available bounds models.",18
"metric distances derived cosine similarity pearson spearman correlations. investigate two classes transformations cosine similarity pearson spearman correlations metric distances, utilising simple tool metric-preserving functions. first class puts anti-correlated objects maximally far apart. previously known transforms fall within class. second class collates correlated anti-correlated objects. example transformation yields metric distance sine function applied centered data.",18
"images lie: transferring deep visual semantic features large-scale multimodal learning rank. search heart modern e-commerce. result, task ranking search results automatically (learning rank) multibillion dollar machine learning problem. traditional models optimize hand-constructed features based item's text. paper, introduce multimodal learning rank model combines traditional features visual semantic features transferred deep convolutional neural network. large scale experiment using data online marketplace etsy, verify moving multimodal representation significantly improves ranking quality. show image features capture fine-grained style information available text-only representation. addition, show concrete examples image information successfully disentangle pairs highly different items ranked similarly text-only model.",3
"approach text steganography based markov chains. text steganography method based markov chains introduced, together reference implementation. method allows information hiding texts automatically generated following given markov model. markov - based systems kind rely big simplifications language model work, produces less natural looking easily detectable texts. method described designed generate texts within good approximation original language model provided.",3
"multidimensional counting grids: inferring word order disordered bags words. models bags words typically assume topic mixing words single bag come limited number topics. show many sets bag words exhibit different pattern variation patterns efficiently captured topic mixing. many cases, one bag words next, words disappear new ones appear theme slowly smoothly shifted across documents (providing documents somehow ordered). examples latent structure describe ordering easily imagined. example, advancement date news stories reflected smooth change theme day certain evolving news stories fall favor new events create new stories. overlaps among stories consecutive days modeled using windows linearly arranged tight distributions words. show strategy extended multiple dimensions cases ordering data readily obvious. demonstrate way modeling covariation word occurrences outperforms standard topic models classification prediction tasks applications biology, text modeling computer vision.",3
"robust discriminative clustering sparse regularizers. clustering high-dimensional data often requires form dimensionality reduction, clustered variables separated ""noise-looking"" variables. cast problem finding low-dimensional projection data well-clustered. yields one-dimensional projection simplest situation two clusters, extends naturally multi-label scenario two clusters. paper, (a) first show joint clustering dimension reduction formulation equivalent previously proposed discriminative clustering frameworks, thus leading convex relaxations problem, (b) propose novel sparse extension, still cast convex relaxation allows estimation higher dimensions, (c) propose natural extension multi-label scenario, (d) provide new theoretical analysis performance formulations simple probabilistic model, leading scalings form $d=o(\sqrt{n})$ affine invariant case $d=o(n)$ sparse case, $n$ number examples $d$ ambient dimension, finally, (e) propose efficient iterative algorithm running-time complexity proportional $o(nd^2)$, improving earlier algorithms quadratic complexity number examples.",18
"inferring strategies limited reconnaissance real-time strategy games. typical real-time strategy (rts) games, enemy units visible within sight range friendly unit. knowledge opponent's disposition limited observed scouting. information costly, since units dedicated scouting unavailable purposes, enemy resist scouting attempts. important infer much possible opponent's current future strategy available observations. present dynamic bayes net model strategies rts game starcraft combines generative model strategies relate observable quantities principled framework incorporating evidence gained via scouting. demonstrate model's ability infer unobserved aspects game realistic observations.",3
"neural network approach railway stand lateral skew control. paper presents study adaptive approach lateral skew control experimental railway stand. preliminary experiments real experimental railway stand simulations 3-d mechanical model, indicates difficulties model-based control device. thus, use neural networks identification control lateral skew shall investigated. paper focuses real-data based modeling railway stand various neural network models, i.e; linear neural unit quadratic neural unit architectures. furthermore, training methods neural architectures such, real-time-recurrent-learning variation back-propagation-through-time examined, accompanied discussion produced experimental results.",3
"simple approach multilingual polarity classification twitter. recently, sentiment analysis received lot attention due interest mining opinions social media users. sentiment analysis consists determining polarity given text, i.e., degree positiveness negativeness. traditionally, sentiment analysis algorithms tailored specific language given complexity number lexical variations errors introduced people generating content. contribution, aim provide simple implement easy use multilingual framework, serve baseline sentiment analysis contests, starting point build new sentiment analysis systems. compare approach eight different languages, three important international contests, namely, semeval (english), tass (spanish), sentipolc (italian). within competitions approach reaches medium high positions rankings; whereas remaining languages approach outperforms reported results.",3
"pairwise mrf calibration perturbation bethe reference point. investigate different ways generating approximate solutions pairwise markov random field (mrf) selection problem. focus mainly inverse ising problem, discuss also somewhat related inverse gaussian problem types mrf suitable inference tasks belief propagation algorithm (bp) certain conditions. approach consists take bethe mean-field solution obtained maximum spanning tree (mst) pairwise mutual information, referred \emph{bethe reference point}, perturbation procedures. consider three different ways following idea: first one, select calibrate iteratively optimal links added starting bethe reference point; second one based observation natural gradient computed analytically bethe point; third one, assuming local field using low temperature expansion develop dual loop joint model based well chosen fundamental cycle basis. indeed identify subclass planar models, refer \emph{bethe-dual graph models}, possibly many loops, characterized singly connected dual factor graph, partition function linear response computed exactly respectively o(n) $o(n^2)$ operations, thanks dual weight propagation (dwp) message passing procedure set up. restricted subclass models, inverse ising problem convex, becomes tractable temperature. experimental tests various datasets refined $l_0$ $l_1$ regularization procedures indicate approaches may competitive useful alternatives existing ones.",2
"adaptive compressive tracking via online vector boosting feature selection. recently, compressive tracking (ct) method attracted much attention due high efficiency, cannot well deal large scale target appearance variations due data-independent random projection matrix results less discriminative features. address issue, paper propose adaptive ct approach, selects discriminative features design effective appearance model. method significantly improves ct three aspects: firstly, discriminative features selected via online vector boosting method. secondly, object representation updated effective online manner, preserves stable features filtering noisy ones. finally, simple effective trajectory rectification approach adopted make estimated location accurate. extensive experiments cvpr2013 tracking benchmark demonstrate superior performance algorithm compared state-of-the-art tracking algorithms.",3
"min-max kernels. min-max kernel generalization popular resemblance kernel (which designed binary data). paper, demonstrate, extensive classification study using kernel machines, min-max kernel often provides effective measure similarity nonnegative data. min-max kernel nonlinear might difficult used industrial applications massive data, show min-max kernel linearized via hashing techniques. allows practitioners apply min-max kernel large-scale applications using well matured linear algorithms linear svm logistic regression. previous remarkable work consistent weighted sampling (cws) produces samples form ($i^*, t^*$) $i^*$ records location (and fact also weights) information analogous samples produced classical minwise hashing binary data. $t^*$ theoretically unbounded, immediately clear effectively implement cws building large-scale linear classifiers. paper, provide simple solution discarding $t^*$ (which refer ""0-bit"" scheme). via extensive empirical study, show 0-bit scheme lose essential information. apply ""0-bit"" cws building linear classifiers approximate min-max kernel classifiers, extensively validated wide range publicly available classification datasets. expect work generate interests among data mining practitioners would like efficiently utilize nonlinear information non-binary nonnegative data.",18
"towards personalized modeling female hormonal cycle: experiments mechanistic models gaussian processes. paper, introduce novel task machine learning healthcare, namely personalized modeling female hormonal cycle. motivation work model hormonal cycle predict phases time, healthy individuals disorders reproductive system. individual differences menstrual cycle, particularly interested personalized models account individual idiosyncracies, towards identifying phenotypes menstrual cycles. first step, consider hormonal cycle set observations time. use previously validated mechanistic model generate realistic hormonal patterns, experiment gaussian process regression estimate values time. specifically, interested feasibility predicting menstrual cycle phases varying learning conditions: number cycles used training, hormonal measurement noise sampling rates, informed vs. agnostic sampling hormonal measurements. results indicate gaussian processes help model female menstrual cycle. discuss implications experiments context modeling female menstrual cycle.",18
"penalized orthogonal-components regression large p small n data. propose penalized orthogonal-components regression (pocre) large p small n data. orthogonal components sequentially constructed maximize, upon standardization, correlation response residuals. new penalization framework, implemented via empirical bayes thresholding, presented effectively identify sparse predictors component. pocre computationally efficient owing sequential construction leading sparse principal components. addition, construction offers properties grouping highly correlated predictors allowing collinear nearly collinear predictors. multivariate responses, pocre construct common components thus build latent-variable models large p small n data.",18
"improving gaussian process sparse spectrum approximation representing uncertainty frequency inputs. standard sparse pseudo-input approximations gaussian process (gp) cannot handle complex functions well. sparse spectrum alternatives attempt answer known over-fit. suggest use variational inference sparse spectrum approximation avoid issues. model covariance function finite fourier series approximation treat random variable. random covariance function posterior, variational distribution placed. variational distribution transforms random covariance function fit data. study properties approximate inference, compare alternative ones, extend distributed stochastic domains. approximation captures complex functions better standard approaches avoids over-fitting.",18
"feature learning incomplete eeg denoising autoencoder. alternative pathway human brain communicate outside world means brain computer interface (bci). bci decode electroencephalogram (eeg) signals brain activities, send command intent external interactive device, wheelchair. effectiveness bci depends performance decoding eeg. usually, eeg contaminated different kinds artefacts (e.g., electromyogram (emg), background activity), leads low decoding performance. number filtering methods utilized remove weaken effects artefacts, generally fail eeg contains extreme artefacts. cases, common approach discard whole data segment containing extreme artefacts. causes fatal drawback bci cannot output decoding results time. order solve problem, employ lomb-scargle periodogram estimate spectral power incomplete eeg (after removing parts contaminated artefacts), denoising autoencoder (dae) learning. proposed method evaluated motor imagery eeg data. results show method successfully decode incomplete eeg good effect.",3
"quantum clustering gaussian mixtures. mixture gaussian distributions, soft version k-means , considered state-of-the-art clustering algorithm. widely used computer vision selecting classes, e.g., color, texture, shapes. algorithm, class described gaussian distribution, defined mean covariance. data described weighted sum gaussian distributions. propose new method, inspired quantum interference physics. instead modeling class distribution directly, model class wave function magnitude square class gaussian distribution. mix class wave functions create mixture wave function. final mixture distribution magnitude square mixture wave function. result, observe quantum class interference phenomena, present gaussian mixture model. show quantum method outperforms gaussian mixture method every aspect estimations. provides accurate estimations distribution parameters, much less fluctuations, also robust data deformations gaussian assumptions. illustrate method color segmentation example application.",18
"machine learning methods computational biology cancer. objectives ""perspective"" paper review recent advances sparse feature selection regression classification, well compressed sensing, discuss might used develop tools advance personalized cancer therapy. illustration possibilities, new algorithm sparse regression presented, applied predict time tumor recurrence ovarian cancer. new algorithm sparse feature selection classification problems presented, validation endometrial cancer briefly discussed. open problems also presented.",15
"analyzing sparse dictionaries online learning kernels. many signal processing machine learning methods share essentially linear-in-the-parameter model, many parameters available samples kernel-based machines. sparse approximation essential many disciplines, new challenges emerging online learning kernels. end, several sparsity measures proposed literature quantify sparse dictionaries constructing relevant ones, prolific ones distance, approximation, coherence babel measures. paper, analyze sparse dictionaries based measures. conducting eigenvalue analysis, show sparsity measures share many properties, including linear independence condition inducing well-posed optimization problem. furthermore, prove exists quasi-isometry parameter (i.e., dual) space dictionary's induced feature space.",18
"distant supervision emotion detection using facebook reactions. exploit facebook reaction feature distant supervised fashion train support vector machine classifier emotion detection, using several feature combinations combining different facebook pages. test models existing benchmarks emotion detection show employing information derived completely automatically, thus without relying handcrafted lexicon usually done, achieve competitive results. results also show large room improvement, especially gearing collection facebook pages, view target domain.",3
"dynamic clustering algorithms via small-variance analysis markov chain mixture models. bayesian nonparametrics class probabilistic models model size inferred data. recently developed methodology field small-variance asymptotic analysis, mathematical technique deriving learning algorithms capture much flexibility bayesian nonparametric inference algorithms, simpler implement less computationally expensive. past work small-variance analysis bayesian nonparametric inference algorithms exclusively considered batch models trained single, static dataset, incapable capturing time evolution latent structure data. work presents small-variance analysis maximum posteriori filtering problem temporally varying mixture model markov dependence structure, captures temporally evolving clusters within dataset. two clustering algorithms result analysis: d-means, iterative clustering algorithm linearly separable, spherical clusters; sd-means, spectral clustering algorithm derived kernelized, relaxed version clustering problem. empirical results experiments demonstrate advantages using d-means sd-means contemporary clustering algorithms, terms computational cost clustering accuracy.",18
"probabilistic framework location inference social media. study extent infer users' geographical locations social media. location inference social media benefit many applications, disaster management, targeted advertising, news content tailoring. recent years, number algorithms proposed identifying user locations social media platforms twitter facebook message contents, friend networks, interactions users. paper, propose novel probabilistic model based factor graphs location inference offers several unique advantages task. first, model generalizes previous methods incorporating content, network, deep features learned social context. model also flexible enough support supervised learning semi-supervised learning. second, explore several learning algorithms proposed model, present two-chain metropolis-hastings (mh+) algorithm, improves inference accuracy. third, validate proposed model three different genres data - twitter, weibo, facebook - demonstrate proposed model substantially improve inference accuracy (+3.3-18.5% f1-score) several state-of-the-art methods.",3
"local differential privacy evolving data. several large scale deployments differential privacy used track statistical information users. however, systems periodically recollect data recompute statistics using algorithms designed single use result provide meaningful privacy guarantees long time scales. moreover, existing techniques mitigate effect apply ""local"" model differential privacy systems use. paper, introduce new local differential privacy technique maintain persistently up-to-date statistics time, privacy guarantees scaling number changes underlying distribution rather number collection periods. key ideas include batching time epochs -- varying epoch size allows us trade accuracy frequency updates -- protocol users ""vote"" update out-of-date statistics losing little privacy. prove main results setting users hold single bit, redrawn every time period, common (but changing) distribution; however, framework quite general give application frequency heavy-hitter estimation.",3
"shopper analytics: customer activity recognition system using distributed rgb-d camera network. aim paper present integrated system consisted rgb-d camera software able monitor shoppers intelligent retail environments. propose innovative low cost smart system understand shoppers' behavior and, particular, interactions products shelves, aim develop automatic rgb-d technique video analysis. system cameras detects presence people univocally identifies them. depth frames, system detects interactions shoppers products shelf determines product picked product taken put back finally, contact products. system low cost easy install, experimental results demonstrated performances satisfactory also real environments.",3
"modular traffic sign recognition applied on-vehicle real-time visual detection american european speed limit signs. present new modular traffic signs recognition system, successfully applied american european speed limit signs. sign detection step based shape-detection (rectangles circles). enables work grayscale images, contrary european competitors, eases robustness illumination conditions (notably night operation). speed sign candidates classified (or rejected) segmenting potential digits inside (which rather original several advantages), applying neural digit recognition. global detection rate ~90% (standard) u.s. e.u. speed signs, misclassification rate <1%, validated false alarm >150 minutes video. system processes real-time ~20 frames/s standard high-end laptop.",3
"extending gossip algorithms distributed estimation u-statistics. efficient robust algorithms decentralized estimation networks essential many distributed systems. whereas distributed estimation sample mean statistics subject good deal attention, computation $u$-statistics, relying expensive averaging pairs observations, less investigated area. yet, data functionals essential describe global properties statistical population, important examples including area curve, empirical variance, gini mean difference within-cluster point scatter. paper proposes new synchronous asynchronous randomized gossip algorithms simultaneously propagate data across network maintain local estimates $u$-statistic interest. establish convergence rate bounds $o(1/t)$ $o(\log / t)$ synchronous asynchronous cases respectively, $t$ number iterations, explicit data network dependent terms. beyond favorable comparisons terms rate analysis, numerical experiments provide empirical evidence proposed algorithms surpasses previously introduced approach.",18
"supervised functional classification: theoretical remark comparisons. problem supervised classification (or discrimination) functional data considered, special interest popular k-nearest neighbors (k-nn) classifier. first, relying recent result cerou guyader (2006), prove consistency k-nn classifier functional data whose distribution belongs broad family gaussian processes triangular covariance functions. second, practical side, check behavior k-nn method compared functional classifiers. carried small simulation study analysis several real functional data sets. global ""uniform"" winner emerges comparisons, overall performance k-nn method, together sound intuitive motivation relative simplicity, suggests could represent reasonable benchmark classification problem functional data.",18
"parallel tracking verifying. intensively studied, visual object tracking witnessed great advances either speed (e.g., correlation filters) accuracy (e.g., deep features). real-time high accuracy tracking algorithms, however, remain scarce. paper study problem new perspective present novel parallel tracking verifying (ptav) framework, taking advantage ubiquity multi-thread techniques borrowing ideas success parallel tracking mapping visual slam. proposed ptav framework typically composed two components, (base) tracker verifier v, working parallel two separate threads. tracker aims provide super real-time tracking inference expected perform well time; contrast, verifier v validates tracking results corrects needed. key innovation that, v work every frame upon requests t; end, may adjust tracking according feedback v. collaboration, ptav enjoys high efficiency provided strong discriminative power v. meanwhile, adapt v object appearance changes time, maintain dynamic target template pool adaptive verification, resulting performance improvements. extensive experiments popular benchmarks including otb2015, tc128, uav20l vot2016, ptav achieves best tracking accuracy among real-time trackers, fact even outperforms many deep learning based algorithms. moreover, general framework, ptav flexible great potentials future improvement generalization.",3
"ensembles random sphere cover classifiers. propose evaluate alternative ensemble schemes new instance based learning classifier, randomised sphere cover (rsc) classifier. rsc fuses instances spheres, bases classification distance spheres rather distance instances. randomised nature rsc makes ideal use ensembles. propose two ensemble methods tailored rsc classifier; $\alpha \beta$rse, ensemble based instance resampling $\alpha$rsse, subspace ensemble. compare $\alpha \beta$rse $\alpha$rsse tree based ensembles set uci datasets demonstrates rsc ensembles perform significantly better ensembles, significantly worse others. demonstrate via case study six gene expression data sets $\alpha$rsse outperform subspace ensemble methods high dimensional data used conjunction attribute filter. finally, perform set bias/variance decomposition experiments analyse source improvement comparison base classifier.",3
"unbounded bayesian optimization via regularization. bayesian optimization recently emerged popular efficient tool global optimization hyperparameter tuning. currently, established bayesian optimization practice requires user-defined bounding box assumed contain optimizer. however, little known probed objective function, difficult prescribe bounds. work modify standard bayesian optimization framework principled way allow automatic resizing search space. introduce two alternative methods compare two common synthetic benchmarking test functions well tasks tuning stochastic gradient descent optimizer multi-layered perceptron convolutional neural network mnist.",18
"high efficiency compression object detection. image video compression traditionally tailored human vision. however, modern applications visual analytics surveillance rely computers seeing analyzing images (or instead of) humans. applications, important adjust compression computer vision. paper present bit allocation rate control strategy tailored object detection. using initial convolutional layers state-of-the-art object detector, create importance map guide bit allocation areas important object detection. proposed method enables bit rate savings 7% compared default hevc, equivalent object detection rate.",5
"functional gradient boosting based residual network perception. residual networks (resnets) become state-of-the-art models deep learning several theoretical studies devoted understanding resnet works well. one attractive viewpoint resnet optimizing risk functional space combining ensemble effective features. paper, adopt viewpoint construct new gradient boosting method, known powerful data analysis. so, formalize gradient boosting perspective resnet mathematically using notion functional gradients propose new method called resfgb classification tasks leveraging resnet perception. two types generalization guarantees provided optimization perspective: one margin bound expected risk bound sample-splitting technique. experimental results show superior performance proposed method state-of-the-art methods lightgbm.",18
"tweet2vec: learning tweet embeddings using character-level cnn-lstm encoder-decoder. present tweet2vec, novel method generating general-purpose vector representation tweets. model learns tweet embeddings using character-level cnn-lstm encoder-decoder. trained model 3 million, randomly selected english-language tweets. model evaluated using two methods: tweet semantic similarity tweet sentiment categorization, outperforming previous state-of-the-art tasks. evaluations demonstrate power tweet embeddings generated model various tweet categorization tasks. vector representations generated model generic, hence applied variety tasks. though model presented paper trained english-language tweets, method presented used learn tweet embeddings different languages.",3
"online optimization smoothed piecewise constant functions. study online optimization smoothed piecewise constant functions domain [0, 1). motivated problem adaptively picking parameters learning algorithms recently introduced framework gupta roughgarden (2016). majority machine learning literature focused lipschitz-continuous functions functions bounded gradients. 1 good reason---any learning algorithm suffers linear regret even piecewise constant functions chosen adversarially, arguably simplest non-lipschitz continuous functions. smoothed setting consider inspired seminal work spielman teng (2004) recent work gupta roughgarden---in setting, sequence functions may chosen adversary, however, uncertainty location discontinuities. give algorithms achieve sublinear regret full information bandit settings.",3
"foreign-language reviews: help hindrance?. number quality user reviews greatly affects consumer purchasing decisions. reviews languages increasing, still often case (especially non-english speakers) reviews person's first language. using online experiment, examine value potential purchasers receive interfaces showing additional reviews second language. results paint complicated picture positive negative reactions inclusion foreign-language reviews. roughly 26-28% subjects clicked see translations foreign-language content given opportunity, likely select product foreign-language reviews not.",3
"shamela: large-scale historical arabic corpus. arabic widely-spoken language rich long history spanning fourteen centuries. yet existing arabic corpora largely focus modern period lack sufficient diachronic information. develop large-scale, historical corpus arabic 1 billion words diverse periods time. clean corpus, process morphological analyzer, enhance detecting parallel passages automatically dating undated texts. demonstrate utility selected case-studies show application digital humanities.",3
"competition selective inhibitory synchrony. models cortical neuronal circuits commonly depend inhibitory feedback control gain, provide signal normalization, selectively amplify signals using winner-take-all (wta) dynamics. models generally assume excitatory inhibitory neurons able interact easily, axons dendrites co-localized small volume. however, quantitative neuroanatomical studies dimensions axonal dendritic trees neurons neocortex show co-localization assumption valid. paper describe simple modification wta circuit design permits effects distributed inhibitory neurons coupled synchronization, allows single wta distributed widely cortical space, well beyond arborization single inhibitory neuron, even across different cortical areas. prove non-linear contraction analysis, demonstrate simulation distributed wta sub-systems combined inhibitory synchrony inherently stable. show analytically synchronization substantially faster winner selection. circuit mechanism allows networks independent wtas fully partially compete other.",15
"time-series scenario forecasting. many applications require ability judge uncertainty time-series forecasts. uncertainty often specified point-wise error bars around mean median forecast. due temporal dependencies, method obscures information. would ideally way query posterior probability entire time-series given predictive variables, minimum, able draw samples distribution. use bayesian dictionary learning algorithm statistically generate ensemble forecasts. show algorithm performs well physics-based ensemble method temperature forecasts houston. conclude method shows promise scenario forecasting physics-based methods absent.",18
"throwing away data improves worst-class error imbalanced classification. class imbalances pervade classification problems, yet treatment differs theory practice. one hand, learning theory instructs us \emph{more data better}, sample size relates inversely average test error entire data distribution. hand, practitioners long developed plethora tricks improve performance learning machines imbalanced data. include data reweighting subsampling, synthetic construction additional samples minority classes, ensembling expensive one-versus architectures, tweaking classification losses thresholds. efforts minimize worst-class error, often associated minority group training data, finds additional motivation robustness, fairness, out-of-distribution literatures. take challenge developing learning theory able describe worst-class error classifiers linearly-separable data fitted either (i) full training set, (ii) subset majority class subsampled match size minority class. borrow tools extreme value theory show that, distributions certain tail properties, \emph{throwing away data majority class leads better worst-class error}.",18
"extremal optimization approach parallel resonance constrained capacitor placement problem. installation capacitors distribution networks one used procedure compensate reactive power generated loads and, consequently, reduce technical losses. so, problem consists identifying optimal placement sizing capacitors. problem known literature optimal capacitor placement problem. neverthless, depending location size capacitor, may become harmonic source, allowing capacitor enter resonance distribution network, causing several undesired side effects. work propose parsimonious method deal capacitor placement problem incorporates resonance constraints, ensuring every allocated capacitor act harmonic source. proposed algorithm based upon physical inspired metaheuristic known extremal optimization. results achieved showed proposal reached significant gains compared proposals attempt repair, post-optimization stage, already obtained solutions violate resonance constraints.",10
"pixels object sequences: recurrent semantic instance segmentation. present recurrent model semantic instance segmentation sequentially generates binary masks associated class probabilities every object image. proposed system trainable end-to-end input image sequence labeled masks and, compared methods relying object proposals, require post-processing steps output. study suitability recurrent model three different instance segmentation benchmarks, namely pascal voc 2012, cvppp plant leaf segmentation cityscapes. further, analyze object sorting patterns generated model observe learns follow consistent pattern, correlates activations learned encoder part network.",3
"artificial neural networks detection malaria rbcs. malaria one common diseases caused mosquitoes great public health problem worldwide. currently, malaria diagnosis standard technique microscopic examination stained blood film. propose use artificial neural networks (ann) diagnosis disease red blood cell. purpose features / parameters computed data obtained digital holographic images blood cells given input ann classifies cell infected one otherwise.",14
"detailed surface geometry albedo recovery rgb-d video natural illumination. paper present novel approach depth map enhancement rgb-d video sequence. basic idea exploit shading information color image. instead making assumption surface albedo controlled object motion lighting, use lighting variations introduced casual object movement. effectively calculating photometric stereo moving object natural illuminations. key technical challenge establish correspondences entire image set. therefore develop lighting insensitive robust pixel matching technique out-performs optical flow method presence lighting variations. addition present expectation-maximization framework recover surface normal albedo simultaneously, without regularization term. validated method synthetic real datasets show superior performance surface details recovery intrinsic decomposition.",3
"stochastic frank-wolfe methods nonconvex optimization. study frank-wolfe methods nonconvex stochastic finite-sum optimization problems. frank-wolfe methods (in convex case) gained tremendous recent interest machine learning optimization communities due projection-free property ability exploit structured constraints. however, understanding algorithms nonconvex setting fairly limited. paper, propose nonconvex stochastic frank-wolfe methods analyze convergence properties. objective functions decompose finite-sum, leverage ideas variance reduction techniques convex optimization obtain new variance reduced nonconvex frank-wolfe methods provably faster convergence classical frank-wolfe method. finally, show faster convergence rates variance reduced methods also translate improved convergence rates stochastic setting.",10
"plda two sources inter-session variability. speaker recognition scenarios find conversations recorded simultaneously multiple channels. case interviews nist sre dataset. take advantage that, propose modification plda model considers two different inter-session variability terms. first term tied recordings belonging conversation whereas second not. thus, former mainly intends capture variability due phonetic content conversation latter tries capture channel variability. document, derive equations model. model applied paper ""handling recordings acquired simultaneously multiple channels plda"" published interspeech 2013.",18
"idea dynamic semantics. ""natural languages programming languages minds."" take slogan seriously? so, how? answers found looking various ""dynamic"" treatments natural language developed last decade so, mostly response problems associated donkey anaphora? dynamic logic programs, meaning program binary relation set states abstract machine. relation meant model aspects effects execution program, particular input-output behavior. what, anything, dynamic aspects various proposed dynamic semantics natural languages supposed model? anything dynamic modeled? not, full about? shall try answer some, least, questions provide materials answers others.",1
"syntax aware lstm model chinese semantic role labeling. semantic role labeling (srl) task, comes utilizing parsing information, traditional methods recent recurrent neural network (rnn) based methods use feature engineering way. paper, propose syntax aware long short time memory(sa-lstm). structure sa-lstm modifies according dependency parsing information order model parsing information directly architecture engineering way instead feature engineering way. experimentally demonstrate sa-lstm gains improvement model architecture. furthermore, sa-lstm outperforms state-of-the-art cpb 1.0 significantly according student t-test ($p<0.05$).",3
"consensus measure rankings. ranking ordered sequence items, item higher ranking score preferred items lower ranking scores. many information systems, rankings widely used represent preferences set items candidates. consensus measure rankings problem evaluate degree rankings agree. consensus measure used evaluate rankings many information systems, quite often ground truth available evaluation. paper introduces novel approach consensus measure rankings using graph representation, vertices nodes items edges relationship items rankings. representation leads various algorithms consensus measure terms different aspects rankings, including number common patterns, number common patterns fixed length length longest common patterns. proposed measure adopted various types rankings, full rankings, partial rankings rankings ties. paper demonstrates proposed approaches used evaluate quality rank aggregation quality top-$k$ rankings google bing search engines.",3
"asynchronous decentralized parallel stochastic gradient descent. commonly used distributed machine learning systems either synchronous centralized asynchronous. synchronous algorithms like allreduce-sgd perform poorly heterogeneous environment, asynchronous algorithms using parameter server suffer 1) communication bottleneck parameter servers workers many, 2) significantly worse convergence traffic parameter server congested. design algorithm robust heterogeneous environment, communication efficient maintaining best-possible convergence rate? paper, propose asynchronous decentralized stochastic gradient decent algorithm (ad-psgd) satisfying expectations. theoretical analysis shows ad-psgd converges optimal $o(1/\sqrt{k})$ rate sgd linear speedup w.r.t. number workers. empirically, ad-psgd outperforms best decentralized parallel sgd (d-psgd), asynchronous parallel sgd (a-psgd), standard data parallel sgd (allreduce-sgd), often orders magnitude heterogeneous environment. training resnet-50 imagenet 128 gpus, ad-psgd converges (w.r.t epochs) similarly allreduce-sgd, epoch 4-8x faster synchronous counterparts network-sharing hpc environment. best knowledge, ad-psgd first asynchronous algorithm achieves similar epoch-wise convergence rate allreduce-sgd, 100-gpu scale.",10
"analysis implementation asynchronous optimization algorithm parameter server. paper presents asynchronous incremental aggregated gradient algorithm implementation parameter server framework solving regularized optimization problems. algorithm handle general convex (possibly non-smooth) regularizers general convex constraints. empirical data loss strongly convex, establish linear convergence rate, give explicit expressions step-size choices guarantee convergence optimum, bound associated convergence factors. expressions explicit dependence degree asynchrony recover classical results synchronous operation. simulations implementations commercial compute clouds validate findings.",10
"hadamard tail initial data light cone. field perturbations curved background spacetime generally propagate speed light also smaller velocities. so-called $hadamard\,tail$ contribution wave propagation relevant various settings, classical self-force calculations communication quantum particle detectors. one method calculating tail contribution integrating homogeneous wave equation using characteristic initial data light cone. however, best knowledge, method never implemented except flat conformally-flat spacetimes, null geodesics emanating point cross. work, implement method black hole toy model pleba\'nski-hacyan spacetime, $\mathbb{m}_2\times\mathbb{s}^2$. obtain new results spacetime calculating hadamard tail scalar field everywhere defined (namely, maximal normal neighbourhood arbitrary point) investigate varies various values coupling constant. serves proof-of-concept characteristic initial data method spacetimes null geodesics emanating point $do$ cross.",6
"multiscale markov decision problems: compression, solution, transfer learning. many problems sequential decision making stochastic control often natural multiscale structure: sub-tasks assembled together accomplish complex goals. systematically inferring leveraging hierarchical structure, particularly beyond single level abstraction, remained longstanding challenge. describe fast multiscale procedure repeatedly compressing, homogenizing, markov decision processes (mdps), wherein hierarchy sub-problems different scales automatically determined. coarsened mdps independent, deterministic mdps, may solved using existing algorithms. multiscale representation delivered procedure decouples sub-tasks lead substantial improvements convergence rates locally within sub-problems globally across sub-problems, yielding significant computational savings. second fundamental aspect work multiscale decompositions yield new transfer opportunities across different problems, solutions sub-tasks different levels hierarchy may amenable transfer new problems. localized transfer policies potential operators arbitrary scales emphasized. finally, demonstrate compression transfer collection illustrative domains, including examples involving discrete continuous statespaces.",3
"reliable decision support using counterfactual models. decision-makers faced challenge estimating likely happen take action. instance, choose treat patient, likely die? practitioners commonly use supervised learning algorithms fit predictive models help decision-makers reason likely future outcomes, show approach unreliable, sometimes even dangerous. key issue supervised learning algorithms highly sensitive policy used choose actions training data, causes model capture relationships generalize. propose using different learning objective predicts counterfactuals instead predicting outcomes existing action policy supervised learning. support decision-making temporal settings, introduce counterfactual gaussian process (cgp) predict counterfactual future progression continuous-time trajectories sequences future actions. demonstrate benefits cgp two important decision-support tasks: risk prediction ""what if?"" reasoning individualized treatment planning.",18
"neur2sp: neural two-stage stochastic programming. stochastic programming powerful modeling framework decision-making uncertainty. work, tackle two-stage stochastic programs (2sps), widely applied studied class stochastic programming models. solving 2sps exactly requires evaluation expected value function computationally intractable. additionally, mixed-integer linear program (mip) nonlinear program (nlp) second stage aggravates problem difficulty. cases, solving prohibitively expensive even specialized algorithms exploit problem structure employed. finding high-quality (first-stage) solutions -- without leveraging problem structure -- crucial settings. develop neur2sp, new method approximates expected value function via neural network obtain surrogate model solved efficiently traditional extensive formulation approach. moreover, neur2sp makes assumptions problem structure, particular second-stage problem, implemented using off-the-shelf solver open-source libraries. extensive computational experiments benchmark 2sp datasets four problem classes different structures (containing mip nlp second-stage problems) show efficiency (time) efficacy (solution quality) neur2sp. specifically, proposed method takes less 1.66 seconds across problems, achieving high-quality solutions even number scenarios increases, ideal property difficult traditional 2sp solution techniques. namely, generic baseline method typically requires minutes hours find solutions comparable quality.",10
"houdini: fooling deep structured prediction models. generating adversarial examples critical step evaluating improving robustness learning machines. far, existing methods work classification designed alter true performance measure problem hand. introduce novel flexible approach named houdini generating adversarial examples specifically tailored final performance measure task considered, combinatorial non-decomposable. successfully apply houdini range applications speech recognition, pose estimation semantic segmentation. cases, attacks based houdini achieve higher success rate based traditional surrogates used train models using less perceptible adversarial perturbation.",18
"fused dnn: deep neural network fusion approach fast robust pedestrian detection. propose deep neural network fusion architecture fast robust pedestrian detection. proposed network fusion architecture allows parallel processing multiple networks speed. single shot deep convolutional network trained object detector generate possible pedestrian candidates different sizes occlusions. network outputs large variety pedestrian candidates cover majority ground-truth pedestrians also introducing large number false positives. next, multiple deep neural networks used parallel refinement pedestrian candidates. introduce soft-rejection based network fusion method fuse soft metrics networks together generate final confidence scores. method performs better existing state-of-the-arts, especially detecting small-size occluded pedestrians. furthermore, propose method integrating pixel-wise semantic segmentation network network fusion architecture reinforcement pedestrian detector. approach outperforms state-of-the-art methods protocols caltech pedestrian dataset, significant boosts several protocols. also faster methods.",3
"optimal convergence rate feed forward neural networks using hjb equation. control theoretic approach presented paper batch instantaneous updates weights feed-forward neural networks. popular hamilton-jacobi-bellman (hjb) equation used generate optimal weight update law. remarkable contribution paper closed form solutions optimal cost weight update achieved feed-forward network using hjb equation simple yet elegant manner. proposed approach compared existing best performing learning algorithms. found expected proposed approach faster convergence terms computational time. benchmark test data 8-bit parity, breast cancer credit approval, well 2d gabor function used validate claims. paper also discusses issues related global optimization. limitations popular deterministic weight update laws critiqued possibility global optimization using hjb formulation discussed. hoped proposed algorithm bring lot interest researchers working developing fast learning algorithms global optimization.",3
"evolutionary multiobjective optimization multi-location transshipment problem. consider multi-location inventory system inventory choices location centrally coordinated. lateral transshipments allowed recourse actions within echelon inventory system reduce costs improve service level. however, transshipment process usually causes undesirable lead times. paper, propose multiobjective model multi-location transshipment problem addresses optimizing three conflicting objectives: (1) minimizing aggregate expected cost, (2) maximizing expected fill rate, (3) minimizing expected transshipment lead times. apply evolutionary multiobjective optimization approach using strength pareto evolutionary algorithm (spea2), approximate optimal pareto front. simulation wide choice model parameters shows different trades-off conflicting objectives.",3
"calibration phone likelihoods automatic speech recognition. paper study probabilistic properties posteriors speech recognition system uses deep neural network (dnn) acoustic modeling. reducing kaldi's dnn shared pdf-id posteriors phone likelihoods, using test set forced alignments evaluate using calibration sensitive metric. individual frame posteriors principle well-calibrated, dnn trained using cross entropy objective function, proper scoring rule. entire phones assessed, observe best average log likelihoods duration phone. scaling average log likelihoods logarithm duration slightly improves calibration, improvement retained tested independent test data.",18
"online learning rested restless bandits. paper study online learning problem involving rested restless multiarmed bandits multiple plays. system consists single player/user set k finite-state discrete-time markov chains (arms) unknown state spaces statistics. time step player play arms. objective user decide step k arms play sequence trials maximize long term reward. restless multiarmed bandit particularly relevant application opportunistic spectrum access (osa), (secondary) user access set k channels, time-varying condition result random fading and/or certain primary users' activities.",10
"video frame interpolation via adaptive separable convolution. standard video frame interpolation methods first estimate optical flow input frames synthesize intermediate frame guided motion. recent approaches merge two steps single convolution process convolving input frames spatially adaptive kernels account motion re-sampling simultaneously. methods require large kernels handle large motion, limits number pixels whose kernels estimated due large memory demand. address problem, paper formulates frame interpolation local separable convolution input frames using pairs 1d kernels. compared regular 2d kernels, 1d kernels require significantly fewer parameters estimated. method develops deep fully convolutional neural network takes two input frames estimates pairs 1d kernels pixels simultaneously. since method able estimate kernels synthesizes whole video frame once, allows incorporation perceptual loss train neural network produce visually pleasing frames. deep neural network trained end-to-end using widely available video data without human annotation. qualitative quantitative experiments show method provides practical solution high-quality video frame interpolation.",3
"learning algebraic invariances, invariant kernel trick. solving data analysis problems important integrate prior knowledge and/or structural invariances. paper contributes novel framework incorporating algebraic invariance structure kernels. particular, show algebraic properties sign symmetries data, phase independence, scaling etc. included easily essentially performing kernel trick twice. demonstrate usefulness theory simulations selected applications sign-invariant spectral clustering underdetermined ica.",18
"composite gaussian processes: scalable computation performance analysis. gaussian process (gp) models provide powerful tool prediction computationally prohibitive using large data sets. scenarios, one resort approximate methods. derive approximation based composite likelihood approach using general belief updating framework, leads recursive computation predictor well learning hyper-parameters. provide analysis derived composite gp model predictive information-theoretic terms. finally, evaluate approximation synthetic data real-world application.",18
"cross-validated variable selection tree-based methods improves predictive performance. recursive partitioning approaches producing tree-like models long standing staple predictive modeling, last decade mostly ``sub-learners'' within state art ensemble methods like boosting random forest. however, fundamental flaw partitioning (or splitting) rule commonly used tree building methods precludes treating different types variables equally. clearly manifests methods' inability properly utilize categorical variables large number categories, ubiquitous new age big data. variables often informative, current tree methods essentially leave us choice either using them, exposing models severe overfitting. propose conceptual framework splitting using leave-one-out (loo) cross validation selecting splitting variable, performing regular split (in case, following cart's approach) selected variable. important consequence approach categorical variables many categories safely used tree building chosen contribute predictive power. demonstrate extensive simulation real data analysis novel splitting approach significantly improves performance single tree models ensemble methods utilize trees. importantly, design algorithm loo splitting variable selection reasonable assumptions increase overall computational complexity compared cart two-class classification. regression tasks, approach carries increased computational burden, replacing o(log(n)) factor cart splitting rule search o(n) term.",18
"multivariate gaussian student$-t$ process regression multi-output prediction. gaussian process vector-valued function model shown useful method multi-output prediction. existing method model re-formulate matrix-variate gaussian distribution multivariate normal distribution. although effective many cases, re-formulation always workable difficult extend matrix-variate distributions transformed related multivariate distributions, case matrix-variate student$-t$ distribution. paper, propose new derivation multivariate gaussian process regression (mv-gpr), model settings, derivations computations directly performed matrix form, rather vectorizing matrices done existing methods. furthermore, introduce multivariate student$-t$ process derive new method, multivariate student$-t$ process regression (mv-tpr) multi-output prediction. mv-gpr mv-tpr closed-form expressions marginal likelihoods predictive distributions. usefulness proposed methods illustrated several simulated examples. particular, verify empirically mv-tpr superiority datasets considered, including air quality prediction bike rent prediction. last, proposed methods shown produce profitable investment strategies stock markets.",18
"cccp algorithms minimize bethe free energy 3-sat problem. k-sat problem prototypical constraint satisfaction problem. many algorithms study k-sat problem, bp algorithm famous one them. bp algorithm converge $\alpha$(constraint density)is bigger threshold value. paper use cccp (concave convex procedure) algorithm study 3-sat problem get better results bp algorithm cccp algorithm still converges bp algorithm converge. work almost builds recent results yuille \cite{yuille2002} apply cccp algorithm bethe kikuchi free energies obtained two algorithms 2d 3d spin glasses. implementation cccp algorithm 3-sat problem different implementation different views cccp algorithm's properties. difference maybe cccp algorithm different properties implementation process different problem others related cccp algorithm itself. work indicates cccp algorithm learning inference applications.",14
"systematic analysis single heavy baryons $λ_{q}$, $σ_{q}$ $ω_{q}$. motivated great progresses experiments searching heavy baryons, systematically analyze mass spectra strong decay behaviors single heavy baryons $\lambda_{q}$, $\sigma_{q}$ $\omega_{q}$. calculations mass spectra carried frame work godfrey-isgur (gi) relativized quark model, baryon regarded three-body system quarks. results show mass single heavy baryon $\lambda$-mode lower $\rho$-mode $\lambda$-$\rho$ mixing mode, indicates lowest state dominated $\lambda$-mode. basing research, systematically calculate mass spectra baryons $\lambda$ excited mode. predicated mass spectra, regge trajectories ($j$,$m^{2}$) plane constructed, slopes, intercepts regge trajectories obtained linear fitting. found available experimental data well reproduced model predictions fit nicely constructed regge trajectories. order make confirmation assignments observed baryons, provide valuable infirmations searching model predicted states, perform systematic study strong decay behaviors $s$-wave, $p$-wave $d$-wave baryons $^{3}p_{0}$ model.",8
"universal algorithms, mathematics semirings parallel computations. survey paper applications mathematics semirings numerical analysis computing. concepts universal algorithm generic program discussed. relations concepts mathematics semirings examined. brief introduction mathematics semirings (including idempotent tropical mathematics) presented. concrete applications optimization problems, idempotent linear algebra interval analysis indicated. known nonlinear problems (and especially optimization problems) become linear appropriate semirings idempotent addition (the so-called idempotent superposition principle). linearity semirings convenient parallel computations.",10
"robust multimodal graph matching: sparse coding meets graph matching. graph matching challenging problem important applications wide range fields, image video analysis biological biomedical problems. propose robust graph matching algorithm inspired sparsity-related techniques. cast problem, resembling group collaborative sparsity formulations, non-smooth convex optimization problem efficiently solved using augmented lagrangian techniques. method deal weighted unweighted graphs, well multimodal data, different graphs represent different types data. proposed approach also naturally integrated collaborative graph inference techniques, solving general network inference problems observed variables, possibly coming different modalities, correspondence. algorithm tested compared state-of-the-art graph matching techniques synthetic real graphs. also present results multimodal graphs applications collaborative inference brain connectivity alignment-free functional magnetic resonance imaging (fmri) data. code publicly available.",10
"effect tuned parameters lsa mcq answering model. paper presents current state work progress, whose objective better understand effects factors significantly influence performance latent semantic analysis (lsa). difficult task, consists answering (french) biology multiple choice questions, used test semantic properties truncated singular space study relative influence main parameters. dedicated software designed fine tune lsa semantic space multiple choice questions task. optimal parameters, performances simple model quite surprisingly equal superior 7th 8th grades students. indicates semantic spaces quite good despite low dimensions small sizes training data sets. besides, present original entropy global weighting answers' terms question multiple choice questions necessary achieve model's success.",3
"stochastic training graph convolutional networks variance reduction. graph convolutional networks (gcns) powerful deep neural networks graph-structured data. however, gcn computes representation node recursively neighbors, making receptive field size grow exponentially number layers. previous attempts reducing receptive field size subsampling neighbors convergence guarantee, receptive field size per node still order hundreds. paper, develop control variate based algorithms allow sampling arbitrarily small neighbor size. furthermore, prove new theoretical guarantee algorithms converge local optimum gcn. empirical results show algorithms enjoy similar convergence exact algorithm using two neighbors per node. runtime algorithms large reddit dataset one seventh previous neighbor sampling algorithms.",18
"note spice method. article, analyze spice method developed [1], establish connections standard sparse estimation methods lasso lad-lasso. result positions spice computationally efficient technique calculation lasso-type estimators. conversely, connection useful establishing asymptotic properties spice several problem scenarios suggesting suitable modifications cases naive version spice would work.",18
"feature selection based pca pso multimodal medical image fusion using dtcwt. multimodal medical image fusion helps increase efficiency medical diagnosis. paper presents multimodal medical image fusion selecting relevant features using principle component analysis (pca) particle swarm optimization techniques (pso). dtcwt used decomposition images low high frequency coefficients. fusion rules combination minimum, maximum simple averaging applied approximate detailed coefficients. fused image reconstructed inverse dtcwt. performance metrics evaluated shows dtcwt-pca performs better dtcwt-pso terms structural similarity index measure (ssim) cross correlation (cc). computation time feature vector size reduced dtcwt-pca compared dtcwt-pso feature selection proves robustness storage capacity.",3
"deep learning identifying metastatic breast cancer. international symposium biomedical imaging (isbi) held grand challenge evaluate computational systems automated detection metastatic breast cancer whole slide images sentinel lymph node biopsies. team competitions grand challenge, obtaining area receiver operating curve (auc) 0.925 task whole slide image classification score 0.7051 tumor localization task. pathologist independently reviewed images, obtaining whole slide image classification auc 0.966 tumor localization score 0.733. combining deep learning system's predictions human pathologist's diagnoses increased pathologist's auc 0.995, representing approximately 85 percent reduction human error rate. results demonstrate power using deep learning produce significant improvements accuracy pathological diagnoses.",15
"nature-inspired mateheuristic algorithms: success new challenges. despite increasing popularity metaheuristics, many crucially important questions remain unanswered. two important issues: theoretical framework gap theory applications. moment, practice metaheuristics like heuristic itself, extent, trial error. mathematical analysis lags far behind, apart few, limited, studies convergence analysis stability, theoretical framework analyzing metaheuristic algorithms. believe mathematical statistical methods using markov chains dynamical systems useful future work. doubt theoretical progress provide potentially huge insightful meteheuristic algorithms.",10
"human expert fusion image classification. image classification, merging opinion several human experts important different tasks evaluation training. indeed, ground truth rarely known scene imaging. propose different models order fuse informations given two experts. considered unit classification, small tile image, contain one kind considered classes given experts. second problem take account, amount certainty expert pixel tile. order solve problems define five models context dempster-shafer theory context dezert-smarandache theory study possible decisions models.",3
"high cadence tess ground-based data sn 2019esa, less energetic sibling sn 2006gy. present photometric spectroscopic observations nearby ($d\approx28$ mpc) interacting supernova (sn) 2019esa, discovered within hours explosion serendipitously observed transiting exoplanet survey satellite (tess). early, high cadence light curves tess dlt40 survey tightly constrain time explosion, show 30 day rise maximum light followed near constant linear decline luminosity. optical spectroscopy first 40 days revealed highly reddened object narrow balmer emission lines seen type iin supernovae. slow rise maximum optical lightcurve combined lack broad h$\alpha$ emission suggest presence optically thick close circumstellar material (csm) quickly decelerated supernova ejecta. csm likely created massive star progenitor $\dot{m}$ $\sim$ 0.3 m$_{\odot}$ yr$^{-1}$ lost previous eruptive episode 3--4 years eruption, similar giant eruptions luminous blue variable stars. late times, strong intermediate-width ca ii, fe i, fe ii lines seen optical spectra, identical seen superluminous interacting sn 2006gy. strong csm interaction masks underlying explosion mechanism sn 2019esa, combination luminosity, strength h$\alpha$ lines, mass loss rate progenitor point core collapse origin.",0
"multiple attractor cellular automata (maca) addressing major problems bioinformatics. ca grown potential classifier addressing major problems bioinformatics. lot bioinformatics problems like predicting protein coding region, finding promoter region, predicting structure protein many problems bioinformatics addressed cellular automata. even though prediction techniques addressing problems, approximate accuracy level less. automated procedure proposed maca (multiple attractor cellular automata) address problems. genetic algorithm also used find rules good fitness values. extensive experiments conducted reporting accuracy proposed tool. average accuracy maca tested encode, bg570, hmr195, fickett tongue, asp67 datasets 78%.",3
"band degeneration evolution nonlinear triatomic chain superlattices. nonlinear superlattices exhibit unique features allowing wave manipulations. despite increasing attention received, underlying physical mechanisms evolution process band structures bandgaps strongly nonlinear superlattices remain unclear. establish examine strongly nonlinear superlattice models (three triatomic models) show evolution process typical nonlinear band structures based analytical numerical approaches. find strongly nonlinear superlattices present particular band degeneration bifurcation, accompanied vibration mode transfer unit cells. evolution processes physical mechanisms band degeneration different models clarified consideration mode transfer. observed degeneration may occur shifting, bifurcating, shortening, merging disappearing dispersion curves, depending arrangement coupled nonlinear elements. meanwhile, dimension unit cell reduces, alongside changes frequency range mechanisms (bragg local resonance) bandgaps. findings answer foundamental questions peritinent study nonlinear periodic structures, nonlinear crystals nonlinear metamaterials, interest broad community physics",2
"reproduction images gamut mapping creation new test charts prepress process. advent digital images problem keeping picture visualization uniformity arises printing scanning device color chart. so, universal color profiles made icc bring uniformity various types devices. keeping color profile mind various new color charts created calibrated help standard it8 test charts available market. main objective color reproduction produce identical picture device output. principles gamut mapping designed",3
"measurement antiproton production antihyperon decays phe collisions $\sqrt{s_{nn}}=110$ gev. interpretation cosmic antiproton flux measurements space-borne experiments currently limited knowledge antiproton production cross-section collisions primary cosmic rays interstellar medium. using collisions protons energy 6.5 tev incident helium nuclei rest proximity interaction region lhcb experiment, ratio antiprotons originating antihyperon decays prompt production measured antiproton momenta 12 110 gev/c. dominant antihyperon contribution, namely $\bar{\lambda} \to \bar{p} \pi^+$ decays promptly produced $\bar{\lambda}$ particles, also exclusively measured. results complement measurement prompt antiproton production obtained data sample. energy scale measurement, antihyperon contributions antiproton production observed significantly larger predictions commonly used hadronic production models.",7
araguaia medical vision lab isic 2017 skin lesion classification challenge. paper describes participation araguaia medical vision lab international skin imaging collaboration 2017 skin lesion challenge. describe use deep convolutional neural networks attempt classify images melanoma seborrheic keratosis lesions. use finetuned googlenet alexnet attained results 0.950 0.846 auc seborrheic keratosis melanoma respectively.,3
"finding significant combinations continuous features. present efficient feature selection method find multiplicative combinations continuous features statistically significantly associated class variable, rigorously correcting multiple testing. key overcome combinatorial explosion number candidates derive lower bound $p$-value feature combination, enables us massively prune combinations never significant gain statistical power. problem addressed binary features past, present first solution continuous features. experiments, novel approach detects true feature combinations higher precision recall competing methods require prior binarization data.",18
"learning similarity preserving representations neural similarity encoders. many dimensionality reduction manifold learning algorithms optimize retaining pairwise similarities, distances, local neighborhoods data points. spectral methods like kernel pca (kpca) isomap achieve computing singular value decomposition (svd) similarity matrix obtain low dimensional representation original data. however, computationally expensive lot training examples available and, additionally, representations new (out-of-sample) data points created similarities original training examples computed. introduce similarity encoders (simec), learn similarity preserving representations using feed-forward neural network map data embedding space original similarities approximated linearly. model optimizes objective kpca process learns linear non-linear embedding function (in form tuned neural network), representations novel data points computed - even original pairwise similarities training set generated unknown process human ratings. creating embeddings image text datasets, demonstrate simec can, one hand, reach solution spectral methods, and, hand, obtain meaningful embeddings similarities based human labels.",18
"learning structural weight uncertainty sequential decision-making. learning probability distributions weights neural networks (nns) recently proven beneficial many applications. bayesian methods, stein variational gradient descent (svgd), offer elegant framework reason nn model uncertainty. however, assuming independent gaussian priors individual nn weights (as often applied), svgd impose prior knowledge often structural information (dependence) among weights. propose efficient posterior learning structural weight uncertainty, within svgd framework, employing matrix variate gaussian priors nn parameters. investigate learned structural uncertainty sequential decision-making problems, including contextual bandits reinforcement learning. experiments several synthetic real datasets indicate superiority model, compared state-of-the-art methods.",18
"cluster forests. inspiration random forests (rf) context classification, new clustering ensemble method---cluster forests (cf) proposed. geometrically, cf randomly probes high-dimensional data cloud obtain ""good local clusterings"" aggregates via spectral clustering obtain cluster assignments whole dataset. search good local clusterings guided cluster quality measure kappa. cf progressively improves local clustering fashion resembles tree growth rf. empirical studies several real-world datasets two different performance metrics show cf compares favorably competitors. theoretical analysis reveals kappa measure makes possible grow local clustering desirable way---it ""noise-resistant"". closed-form expression obtained mis-clustering rate spectral clustering perturbation model, yields new insights aspects spectral clustering.",18
"xml framework concept description knowledge representation. xml framework concept description given, based upon fact tree structure xml implies logical structure concepts defined attributional calculus. especially, attribute-value representation implementable xml framework. since attribute-value representation important way represent knowledge ai, framework offers simpler way powerful rdf technology.",3
"learning disambiguate asking discriminative questions. ability ask questions powerful tool gather information order learn world resolve ambiguities. paper, explore novel problem generating discriminative questions help disambiguate visual instances. work seen complement new extension rich research studies image captioning question answering. introduce first large-scale dataset 10,000 carefully annotated images-question tuples facilitate benchmarking. particular, tuple consists pair images 4.6 discriminative questions (as positive samples) 5.9 non-discriminative questions (as negative samples) average. addition, present effective method visual discriminative question generation. method trained weakly supervised manner without discriminative images-question tuples existing visual question answering datasets. promising results shown representative baselines quantitative evaluations user studies.",3
"parameter compression recurrent neural networks degradation short-term memory. significant computational costs deploying neural networks large-scale resource constrained environments, data centers mobile devices, spurred interest model compression, achieve reduction arithmetic operations storage memory. several techniques proposed reducing compressing parameters feed-forward convolutional neural networks, less understood effect parameter compression recurrent neural networks (rnn). particular, extent recurrent parameters compressed impact short-term memory performance, well understood. paper, study effect complexity reduction, singular value decomposition rank reduction, rnn minimal gated recurrent unit (mgru) networks several tasks. show considerable rank reduction possible compressing recurrent weights, even without fine tuning. furthermore, propose perturbation model effect general perturbations, compression, recurrent parameters rnns. model tested noiseless memorization experiment elucidates short-term memory performance. way, demonstrate effect compression recurrent parameters dependent degree temporal coherence present data task. work guide on-the-fly rnn compression novel environments tasks, provides insight applying rnn compression low-power devices, hearing aids.",3
"mall bundles flat connections hopf manifolds. mall bundle hopf manifold h holomorphic vector bundle whose pullback universal cover h trivial. define resonant non-resonant mall bundles, generalizing notion resonance ode, prove non-resonant mall bundle always admits flat holomorphic connection. use observation prove version poincare-dulac linearization theorem, showing non-resonant invertible holomorphic contraction complex space linear appropriate holomorphic coordinates. define notion resonance hopf manifolds, show non-resonant hopf manifolds linear; previously, result obtained kodaira using poincare-dulac theorem.",10
"neural semantic parsing character-based translation: experiments abstract meaning representations. evaluate character-level translation method neural semantic parsing large corpus sentences annotated abstract meaning representations (amrs). using sequence-to-sequence model, trivial preprocessing postprocessing amrs, obtain baseline accuracy 53.1 (f-score amr-triples). examine five different approaches improve baseline result: (i) reordering amr branches match word order input sentence increases performance 58.3; (ii) adding part-of-speech tags (automatically produced) input shows improvement well (57.2); (iii) introduction super characters (conflating frequent sequences characters single character), reaching 57.4; (iv) optimizing training process using pre-training averaging set models increases performance 58.7; (v) adding silver-standard training data obtained off-the-shelf parser yields biggest improvement, resulting f-score 64.0. combining five techniques leads f-score 71.0 holdout data, state-of-the-art amr parsing. remarkable relative simplicity approach.",3
"physician advisory system chronic heart failure management based knowledge patterns. management chronic diseases heart failure, diabetes, chronic obstructive pulmonary disease (copd) major problem health care. standard approach medical community devised manage widely prevalent chronic diseases chronic heart failure (chf) committee experts develop guidelines physicians follow. guidelines typically consist series complex rules make recommendations based patient's information. due complexity, often guidelines either ignored complied all, result poor medical practices. even clear whether humanly possible follow guidelines due length complexity. case chf management, guidelines run nearly 80 pages. paper describe physician-advisory system chf management codes entire set clinical practice guidelines chf using answer set programming. approach based developing reasoning templates (that call knowledge patterns) using patterns systemically code clinical guidelines chf asp rules. use knowledge patterns greatly facilitates development system. given patient's medical information, system generates recommendation treatment human physician would, using guidelines. system work even presence incomplete information. work makes two contributions: (i) shows highly complex guidelines successfully coded asp rules, (ii) develops series knowledge patterns facilitate coding knowledge expressed natural language used application domains. paper consideration acceptance tplp.",3
"topological tensor representations gl(v) space v countable dimension. lie algebra $gl(v)$ lie algebra endomorphisms countable-dimensional complex vector space $v$. define tensor category topological representations lie algebra $gl(v)$, $v$, dual adjoint representation $gl(v)$ objects category. makes analogue category finite-dimensional modules finite-dimensional lie algebra $gl(n)$. main result category antiequivalent symmetric monoidal category category tensor representations lie algebra finitary infinite matrices.",10
"back basics: unsupervised learning optical flow via brightness constancy motion smoothness. recently, convolutional networks (convnets) proven useful predicting optical flow. much success predicated availability large datasets require expensive involved data acquisition laborious la- beling. bypass challenges, propose unsuper- vised approach (i.e., without leveraging groundtruth flow) train convnet end-to-end predicting optical flow be- tween two images. use loss function combines data term measures photometric constancy time spatial term models expected variation flow across image. together losses form proxy measure losses based groundtruth flow. empiri- cally, show strong convnet baseline trained proposed unsupervised approach outperforms network trained supervision kitti dataset.",3
"multisensor images fusion based feature-level. now, highest relevance remote sensing data processing analysis techniques pixel level image fusion. so, paper attempts undertake study feature-level based image fusion. purpose, feature based fusion techniques, usually based empirical heuristic rules, employed. hence, paper consider feature extraction (fe) fusion. aims finding transformation original space would produce new features, preserve improve much possible. study introduces three different types image fusion techniques including principal component analysis based feature fusion (pca), segment fusion (sf) edge fusion (ef). paper also devotes concentrate analytical techniques evaluating quality image fusion (f) using various methods including (sd), (en), (cc), (snr), (nrmse) (di) estimate quality degree information improvement fused image quantitatively.",3
"sex gibbs sampling: probability model evolution. show evolutionary computation implemented standard markov-chain monte-carlo (mcmc) sampling. care, `genetic algorithms' constructed reversible markov chains satisfy detailed balance; follows stationary distribution populations gibbs distribution simple factorised form. standard popular nonparametric probability models, exhibit gibbs-sampling procedures plausible genetic algorithms. mutation-selection equilibrium, population genomes analogous sample bayesian posterior, genomes analogous latent variables. suggest general, tractable, insightful formulation evolutionary computation terms standard machine learning concepts techniques. addition, show evolutionary processes selection acts differences fecundity reversible, also possible construct reversible evolutionary models child produced two parents.",15
"freeze-thaw bayesian optimization. paper develop dynamic form bayesian optimization machine learning models goal rapidly finding good hyperparameter settings. method uses partial information gained training machine learning model order decide whether pause training start new model, resume training previously-considered model. specifically tailor method machine learning problems developing novel positive-definite covariance kernel capture variety training curves. furthermore, develop gaussian process prior scales gracefully additional temporal observations. finally, provide information-theoretic framework automate decision process. experiments several common machine learning models show approach extremely effective practice.",18
"interpretation prediction models using input gradient. state art machine learning algorithms highly optimized provide optimal prediction possible, naturally resulting complex models. models often outperform simpler interpretable models order magnitudes, terms understanding way model functions, often facing ""black box"". paper suggest simple method interpret behavior predictive model, regression classification. given particular model, information required interpret obtained studying partial derivatives model respect input. exemplify insight interpreting convolutional multi-layer neural networks field natural language processing.",18
"interaction entropy-based discretization sample size: empirical study. empirical investigation interaction sample size discretization - case entropy-based method caim (class-attribute interdependence maximization) - undertaken evaluate impact potential bias introduced data mining performance metrics due variation sample size impacts discretization process. particular interest effect discretizing within cross-validation folds averse outside discretization folds. previous publications suggested discretizing externally bias performance results; however, thorough review literature found empirical evidence support assertion. investigation involved construction 117,000 models seven distinct datasets uci (university california-irvine) machine learning library multiple modeling methods across variety configurations sample size discretization, unique ""setup"" independently replicated ten times. analysis revealed significant optimistic bias sample sizes decreased discretization employed. study also revealed may relationship interaction produces bias numbers types predictor attributes, extending ""curse dimensionality"" concept feature selection discretization realm. directions exploration laid out, well general guidelines proper application discretization light results.",18
"learning phrase embeddings paraphrases grus. learning phrase representations widely explored many natural language processing (nlp) tasks (e.g., sentiment analysis, machine translation) shown promising improvements. previous studies either learn non-compositional phrase representations general word embedding learning techniques learn compositional phrase representations based syntactic structures, either require huge amounts human annotations cannot easily generalized phrases. work, propose take advantage large-scaled paraphrase database present pair-wise gated recurrent units (pairwise-gru) framework generate compositional phrase representations. framework re-used generate representations phrases. experimental results show framework achieves state-of-the-art results several phrase similarity tasks.",3
"interactive visual data exploration subjective feedback: information-theoretic approach. visual exploration high-dimensional real-valued datasets fundamental task exploratory data analysis (eda). existing methods use predefined criteria choose representation data. lack methods (i) elicit user learned data (ii) show patterns know yet. construct theoretical model identified patterns input knowledge system. knowledge syntax intuitive, ""this set points forms cluster"", requires knowledge maths. background knowledge used find maximum entropy distribution data, system provides user data projections data maximum entropy distribution differ most, hence showing user aspects data maximally informative given user's current knowledge. provide open source eda system tailored interactive visualizations demonstrate concepts. study performance system present use cases synthetic real data. find model prototype system allow user learn information efficiently various data sources system works sufficiently fast practice. conclude information theoretic approach exploratory data analysis patterns observed user formalized constraints provides principled, intuitive, efficient basis constructing eda system.",18
"sparse linear isotonic models. machine learning data mining, linear models widely used model response parametric linear functions predictors. relax stringent assumptions made parametric linear models, additive models consider response summation unknown transformations applied predictors; particular, additive isotonic models (aims) assume unknown transformations monotone. paper, introduce sparse linear isotonic models (slims) highdimensional problems hybridizing ideas parametric sparse linear models aims, enjoy appealing advantages both. high-dimensional setting, two-step algorithm proposed estimating sparse parameters well monotone functions predictors. mild statistical assumptions, show algorithm accurately estimate parameters. promising preliminary experiments presented support theoretical results.",18
"network-guided biomarker discovery. identifying measurable genetic indicators (or biomarkers) specific condition biological system key element precision medicine. indeed allows tailor diagnostic, prognostic treatment choice individual characteristics patient. machine learning terms, biomarker discovery framed feature selection problem whole-genome data sets. however, classical feature selection methods usually underpowered process data sets, contain orders magnitude features samples. addressed making assumption genetic features linked biological network likely work jointly towards explaining phenotype interest. review three families methods feature selection integrate prior knowledge form networks.",18
"saccadic eye movements generalized pareto distribution. describe statistical analysis eye tracker measurements database 15 observers viewing 1003 images free-viewing conditions. contrast common approach investigating properties fixation points analyze properties transition phases fixations. introduce hyperbolic geometry tool measure step length consecutive eye positions. show step lengths, measured hyperbolic euclidean geometry, follow generalized pareto distribution. results based hyperbolic distance robust based euclidean geometry. show structure space generalized pareto distributions used characterize identify individual observers.",3
"dropout bayesian approximation: appendix. show neural network arbitrary depth non-linearities, dropout applied every weight layer, mathematically equivalent approximation well known bayesian model. interpretation might offer explanation dropout's key properties, robustness over-fitting. interpretation allows us reason uncertainty deep learning, allows introduction bayesian machinery existing deep learning frameworks principled way. document appendix main paper ""dropout bayesian approximation: representing model uncertainty deep learning"" gal ghahramani, 2015.",18
"benchmark iris location deep learning detector evaluation. iris considered biometric trait highest unique probability. iris location important task biometrics systems, affecting directly results obtained specific applications iris recognition, spoofing contact lenses detection, among others. work defines iris location problem delimitation smallest squared window encompass iris region. order build benchmark iris location annotate (iris squared bounding boxes) four databases different biometric applications make publicly available community. besides 4 annotated databases, include 2 literature, perform experiments six databases, five obtained near infra-red sensors one visible light sensor. compare classical outstanding daugman iris location approach two window based detectors: 1) sliding window detector based features histogram gradients (hog) linear support vector machines classifier; 2) deep learning based detector fine-tuned yolo object detector. experimental results showed deep learning based detector outperforms ones terms accuracy runtime (gpus version) chosen whenever possible.",3
remote sensing image classification exploiting multiple kernel learning. propose strategy land use classification exploits multiple kernel learning (mkl) automatically determine suitable combination set features without requiring heuristic knowledge classification task. present novel procedure allows mkl achieve good performance case small training sets. experimental results publicly available datasets demonstrate feasibility proposed approach.,3
"unified framework stochastic matrix factorization via variance reduction. propose unified framework speed existing stochastic matrix factorization (smf) algorithms via variance reduction. framework general subsumes several well-known smf formulations literature. perform non-asymptotic convergence analysis framework derive computational sample complexities algorithm converge $\epsilon$-stationary point expectation. addition, extensive experiments wide class smf formulations demonstrate framework consistently yields faster convergence accurate output dictionary vis-\`a-vis state-of-the-art frameworks.",18
"conductivity electron coupled anharmonic phonons. study impact phonon anharmonicity electronic dynamics soft materials using nonperturbative quantum-classical approach. method applied one-dimensional model doped organic semiconductors low-frequency intermolecular lattice phonons. find anharmonicity leads phonon hardening increases mobility anharmonicity leads phonon softening decreases mobility. also test various approximations, including use adiabatic phonon disorder, effective harmonic model temperature-dependent frequencies, boltzmann transport equation second-order perturbation theory scattering rates. overall, find surprisingly good agreement methods accounting phonon anharmonicity important accurate prediction electronic transport including quantitative mobility values qualitative temperature dependence. model studied, phonon lifetime effects relatively little impact carrier transport, effective frequency shift due anharmonicity essential. cases highly asymmetric, non-gaussian disorder, effective harmonic model cannot quantitatively reproduce mobilities finite-frequency conductivity, especially true acoustic phonons.",2
"learning make predictions graphs autoencoders. examine two fundamental tasks associated graph representation learning: link prediction semi-supervised node classification. present densely connected autoencoder architecture capable learning joint representation local graph structure available external node features multi-task learning link prediction node classification. best knowledge, first architecture efficiently trained end-to-end single learning stage simultaneously perform link prediction node classification. provide comprehensive empirical evaluation models range challenging benchmark graph-structured datasets, demonstrate significant improvement accuracy related methods graph representation learning. code implementation available https://github.com/vuptran/graph-representation-learning",3
"learning virtual codec based deep convolutional neural network compress image. although deep convolutional neural network proved efficiently eliminate coding artifacts caused coarse quantization traditional codec, difficult train neural network front encoder gradient's back-propagation. paper, propose end-to-end image compression framework based convolutional neural network resolve problem non-differentiability quantization function standard codec. first, feature description neural network used get valid description low-dimension space respect ground-truth image amount image data greatly reduced storage transmission. image's valid description, standard image codec jpeg leveraged compress image, leads image's great distortion compression artifacts, especially blocking artifacts, detail missing, blurring, ringing artifacts. then, use post-processing neural network remove artifacts. due challenge directly learning non-linear function standard codec based convolutional neural network, propose learn virtual codec neural network approximate projection valid description image post-processed compressed image, gradient could efficiently back-propagated post-processing neural network feature description neural network training. meanwhile, advanced learning algorithm proposed train deep neural networks compression. obviously, priority proposed method compatible standard existing codecs learning strategy easily extended codecs based convolutional neural network. experimental results demonstrated advances proposed method compared several state-of-the-art approaches, especially low bit-rate.",3
"semantic web takes wing: programming ontologies tawny-owl. tawny-owl library provides fully-programmatic environment ontology building; enables use rich set tools ontology development, recasting development form programming. built clojure - modern lisp dialect, backed owl api. used simply, similar syntax owl manchester syntax, provides arbitrary extensibility abstraction. builds existing facilities clojure, provides rich modern programming tool chain, versioning, distributed development, build, testing continuous integration. paper, describe library, environment potential implications ontology development process.",3
"stochastic model vocabulary growth natural languages. propose stochastic model number different words given database incorporates dependence database size historical changes. main feature model existence two different classes words: (i) finite number core-words higher frequency affect probability new word used; (ii) remaining virtually infinite number noncore-words lower frequency used reduce probability new word used future. model relies careful analysis google-ngram database books published last centuries main consequence generalization zipf's heaps' law two scaling regimes. confirm generalizations yield best simple description data among generic descriptive models two free parameters depend language database. point view model main change historical time scales composition specific words included finite list core-words, observe decay exponentially time rate approximately 30 words per year english.",14
"fine-grained car detection visual census estimation. targeted socioeconomic policies require accurate understanding country's demographic makeup. end, united states spends 1 billion dollars year gathering census data race, gender, education, occupation unemployment rates. compared traditional method collecting surveys across many years costly labor intensive, data-driven, machine learning driven approaches cheaper faster--with potential ability detect trends close real time. work, leverage ubiquity google street view images develop computer vision pipeline predict income, per capita carbon emission, crime rates city attributes single source publicly available visual data. first detect cars 50 million images across 200 largest us cities train model predict demographic attributes using detected cars. facilitate work, collected largest challenging fine-grained dataset reported date consisting 2600 classes cars comprised images google street view web sources, classified car experts account even subtle visual differences. use data construct largest scale fine-grained detection system reported date. prediction results correlate well ground truth income data (r=0.82), massachusetts department vehicle registration, sources investigating crime rates, income segregation, per capita carbon emission, market research. finally, learn interesting relationships cars neighborhoods allowing us perform first large scale sociological analysis cities using computer vision techniques.",3
"fiber orientation estimation guided deep network. diffusion magnetic resonance imaging (dmri) currently tool noninvasively imaging brain's white matter tracts. fiber orientation (fo) key feature computed dmri fiber tract reconstruction. number fos voxel usually small, dictionary-based sparse reconstruction used estimate fos relatively small number diffusion gradients. however, accurate fo estimation regions complex fo configurations presence noise still challenging. work explore use deep network fo estimation dictionary-based framework propose algorithm named fiber orientation reconstruction guided deep network (fordn). fordn consists two steps. first, use smaller dictionary encoding coarse basis fos represent diffusion signals. estimate mixture fractions dictionary atoms (and thus coarse fos), deep network designed specifically solving sparse reconstruction problem. here, smaller dictionary used reduce computational cost training. second, coarse fos inform final fo estimation, larger dictionary encoding dense basis fos used weighted l1-norm regularized least squares problem solved encourage fos consistent network output. fordn evaluated compared state-of-the-art algorithms estimate fos using sparse reconstruction simulated real dmri data, results demonstrate benefit using deep network fo estimation.",3
"semantic segmentation human thigh quadriceps muscle magnetic resonance images. paper presents end-to-end solution mri thigh quadriceps segmentation. first attempt deep learning methods used mri thigh segmentation task. use state-of-the-art fully convolutional networks transfer learning approach semantic segmentation regions interest mri thigh scans. improve performance segmentation, propose post-processing technique using basic image processing methods. proposed method, established new benchmark mri thigh quadriceps segmentation mean jaccard similarity index 0.9502 processing time 0.117 second per image.",3
"robust fundamental lemma data-driven control. fundamental lemma willems coauthors facilitates parameterization trajectories linear time-invariant system terms single, measured one. result plays important role data-driven simulation control. hood, fundamental lemma works applying persistently exciting input system. ensures hankel matrix resulting input/output data ""right"" rank, meaning columns span entire subspace trajectories. however, binary rank conditions known fragile sense small additive noise could already cause hankel matrix full rank. therefore, extended abstract present robust version fundamental lemma. idea behind approach guarantee certain lower bounds singular values data hankel matrix, rather mere rank conditions. achieved designing inputs experiment minimum singular value deeper input hankel matrix sufficiently large. inspires new quantitative robust notion persistency excitation. relevance result data-driven control also highlighted comparing predictive control performance varying degrees persistently exciting data.",10
"deep interest network click-through rate prediction. better extract users' interest exploiting rich historical behavior data crucial building click-through rate (ctr) prediction model online advertising system e-commerce industry. two key observations user behavior data: i) \textbf{diversity}. users interested different kinds goods visiting e-commerce site. ii) \textbf{local activation}. whether users click click good depends part related historical behavior. however, traditional ctr models lack capturing structures behavior data. paper, introduce new proposed model, deep interest network (din), developed deployed display advertising system alibaba. din represents users' diverse interests interest distribution designs attention-like network structure locally activate related interests according candidate ad, proven effective significantly outperforms traditional model. overfitting problem easy encounter training industrial deep network large scale sparse inputs. study problem carefully propose useful adaptive regularization technique.",18
"vlbi observations vik j2318-3113, quasar z = 6.44. nature jets active galactic nuclei (agn) early universe feedback host galaxy remain highly topical question. observations radio structure high-redshift agns enabled long baseline interferometry (vlbi) provide indispensable input studies properties role galaxies' evolution. now, five agns redshift $z > 6$ studied vlbi technique. vikj2318-3113 recently discovered quasar z = 6.44 imaged vlbi current work. present first vlbi imaging results high-redshift quasar, aim corroborating high-resolution appearance physical model object. carried vlbi phase-referencing observations vikj2318-3113 using long baseline array two frequencies, 1.6 4.7 ghz, obtained first view radio structure milliarcsecond scale. source clearly detected 1.6 ghz. found almost radio emission comes pc-scale core region. dual-frequency observations constrain spectral index brightness temperature radio core. properties similar known high-redshift radio-loud agns.",0
"method artificial systems. document written intention describe detail method means computer program reason world doing, increase analogue living system. literature rife apparent we, scientists engineers, found solution, document attempt solution grounding intellectual arguments within tenets human cognition western philosophy. result characteristic description method describe artificial system analogous performed human. approach substance master's thesis, explored deeply course postdoc research. focuses primarily context awareness choice set within boundary available epistemology, serves describe it. expanded upon, description strives discover agreement kant's critique reason understand could applied define architecture design. intention never mimic human biological systems, rather, understand profoundly fundamental rules, leveraged correctly, results artificial consciousness noumenon keeping perception phenomenon.",3
"cross-domain semantic parsing via paraphrasing. existing studies semantic parsing mainly focus in-domain setting. formulate cross-domain semantic parsing domain adaptation problem: train semantic parser source domains adapt target domain. due diversity logical forms different domains, problem presents unique intriguing challenges. converting logical forms canonical utterances natural language, reduce semantic parsing paraphrasing, develop attentive sequence-to-sequence paraphrase model general flexible adapt different domains. discover two problems, small micro variance large macro variance, pre-trained word embeddings hinder direct use neural networks, propose standardization techniques remedy. popular overnight dataset, contains eight domains, show cross-domain training standardized pre-trained word embeddings bring significant improvement.",3
"sparse markov decision processes causal sparse tsallis entropy regularization reinforcement learning. paper, sparse markov decision process (mdp) novel causal sparse tsallis entropy regularization proposed.the proposed policy regularization induces sparse multi-modal optimal policy distribution sparse mdp. full mathematical analysis proposed sparse mdp provided.we first analyze optimality condition sparse mdp. then, propose sparse value iteration method solves sparse mdp prove convergence optimality sparse value iteration using banach fixed point theorem. proposed sparse mdp compared soft mdps utilize causal entropy regularization. show performance error sparse mdp constant bound, error soft mdp increases logarithmically respect number actions, performance error caused introduced regularization term. experiments, apply sparse mdps reinforcement learning problems. proposed method outperforms existing methods terms convergence speed performance.",3
"gibbs-ringing artifact removal based local subvoxel-shifts. gibbs-ringing well known artifact manifests spurious oscillations vicinity sharp image transients, e.g. tissue boundaries. origin seen truncation k-space mri data-acquisition. consequently, correction techniques like gegenbauer reconstruction extrapolation methods aim recovering missing data. here, present simple robust method exploits different view gibbs-phenomena. truncation k-space interpreted convolution sinc-function image space. hence, severity artifacts depends sinc-function sampled. propose re-interpolate image based local, subvoxel shifts sample ringing pattern zero-crossings oscillating sinc-function. this, artifact effectively robustly removed minimal amount smoothing.",14
"high-resolution multispectral dataset semantic segmentation. unmanned aircraft decreased cost required collect remote sensing imagery, enabled researchers collect high-spatial resolution data multiple sensor modalities frequently easily. increase data push need semantic segmentation frameworks able classify non-rgb imagery, type algorithmic development requires increase publicly available benchmark datasets class labels. paper, introduce high-resolution multispectral dataset image labels. new benchmark dataset pre-split training/testing folds order standardize evaluation continue push state-of-the-art classification frameworks non-rgb imagery.",3
"entity-specific sentiment classification yahoo news comments. sentiment classification widely used product reviews online social media forums, twitter, blogs. however, problem classifying sentiment user comments news sites addressed yet. news sites cover wide range domains including politics, sports, technology, entertainment, contrast online social sites forums review sites, specific particular domain. user associated news site likely post comments diverse topics (e.g., politics, smartphones, sports) diverse entities (e.g., obama, iphone, google). classifying sentiment users tied various entities may help obtain holistic view personality, could useful applications online advertising, content personalization, political campaign planning. paper, formulate problem entity-specific sentiment classification comments posted news articles yahoo news propose novel features specific news comments. experimental results show models outperform state-of-the-art baselines.",3
"storytelling agents personality adaptivity. explore expression personality adaptivity gestures virtual agents storytelling task. conduct two experiments using four different dialogic stories. manipulate agent personality extraversion scale, whether agents adapt one another gestural performance agent gender. results show subjects able perceive intended variation extraversion different virtual agents, independently story telling gender agent. second study shows subjects also prefer adaptive nonadaptive virtual agents.",3
"supervised learning dynamical system learning. recently substantial interest spectral methods learning dynamical systems. methods popular since often offer good tradeoff computational statistical efficiency. unfortunately, difficult use extend practice: e.g., make difficult incorporate prior information sparsity structure. address problem, present new view dynamical system learning: show learn dynamical systems solving sequence ordinary supervised learning problems, thereby allowing users incorporate prior knowledge via standard techniques l1 regularization. many existing spectral methods special cases new framework, using linear regression supervised learner. demonstrate effectiveness framework showing examples nonlinear regression lasso let us learn better state representations plain linear regression does; correctness instances follows directly general analysis.",18
"soft power diagrams. many applications data analysis begin set points euclidean space partitioned clusters. common tasks devise classifier deciding clusters new point associated to, finding outliers respect clusters, identifying type clustering used partition. one common kinds clusterings (balanced) least-squares assignments respect given set sites. these, 'separating power diagram' cluster lies cell. present paper, aim efficient algorithms outlier detection computation thresholds measure similar clustering least-squares assignment fixed sites. purpose, devise new model computation 'soft power diagram', allows soft separation clusters 'point counting properties'; e.g. able prescribe many points want classify outliers. results hold general non-convex model free sites, describe proofs general way. locally optimal solutions satisfy aforementioned point counting properties. target applications use fixed sites, algorithms efficiently solvable global optimality linear programming.",3
"improving variational auto-encoders using convex combination linear inverse autoregressive flow. paper, propose new volume-preserving flow show performs similarly linear general normalizing flow. idea enrich linear inverse autoregressive flow introducing multiple lower-triangular matrices ones diagonal combining using convex combination. experimental studies mnist histopathology data show proposed approach outperforms volume-preserving flows competitive current state-of-the-art linear normalizing flow.",18
"distributed quasi-newton algorithm empirical risk minimization nonsmooth regularization. paper, propose communication- computation- efficient distributed optimization algorithm using second- order information solving erm problems nonsmooth regularization term. current second-order quasi- newton methods problem either work well distributed setting work specific regularizers. algorithm uses successive quadratic approximations, describe maintain approximation hessian solve subproblems efficiently distributed manner. proposed method enjoys global linear convergence broad range non-strongly convex problems includes commonly used erms, thus requiring lower communication complexity. empirical results also demonstrate method significantly improves communication cost running time current state-of-the-art methods.",10
"convex geometries representable colors, ellipses plane, impossible circles. convex geometry closure system satisfying anti-exchange property. paper, following work k. adaricheva m. bolat (2016) polymath reu 2020 team, continues investigate representations convex geometries 5-element base set. introduces several properties: opposite property, nested triangle property, area q property, separation property, convex geometries circles plane, preventing representation numerous convex geometries 5-element base set. also demonstrates 672 convex geometries 5-element base set representation ellipses, given appendix without known representation circles, introduces method expanding representation circles defining unary predicates, shown colors.",10
"computational model syntactic processing: ambiguity resolution interpretation. syntactic ambiguity abounds natural language, yet humans difficulty coping it. fact, process ambiguity resolution almost always unconscious. infallible, however, example 1 demonstrates. 1. horse raced past barn fell. sentence perfectly grammatical, evident appears following context: 2. two horses shown prospective buyer. one raced past meadow. raced past barn. ... grammatical yet unprocessable sentences 1 called `garden-path sentences.' existence provides opportunity investigate human sentence processing mechanism studying fails. aim thesis construct computational model language understanding predict processing difficulty. data modeled known examples garden path non-garden path sentences, results psycholinguistics. widely believed two distinct loci computation sentence processing: syntactic parsing semantic interpretation. one longstanding controversy two modules bears responsibility immediate resolution ambiguity. claim latter, syntactic processing module simple device blindly faithfully constructs possible analyses sentence current point processing. interpretive module serves filter, occasionally discarding certain analyses deems less appropriate ongoing discourse competitors. document divided three parts. first introductory, reviews selection proposals sentence processing literature. second part explores body data adduced support theory structural preferences --- one inconsistent present claim. show current proposal specified account available data, moreover predict structural preference theories go wrong. third part theoretical investigation well proposed architecture realized using current conceptions linguistic competence. it, present parsing algorithm meaning-based ambiguity resolution method.",1
"computational approach emergence evolution language - evolutionary naming game model. computational modelling multi-agent systems becoming important technique studying language evolution. present brief introduction rapidly developing field, well contributions include analysis evolutionary naming-game model. model communicating agents, try establish common vocabulary, equipped evolutionarily selected learning ability. coupling biological linguistic ingredients results abrupt transition: upon small change model control parameter poorly communicating group linguistically unskilled agents transforms almost perfectly communicating group large learning abilities. genetic imprinting learning abilities proceeds via baldwin effect: initially unskilled communicating agents learn language creates niche evolutionary pressure increase learning ability. assumption communication intensity increases continuously finite speed, transition split several transition-like changes. shows speed cultural changes, sets additional characteristic timescale, might yet another factor affecting evolution language. opinion, model shows linguistic biological processes strong influence effect certainly contributed explosive development species.",14
"group symmetry non-gaussian covariance estimation. consider robust covariance estimation group symmetry constraints. non-gaussian covariance estimation, e.g., tyler scatter estimator multivariate generalized gaussian distribution methods, usually involve non-convex minimization problems. recently, shown underlying principle behind success extended form convexity geodesics manifold positive definite matrices. modern approach improve estimation accuracy exploit prior knowledge via additional constraints, e.g., restricting attention specific classes covariances adhere prior symmetry structures. paper, prove group symmetry constraints also geodesically convex therefore incorporated various non-gaussian covariance estimators. practical examples sets include: circulant, persymmetric complex/quaternion proper structures. provide simple numerical technique finding maximum likelihood estimates constraints, demonstrate performance advantage using synthetic experiments.",18
"survey schemes stochastic gradient descent applications m-estimation. certain situations shall undoubtedly common big data era, datasets available massive computing statistics full sample hardly feasible, unfeasible. natural approach context consists using survey schemes substituting ""full data"" statistics counterparts based resulting random samples, manageable size. main purpose paper investigate impact survey sampling unequal inclusion probabilities stochastic gradient descent-based m-estimation methods large-scale statistical machine-learning problems. precisely, prove that, presence priori information, one may significantly increase asymptotic accuracy choosing appropriate first order inclusion probabilities, without affecting complexity. striking results described limit theorems also illustrated numerical experiments.",18
"preemption overdetermination formal theories causality. one key challenges looking causes complex event determine causal status factors neither individually necessary individually sufficient produce event. order reason factors taken account, need vocabulary distinguish different cases. philosophy, concept overdetermination concept preemption serve important purpose regard, although exact meaning tends remain elusive. paper, provide theory-neutral definitions concepts using structural equations halpern-pearl tradition. definitions presuppose particular causal theory, take theory variable parameter. enables us specify formal constraints theories causality, terms pre-theoretic understanding preemption overdetermination actually mean. demonstrate usefulness presenting arguing call principle presumption. roughly speaking, principle states possible cause regarded preempted independent evidence support inference. conclude showing principle presumption violated two main theories causality formulated halpern-pearl tradition. paper concludes defining class empirical causal theories, characterised terms fixed-point counterfactual reasoning difference-making. argued theories actual causality ought empirical.",3
"deep keyphrase generation. keyphrase provides highly-summative information effectively used understanding, organizing retrieving text content. though previous studies provided many workable solutions automated keyphrase extraction, commonly divided to-be-summarized content multiple text chunks, ranked selected meaningful ones. approaches could neither identify keyphrases appear text, capture real semantic meaning behind text. propose generative model keyphrase prediction encoder-decoder framework, effectively overcome drawbacks. name deep keyphrase generation since attempts capture deep semantic meaning content deep learning method. empirical analysis six datasets demonstrates proposed model achieves significant performance boost extracting keyphrases appear source text, also generate absent keyphrases based semantic meaning text. code dataset available https://github.com/memray/seq2seq-keyphrase.",3
"processing natural language ongoing actions. actions may proceed planned; may interrupted, resumed overridden. challenge handle natural language understanding system. describe extensions existing implementation control autonomous systems natural language, enable systems handle incoming language requests regarding actions. language communication autonomous systems (lcas) extended support x-nets, parameterized executable schemas representing actions. x-nets enable system control actions desired level granularity, providing mechanism language requests processed asynchronously. standard semantics supported include requests stop, continue, override existing action. specific domain demonstrated control motion simulated robot, approach general, could applied domains.",3
"evaluation alzheimer's disease analysis mr images using multilayer perceptrons kohonen som classifiers alternative adc maps. alzheimer's disease common cause dementia, yet hard diagnose precisely without invasive techniques, particularly onset disease. work approaches image analysis classification synthetic multispectral images composed diffusion-weighted magnetic resonance (mr) cerebral images evaluation cerebrospinal fluid area measuring advance alzheimer's disease. clinical 1.5 mr imaging system used acquire images presented. classification methods based multilayer perceptrons kohonen self-organized map classifiers. assume classes interest separated hyperquadrics. therefore, 2-degree polynomial network used classify original image, generating ground truth image. classification results used improve usual analysis apparent diffusion coefficient map.",5
"learning rank based analogical reasoning. object ranking ""learning rank"" important problem realm preference learning. basis training data form set rankings objects represented feature vectors, goal learn ranking function predicts linear order new set objects. paper, propose new approach object ranking based principles analogical reasoning. specifically, inference pattern formalized terms so-called analogical proportions summarized follows: given objects $a,b,c,d$, object $a$ known preferred $b$, $c$ relates $d$ $a$ relates $b$, $c$ (supposedly) preferred $d$. method applies pattern main building block combines ideas techniques instance-based learning rank aggregation. based first experimental results data sets various domains (sports, education, tourism, etc.), conclude approach highly competitive. appears specifically interesting situations objects coming different subdomains, hence require kind knowledge transfer.",18
"optimization convolutional neural network using microcanonical annealing algorithm. convolutional neural network (cnn) one prominent architectures algorithm deep learning. shows remarkable improvement recognition classification objects. method also proven effective variety computer vision machine learning problems. deep learning, however, training cnn interesting yet challenging. recently, metaheuristic algorithms used optimize cnn using genetic algorithm, particle swarm optimization, simulated annealing harmony search. paper, another type metaheuristic algorithms different strategy proposed, i.e. microcanonical annealing optimize convolutional neural network. performance proposed method tested using mnist cifar-10 datasets. although experiment results mnist dataset indicate increase computation time (1.02x - 1.38x), nevertheless proposed method considerably enhance performance original cnn (up 4.60\%). cifar10 dataset, currently, state art 96.53\% using fractional pooling, proposed method achieves 99.14\%.",3
"preserving differential privacy features distributed estimation. privacy crucial many applications machine learning. legal, ethical societal issues restrict sharing sensitive data making difficult learn datasets partitioned many parties. one important instance distributed setting arises information record dataset held different data owners (the design matrix ""vertically-partitioned""). setting approaches exist private data sharing purposes statistical estimation classical setup differential privacy ""trusted curator"" preparing data apply. work notion $(\epsilon,\delta)$-distributed differential privacy extends single-party differential privacy distributed, vertically-partitioned case. propose pride, scalable framework distributed estimation party communicates perturbed random projections locally held features ensuring $(\epsilon,\delta)$-distributed differential privacy preserved. $\ell_2$-penalized supervised learning problems pride bounded estimation error compared optimal estimates obtained without privacy constraints non-distributed setting. confirm empirically real world synthetic datasets.",18
"pac learnability concept class non-atomic measures: problem vidyasagar. response 1997 problem m. vidyasagar, state necessary sufficient condition distribution-free pac learnability concept class $\mathscr c$ family non-atomic (diffuse) measures domain $\omega$. clearly, finiteness classical vapnik-chervonenkis dimension $\mathscr c$ sufficient, longer necessary, condition. besides, learnability $\mathscr c$ non-atomic measures imply uniform glivenko-cantelli property regard non-atomic measures. learnability criterion stated terms combinatorial parameter $\vc({\mathscr c}\,{\mathrm{mod}}\,\omega_1)$ call vc dimension $\mathscr c$ modulo countable sets. new parameter obtained ``thickening up'' single points definition vc dimension uncountable ``clusters''. equivalently, $\vc(\mathscr c\modd\omega_1)\leq d$ every countable subclass $\mathscr c$ vc dimension $\leq d$ outside countable subset $\omega$. new parameter also expressed classical vc dimension $\mathscr c$ calculated suitable subset compactification $\omega$. make measurability assumptions $\mathscr c$, assuming instead validity martin's axiom (ma).",3
"adapting stochastic search real-time dynamic weighted constraint satisfaction. work presents two new algorithms performing constraint satisfaction. first algorithm presented, dmaxwalksat, constraint solver specialized solving dynamic, weighted constraint satisfaction problems. second algorithm, rdmaxwalksat, derivative dmaxwalksat modified anytime algorithm, hence support realtime constraint satisfaction. dmaxwalksat shown offer performance advantages terms solution quality runtime parent constraint solver, maxwalksat. rdmaxwalksat shown support anytime operation. introduction algorithms brings another tool areas computer science naturally represent problems constraint satisfaction problems, example robust coherence algorithm.",3
"density-based region search arbitrary shape object localization. region search widely used object localization. typically, region search methods project score classifier image plane, search region maximal score. recently proposed region search methods, efficient subwindow search efficient region search, %which localize objects score distribution image much efficient sliding window search. however, classifiers tasks, projected scores nearly positive, hence maximizing score region results localizing nearly entire images objects, meaningless. paper, observe large scores mainly concentrated around objects. based observation, propose method, named level set maximum-weight connected subgraph (ls-mwcs), localizes objects arbitrary shapes searching regions densest score rather maximal score. region density controlled parameter flexibly. prove important property proposed ls-mwcs, guarantees region densest score searched. moreover, ls-mwcs efficiently optimized belief propagation. method evaluated problem weakly-supervised object localization, quantitative results demonstrate superiorities ls-mwcs compared state-of-the-art methods.",3
"robust photometric stereo using learned image gradient dictionaries. photometric stereo method estimating normal vectors object images object varying lighting conditions. motivated several recent works extend photometric stereo general objects lighting conditions, study new robust approach photometric stereo utilizes dictionary learning. specifically, propose analyze two approaches adaptive dictionary regularization photometric stereo problem. first, propose image preprocessing step utilizes adaptive dictionary learning model remove noise non-idealities image dataset estimating normal vectors. also propose alternative model directly apply adaptive dictionary regularization normal vectors estimation. study practical performance methods extensive simulations, demonstrate state-of-the-art performance methods presence noise.",3
"explaining logical nature electrical solitons neural circuits. neurons modeled electrically based ferroelectric membranes thin enough permit charge transfer, conjectured tunneling result thermally energetic ions random electrons. membranes triggered produce electrical solitons, main signals brain associative memory logical processing. dendritic circuits modeled, electrical solitons simulated demonstrate nature soliton propagation, soliton reflection, collision solitons, well soliton gates, gates, xor gates gates.",3
"deep learning-based search lensed gravitational-wave signals binary black hole events gwtc-1 -2. present result deep learning-based search beating patterns, one gravitational lensing signatures, observed gravitational-wave signals. search, examine binary black hole events first second gravitational-wave transient catalogs. search first endeavor utilizing deep learning searching lensing signatures gravitational waves. specifically, search identifies beating patterns induced lenses masses $10^3$--$10^5m_\odot$ spectrograms gravitational-wave signals. train deep learning model spectrograms simulated noisy gravitational-wave signals classify binary black hole events two classes, lensed unlensed signals. introduce ensemble learning deep learning model employ majority voting strategy predictions ensemble learners. majority voting-based primary classification classifies one event, gw190707\_093326, forty-six events, lensed class. however, upon estimating $p$-value event, observe uncertainty $p$-value still includes possibility event unlensed. therefore, conclude gravitational-wave signal gw190707\_093326 likely unlensed signal and, consequently, search finds significant evidence beating patterns evaluated binary black hole events.",6
"ism high-redshift galaxies: clues c iii] [c ii] lines. tight relation [c ii] $ 158 \ {\rm{\mu m}}$ line luminosity star formation rate (sfr) observed local galaxies. $z > 5$, galaxies instead deviate downwards local $\sigma_{\rm{[c ii]}}-\sigma_{\rm{sfr}}$ relation. deviation might caused different interstellar medium (ism) properties galaxies early epochs. combine [c ii] sfr data c iii] $\lambda1909$ line observations physical models. additionally investigate ism properties: burstiness $\kappa_s$, total gas density $n$, metallicity $z$, affect deviation $\sigma_{\rm{[c ii]}}-\sigma_{\rm{sfr}}$ relation sources. present vlt/x-shooter observations targeting c iii] $\lambda1909$ line emission three galaxies $5.5<z<7.0$. extend sample galaxies including archival x-shooter muse data eight sources $2<z<7$ eleven star-forming systems $6 < z< 7.5$ c iii] [c ii] data reported literature. detect c iii] $\lambda \lambda1907, 1909$ line emission hz10 derive intrinsic flux c iii] $ \lambda1909$ line. constrain ism properties sample galaxies: $\kappa_s$, $n$, $z$, applying physically motivated model based mcmc algorithm. part, high-$z$, star-forming galaxies show subsolar metallicities. majority sources $\log{(\kappa_s)} \gtrsim 1$, i.e., overshoot kennicutt-schmidt relation one order magnitude. suggests whole ks relation might shifted upwards early times. next, high-$z$ galaxies sample lie $\sigma_{\rm{[c ii]}}-\sigma_{\rm{sfr}}$ local relation. total gas density $n$ shows strongest correlation deviation local $\sigma_{\rm{[c ii]}}-\sigma_{\rm{sfr}}$ relation, agreement theoretical models.",0
"learning non-iid data: fast rates one-vs-all multiclass plug-in classifiers. prove new fast learning rates one-vs-all multiclass plug-in classifiers trained either exponentially strongly mixing data data generated converging drifting distribution. two typical scenarios training data iid. learning rates obtained multiclass version tsybakov's margin assumption, type low-noise assumption, depend number classes. results general include previous result binary-class plug-in classifiers iid data special case. contrast previous works least squares svms binary-class setting, results retain optimal learning rate iid case.",18
"new algorithmic decision categorical syllogisms via caroll's diagrams. paper, deal calculus system slcd (syllogistic logic carroll diagrams), gives formal approach logical reasoning diagrams, representations fundamental aristotelian categorical propositions show closed syllogistic criterion inference deletion middle term. therefore, implemented let formalism comprise synchronically bilateral trilateral diagrammatical appearance naive algorithmic nature. also, need specific knowledge exclusive ability understand well use it. consequently, give effective algorithm used determine whether syllogistic reasoning valid using slcd.",3
"support vector machine classification dimensionally reduced structural mri images dementia. classify very-mild moderate dementia patients (cdr ranging 0 2) using support vector machine classifier acting dimensionally reduced feature set derived mri brain scans 416 subjects available oasis-brains dataset. use image segmentation principal component analysis reduce dimensionality data. resulting feature set contains 11 features subject. performance classifiers evaluated using 10-fold cross-validation. using linear (gaussian) kernels, obtain training classification accuracy 86.4% (90.1%), test accuracy 85.0% (85.7%), test precision 68.7% (68.5%), test recall 68.0% (74.0%), test matthews correlation coefficient 0.594 (0.616).",3
"real-time optical flow-based video stabilization unmanned aerial vehicles. paper describes development novel algorithm tackle problem real-time video stabilization unmanned aerial vehicles (uavs). two main components algorithm: (1) designing suitable model global motion uav, proposed algorithm avoids necessity estimating general motion model, projective transformation, considers simpler motion models, rigid transformation similarity transformation. (2) achieve high processing speed, optical-flow based tracking employed lieu conventional tracking matching methods used state-of-the-art algorithms. two new ideas resulted real-time stabilization algorithm, developed two phases. stage considers processing whole sequence frames video achieving average processing speed 50fps several publicly available benchmark videos. next, stage ii undertakes task real-time video stabilization using multi-threading implementation algorithm designed stage i.",3
"temporal autoencoding improves generative models time series. restricted boltzmann machines (rbms) generative models learn useful representations samples dataset unsupervised fashion. widely employed unsupervised pre-training method machine learning. rbms modified model time series two main ways: temporal rbm stacks number rbms laterally introduces temporal dependencies hidden layer units; conditional rbm, hand, considers past samples dataset conditional bias learns representation takes account. propose new training method trbm crbm, enforces dynamic structure temporal datasets. treating temporal models denoising autoencoders, considering past frames dataset corrupted versions present frame minimizing reconstruction error present data model. call approach temporal autoencoding. leads significant improvement performance models filling-in-frames task across number datasets. error reduction motion capture data 56\% crbm 80\% trbm. taking posterior mean prediction instead single samples improves model's estimates, decreasing error much 91\% crbm motion capture data. also trained model perform forecasting large number datasets found ta pretraining consistently improve performance forecasts. furthermore, looking prediction error across time, see improvement reflects better representation dynamics data opposed bias towards reconstructing observed data short time scale.",18
"multi-agent only-knowing revisited. levesque introduced notion only-knowing precisely capture beliefs knowledge base. also showed only-knowing used formalize non-monotonic behavior within monotonic logic. despite appeal, attempts extend only-knowing many agent case undesirable properties. belief model halpern lakemeyer, instance, appeals proof-theoretic constructs semantics needs axiomatize validity part logic. also clear generalize ideas first-order case. paper, propose new account multi-agent only-knowing which, first time, natural possible-world semantics quantified language equality. provide, propositional fragment, sound complete axiomatization faithfully lifts levesque's proof theory many agent case. also discuss comparisons earlier approach halpern lakemeyer.",3
"gradient descent converges minimizers. show gradient descent converges local minimizer, almost surely random initialization. proved applying stable manifold theorem dynamical systems theory.",18
"developments theory randomized shortest paths comparison graph node distances. lately several suggestions parametrized distances graph generalize shortest path distance commute time resistance distance. need developing distances risen observation above-mentioned common distances many situations fail take account global structure graph. article, develop theory one family graph node distances, known randomized shortest path dissimilarity, foundation statistical physics. show randomized shortest path dissimilarity easily computed closed form pairs nodes graph. moreover, come new definition distance measure call free energy distance. free energy distance seen upgrade randomized shortest path dissimilarity defines metric, addition satisfies graph-geodetic property. derivation computation free energy distance also straightforward. make comparison set generalized distances interpolate shortest path distance commute time, resistance distance. comparison focuses applicability distances graph node clustering classification. comparison, general, shows parametrized distances perform well tasks. particular, see results obtained free energy distance among best experiments.",18
"fast vehicle detection aerial imagery. recent years, several real-time near real-time object detectors developed. however object detectors typically designed first-person view images subject large image directly apply well detecting vehicles aerial imagery. though detectors developed aerial imagery, either slow handle multi-scale imagery well. popular yolov2 detector modified vastly improve performance aerial data. modified detector compared faster rcnn several aerial imagery datasets. proposed detector gives near state art performance 4x speed.",3
"measure concentration random maximum a-posteriori perturbations. maximum a-posteriori (map) perturbation framework emerged useful approach inference learning high dimensional complex models. maximizing randomly perturbed potential function, map perturbations generate unbiased samples gibbs distribution. unfortunately, computational cost generating many high-dimensional random variables prohibitive. efficient algorithms use sequential sampling strategies based expected value low dimensional map perturbations. paper develops new measure concentration inequalities bound number samples needed estimate expected values. applying general result map perturbations yield efficient algorithm approximate sampling gibbs distribution. measure concentration result general interest may applicable areas involving expected estimations.",3
"comal tracking: tracking points object boundaries. traditional point tracking algorithms klt use local 2d information aggregation feature detection tracking, due performance degrades object boundaries separate multiple objects. recently, comal features proposed handle case. however, proposed simple tracking framework points re-detected frame matched. inefficient may also lose many points re-detected next frame. propose novel tracking algorithm accurately efficiently track comal points. this, level line segment associated comal points matched mser segments next frame using shape-based matching matches filtered using texture-based matching. experiments show improvements simple re-detect-and-match framework well klt terms speed/accuracy different real-world applications, especially object boundaries.",3
"distribution-dependent sample complexity large margin learning. obtain tight distribution-specific characterization sample complexity large-margin classification l2 regularization: introduce margin-adapted dimension, simple function second order statistics data distribution, show distribution-specific upper lower bounds sample complexity, governed margin-adapted dimension data distribution. upper bounds universal, lower bounds hold rich family sub-gaussian distributions independent features. conclude new quantity tightly characterizes true sample complexity large-margin classification. prove lower bound, develop several new tools independent interest. include new connections shattering hardness learning, new properties shattering linear classifiers, new lower bound smallest eigenvalue random gram matrix generated sub-gaussian variables. results used quantitatively compare large margin learning learning rules, improve effectiveness methods use sample complexity bounds, active learning.",18
"robot language learning, generation, comprehension. present unified framework supports grounding natural-language semantics robotic driving. framework supports acquisition (learning grounded meanings nouns prepositions human annotation robotic driving paths), generation (using acquired meanings generate sentential description new robotic driving paths), comprehension (using acquired meanings support automated driving accomplish navigational goals specified natural language). evaluate performance three tasks independent human judges rate semantic fidelity sentences associated paths, achieving overall average correctness 94.6% overall average completeness 85.6%.",3
"smiles2vec: interpretable general-purpose deep neural network predicting chemical properties. chemical databases store information text representations, smiles format universal standard used many cheminformatics software. encoded smiles string structural information used predict complex chemical properties. work, develop smiles2vec, deep rnn automatically learns features smiles predict chemical properties, without need additional explicit feature engineering. using bayesian optimization methods tune network architecture, show optimized smiles2vec model serve general-purpose neural network predicting distinct chemical properties including toxicity, activity, solubility solvation energy, also outperforming contemporary mlp neural networks uses engineered features. furthermore, demonstrate proof-of-concept interpretability developing explanation mask localizes important characters used making prediction. tested solubility dataset, identified specific parts chemical consistent established first-principles knowledge accuracy 88%. work demonstrates neural networks learn technically accurate chemical concept provide state-of-the-art accuracy, making interpretable deep neural networks useful tool relevance chemical industry.",18
"abstract argumentation / persuasion / dynamics. act persuasion, key component rhetoric argumentation, may viewed dynamics modifier. modifiers well-known research fields: recall dynamic epistemic logic operators modify possible world accessibilities, recall side effects concurrency programming languages. consider persuasion abstract argumentation undertaking similar role. extend dung's frameworks acts persuasion among agents abstract persuasion argumentation (apa), set forth properties related arguments' admissibilities. show way enriching basic notion admissibility ctl (computation tree logic) encoding, also permits importation theoretical results known logic argumentation frameworks.",3
"offline a/b testing recommender systems. a/b testing online new version recommender system, usual perform offline evaluations historical data. focus evaluation methods compute estimator potential uplift revenue could generate new technology. helps iterate faster avoid losing money detecting poor policies. estimators known counterfactual off-policy estimators. show traditional counterfactual estimators capped importance sampling normalised importance sampling experimentally satisfying bias-variance compromises context personalised product recommendation online advertising. propose two variants counterfactual estimates different modelling bias prove accurate real-world conditions. provide benchmark estimators showing correlation business metrics observed running online a/b tests commercial recommender system.",18
non-isotopic transverse tori engel manifolds. every engel manifold construct infinite family pairwise non-isotopic transverse tori smoothly isotopic. distinguish transverse tori family introduce homological invariant transverse tori similar self-linking number transverse knots contact 3-manifolds. analogous results presented legendrian tori even contact 4-manifolds.,10
"statistical compressed sensing gaussian mixture models. novel framework compressed sensing, namely statistical compressed sensing (scs), aims efficiently sampling collection signals follow statistical distribution, achieving accurate reconstruction average, introduced. scs based gaussian models investigated depth. signals follow single gaussian model, gaussian bernoulli sensing matrices o(k) measurements, considerably smaller o(k log(n/k)) required conventional cs based sparse models, n signal dimension, optimal decoder implemented via linear filtering, significantly faster pursuit decoders applied conventional cs, error scs shown tightly upper bounded constant times best k-term approximation error, overwhelming probability. failure probability also significantly smaller conventional sparsity-oriented cs. stronger yet simpler results show sensing matrix, error gaussian scs upper bounded constant times best k-term approximation probability one, bound constant efficiently calculated. gaussian mixture models (gmms), assume multiple gaussian distributions signal follows one unknown index, piecewise linear estimator introduced decode scs. accuracy model selection, heart piecewise linear decoder, analyzed terms properties gaussian distributions number sensing measurements. maximum posteriori expectation-maximization algorithm iteratively estimates gaussian models parameters, signals model selection, decodes signals, presented gmm-based scs. real image sensing applications, gmm-based scs shown lead improved results compared conventional cs, considerably lower computational cost.",3
"learning generative models sinkhorn divergences. ability compare two degenerate probability distributions (i.e. two probability distributions supported two distinct low-dimensional manifolds living much higher-dimensional space) crucial problem arising estimation generative models high-dimensional observations arising computer vision natural language. known optimal transport metrics represent cure problem, since specifically designed alternative information divergences handle problematic scenarios. unfortunately, training generative machines using ot raises formidable computational statistical challenges, (i) computational burden evaluating ot losses, (ii) instability lack smoothness losses, (iii) difficulty estimate robustly losses gradients high dimension. paper presents first tractable computational method train large scale generative models using optimal transport loss, tackles three issues relying two key ideas: (a) entropic smoothing, turns original ot loss one computed using sinkhorn fixed point iterations; (b) algorithmic (automatic) differentiation iterations. two approximations result robust differentiable approximation ot loss streamlined gpu execution. entropic smoothing generates family losses interpolating wasserstein (ot) maximum mean discrepancy (mmd), thus allowing find sweet spot leveraging geometry ot favorable high-dimensional sample complexity mmd comes unbiased gradient estimates. resulting computational architecture complements nicely standard deep network generative models stack extra layers implementing loss function.",18
"optimistic robust optimization applications machine learning. robust optimization traditionally taken pessimistic, worst-case viewpoint uncertainty motivated desire find sets optimal policies maintain feasibility variety operating conditions. paper, explore optimistic, best-case view uncertainty show fruitful approach. show techniques used address wide variety problems. first, apply methods context robust linear programming, providing method reducing conservatism intuitive ways encode economically realistic modeling assumptions. second, look problems machine learning find approach strongly connected existing literature. specifically, provide new interpretation popular sparsity inducing non-convex regularization schemes. additionally, show successful approaches dealing outliers noise interpreted optimistic robust optimization problems. although many problems resulting approach non-convex, find dca dca-like optimization approaches intuitive efficient.",18
"algebraic approach phase retrieval explicit inversion identifiability threshold. study phase retrieval magnitude measurements unknown signal algebraic estimation problem. indeed, phase retrieval rank-one general linear measurements treated algebraic way. verified certain number generic rank-one generic linear measurements sufficient enable signal reconstruction generic signals, slightly generic measurements yield reconstructability signals. results solve open problems stated recent literature. furthermore, show algebraic estimation problem solved closed-form algebraic estimation technique, termed ideal regression, providing non-asymptotic success guarantees.",10
"stdp allows close-to-optimal spatiotemporal spike pattern detection single coincidence detector neurons. recording multiple cells simultaneously, electrophysiologists found evidence repeating spatiotemporal spike patterns, carry information. information extracted downstream neurons unclear. theoretical paper, investigate extent single cell could detect given spike pattern optimal parameters are, particular membrane time constant $\tau$. using leaky integrate-and-fire (lif) neuron instantaneous synapses homogeneous poisson input, able compute optimum analytically. results indicate relatively small $\tau$ (at tens ms) usually optimal, even pattern much longer. somewhat counter intuitive resulting detector ignores pattern, due fast memory decay. next, wondered spike-timing-dependent plasticity (stdp) could enable neuron reach theoretical optimum. simulated lif neuron equipped additive spike-timing-dependent potentiation homeostatic rate-based depression, repeatedly exposed given input spike pattern. previous studies, lif progressively became selective repeating pattern supervision, even pattern embedded poisson activity. show that, using certain stdp parameters, resulting pattern detector optimal. taken together, results may explain humans learn repeating visual auditory sequences. long sequences could recognized thanks coincidence detectors working much shorter timescale. consistent fact recognition still possible sound sequence compressed, played backward, scrambled using 10ms bins. coincidence detection simple yet powerful mechanism, could main function neurons brain.",3
"omega model human detection counting application smart surveillance system. driven significant advancements technology social issues security management, strong need smart surveillance system society today. one key features smart surveillance system efficient human detection counting system decide label events own. paper propose new, novel robust model, omega model, detecting counting human beings present scene. proposed model employs set four distinct descriptors identifying unique features head, neck shoulder regions person. unique head neck shoulder signature given omega model exploits challenges inter person variations size shape peoples head, neck shoulder regions achieve robust detection human beings even partial occlusion, dynamically changing background varying illumination conditions. experimentation observe analyze influences four descriptors system performance computation speed conclude weight based decision making system produces best results. evaluation results number images indicate validation method actual situation.",3
"learning criticality embodied boltzmann machine. many biological cognitive systems operate deep one regime activity. instead, exploit critical surfaces poised transitions parameter space. pervasiveness criticality natural systems suggests may general principles inducing behaviour. however, lack conceptual models explaining embodied agents propel towards critical points. paper, present learning model driving embodied boltzmann machine towards critical behaviour maximizing heat capacity network. test corroborate model implementing embodied agent mountain car benchmark, controlled boltzmann machine adjust weights according model. find neural controller reaches point criticality, coincides transition point behaviour agent two regimes behaviour, maximizing synergistic information sensors hidden motor neurons. finally, discuss potential learning model study contribution criticality behaviour embodied living systems scenarios necessarily constrained biological restrictions examples criticality find nature.",12
"factored latent analysis far-field tracking data. paper uses factored latent analysis (fla) learn factorized, segmental representation observations tracked objects time. factored latent analysis latent class analysis observation space subdivided aspect original space represented separate latent class model. one could simply treat factors completely independent ignore interdependencies one could concatenate together attempt learn latent class structure complete observation space. alternatively, fla allows interdependencies exploited estimating effective model, also capable representing factored latent state. paper, fla used learn set factored latent classes represent different modalities observations tracked objects. different characteristics state tracked objects represented separate latent class models, including normalized size, normalized speed, normalized direction, position. model also enables effective temporal segmentation sequences. method data-driven, unsupervised using pairwise observation statistics. data-driven unsupervised activity classi- fication technique exhibits good performance multiple challenging environments.",3
"circle-based eye center localization (cecl). propose improved eye center localization method based hough transform, called circle-based eye center localization (cecl) simple, robust, achieves accuracy par typically complex state-of-the-art methods. cecl method relies color shape cues distinguish iris facial structures. accuracy cecl method demonstrated comparison 15 state-of-the-art eye center localization methods five error thresholds, reported literature. cecl method achieved accuracy 80.8% 99.4% ranked first 2 5 thresholds. concluded cecl method offers attractive alternative existing methods automatic eye center localization.",3
"collisional charging low pressure range protoplanetary disks. recent years, collisional charging proposed promote growth pebbles early phases planet formation. ambient pressure protoplanetary disks spans wide range $10^{-9}$ mbar way beyond mbar. yet, experiments collisional charging material surfaces conducted earth atmospheric pressure, martian pressure generally $10^{-2}$ mbar thus far. work presents first pressure dependent charge measurements material collisions $10^{-8}$ $10^3$ mbar. strong charging occurs lowest pressure. detail, observations show strong similarity pressure dependence breakdown voltage two electrodes suggest breakdown also determines maximum charge colliding grains protoplanetary disks. conclude collisional charging occur parts protoplanetary disks relevant planet formation.",0
"grounded language learning simulated 3d world. increasingly surrounded artificially intelligent technology takes decisions executes actions behalf. creates pressing need general means communicate with, instruct guide artificial agents, human language compelling means communication. achieve scalable fashion, agents must able relate language world actions; is, understanding language must grounded embodied. however, learning grounded language notoriously challenging problem artificial intelligence research. present agent learns interpret language simulated 3d environment rewarded successful execution written instructions. trained via combination reinforcement unsupervised learning, beginning minimal prior knowledge, agent learns relate linguistic symbols emergent perceptual representations physical surroundings pertinent sequences actions. agent's comprehension language extends beyond prior experience, enabling apply familiar language unfamiliar situations interpret entirely novel instructions. moreover, speed agent learns new words increases semantic knowledge grows. facility generalising bootstrapping semantic knowledge indicates potential present approach reconciling ambiguous natural language complexity physical world.",3
"first-order non-homogeneous markov model response spiking neurons stimulated small phase-continuous signals. present first-order non-homogeneous markov model interspike-interval density continuously stimulated spiking neuron. model allows conditional interspike-interval density stationary interspike-interval density expressed products two separate functions, one describes neuron characteristics, describes signal characteristics. allows use model predict response underlying neuron model known well determined. approximation shows particularly clearly signal autocorrelations cross-correlations arise natural features interspike-interval density, particularly clear small signals moderate noise. show model simplifies design spiking neuron cross-correlation systems, describe four-neuron mutual inhibition network generates cross-correlation output two input signals.",15
"roxybot-06: stochastic prediction optimization tac travel. paper, describe autonomous bidding agent, roxybot, emerged victorious travel division 2006 trading agent competition photo finish. high level, design many successful trading agents summarized follows: (i) price prediction: build model market prices; (ii) optimization: solve approximately optimal set bids, given model. predict, roxybot builds stochastic model market prices simulating simultaneous ascending auctions. optimize, roxybot relies sample average approximation method, stochastic optimization technique.",3
"un résultat intrigant en commande sans modèle. elementary mathematical example proves, thanks routh-hurwitz criterion, result intriguing respect today's practical understanding model-free control, i.e., ""intelligent"" proportional controller (ip) may turn difficult tune intelligent proportional-derivative one (ipd). vast superiority ipds compared classic pids shown via computer simulations. introduction well conclusion analyse model-free control light recent advances.",3
"galaxy mass assembly (gama): probing galaxy-group correlations redshift space halo streaming model. studied galaxy-group cross-correlations redshift space galaxy mass assembly (gama) survey. use set mock gama galaxy group catalogues develop test novel 'halo streaming' model redshift-space distortions. treats 2-halo correlations via streaming model, plus empirical 1-halo term derived mocks, allowing accurate modelling nonlinear regime. order probe robustness growth rate inferred redshift-space distortions, divide galaxies colour, divide groups according total stellar mass, calibrated total mass via gravitational lensing. fit model correlation data, obtain estimates perturbation growth rate, $f\sigma_8$, validating parameter errors via dispersion different mock realizations. mocks real data, demonstrate results closely consistent different subsets group galaxy populations, considering use correlation data minimum projected radius, $r_{\rm min}$. mock data, use halo streaming model $r_{\rm min} = 5h^{-1}$ mpc, finding subsets yield growth rates within 3% other, consistent true value. actual gama data, results limited cosmic variance: $f\sigma_8=0.29\pm 0.10$ effective redshift 0.20; every reason expect method yield precise constraints larger datasets type, desi bright galaxy survey.",0
"natural images birthplace human eye. introduce database calibrated natural images publicly available easy-to-use web interface. using nikon d70 digital slr camera, acquired 5000 six-megapixel images okavango delta botswana, tropical savanna habitat similar human eye thought evolved. sequences images captured unsystematically following baboon troop, others designed vary single parameter aperture, object distance, time day position horizon. images available raw rgb format grayscale. images also available units relevant physiology human cone photoreceptors, pixel values represent expected number photoisomerizations per second cones sensitive long (l), medium (m) short (s) wavelengths. database distributed creative commons attribution-noncommercial unported license facilitate research computer vision, psychophysics perception, visual neuroscience.",15
"three-dirac-fermion approach unexpected gapless surface states van der waals magnetic topological insulators. diverse range topological quantum phenomena potential applications three-dimensional topological insulators (tis) hinge opening energy gap dirac-cone surface states. layered van der waals (vdw) topological materials, especially recently discovered mnbi$_2$te$_4$-family magnetic tis, aroused great attention, interlayer vdw gap expected play crucial role topological surface states. however, remains serious controversy whether surface states gapped gapless magnetic ti mnbi$_2$te$_4$, crucial issue prospect various magnetic topological states. here, 3-dirac-fermion approach developed generally describe surface states nonmagnetic/magnetic vdw tis interlayer vdw gap modulation. particular, apply approach solving controversial issues surface states vdw antiferromagnetic (afm) tis. remarkably, unexpected topologically protected gapless dirac-cone surface states found arise due interlayer vdw gap expansion surface, surface ferromagnetic layer zero chern number, surface states remain gapped cases. results confirmed first-principles calculations afm ti mnbi$_2$te$_4$. unexpected gapless dirac-cone states invaluable solving puzzle observed gapless surface states mnbi$_2$te$_4$. work also provides promising way experiments realize intrinsic magnetic quantum anomalous hall effect large energy gap mnbi$_2$te$_4$ films.",2
"equitability analysis maximal information coefficient, comparisons. measure dependence said equitable gives similar scores equally noisy relationships different types. equitability important data exploration goal identify relatively small set strongest associations within dataset opposed finding many non-zero associations possible, often many sift through. thus equitable statistic, maximal information coefficient (mic), useful analyzing high-dimensional data sets. here, explore equitability properties mic, discuss several aspects theory practice mic. begin presenting intuition behind equitability mic exploration maximization normalization steps definition. examine speed optimality approximation algorithm used compute mic, suggest directions improving both. finally, demonstrate range noise models sample sizes mic equitable natural alternatives, mutual information estimation distance correlation.",3
"rayleigh-taylor instability multi-mode perturbation: discrete boltzmann modeling tracers. rayleigh-taylor instability (rti) multi-mode perturbation compressible flow probed via discrete boltzmann modeling (dbm) tracers. distribution tracers provides clear boundaries light heavy fluids position space. besides, position-velocity phase space offers new perspective understanding flow behavior rti intuitive geometrical correspondence. effects viscosity, acceleration, compressibility, atwood number mixing material momentum mean non-equilibrium strength interfaces investigated separately based mixedness defined tracers non-equilibrium strength defined dbm. mixedness increases viscosity early stage decreases viscosity later stage. acceleration, compressibility, atwood number show enhancement effects mixing based different mechanisms. system relaxes initial state, mean non-equilibrium strength interfaces presents initially increasing declining trend, jointly determined interface length macroscopic physical quantity gradient. conclude four factors investigated significantly affect early evolution behavior rti system, competition interface length macroscopic physical quantity gradient. results contribute understanding multi-mode rti evolutionary mechanism accompanied kinetic effects.",14
"point-wise map recovery refinement functional correspondence. since introduction shape analysis community, functional maps met considerable success due ability compactly represent dense correspondences deformable shapes, applications ranging shape matching image segmentation, exploration large shape collections. despite numerous advantages representation, however, problem converting given functional map back point-to-point map received surprisingly limited interest. paper analyze general problem point-wise map recovery arbitrary functional maps. so, rule many assumptions required currently established approach -- notably, limiting requirement input shapes nearly-isometric. devise efficient recovery process based simple probabilistic model. experiments confirm approach achieves remarkable accuracy improvements challenging cases.",3
"log-euclidean bag words human action recognition. representing videos densely extracted local space-time features recently become popular approach analysing actions. paper, tackle problem categorising human actions devising bag words (bow) models based covariance matrices spatio-temporal features, features formed histograms optical flow. since covariance matrices form special type riemannian manifold, space symmetric positive definite (spd) matrices, non-euclidean geometry taken account discriminating covariance matrices. end, propose embed spd manifolds euclidean spaces via diffeomorphism extend bow approach riemannian version. proposed bow approach takes account manifold geometry spd matrices generation codebook histograms. experiments challenging human action datasets show proposed method obtains notable improvements discrimination accuracy, comparison several state-of-the-art methods.",3
"illustrating neural model logic computations: case sherlock holmes' old maxim. natural languages express logical propositions humans able understand. illustrate fact famous text conan doyle attributed holmes: 'it old maxim mine excluded impossible, whatever remains, however improbable, must truth'. subtle logical statement usually felt evident truth. problem trying solve cognitive reason feeling. postulate accept holmes' maxim true adult brains equipped neural modules naturally perform modal logical computations.",15
"optimization sparse symmetric sets via nonmonotone projected gradient method. consider problem minimizing lipschitz differentiable function class sparse symmetric sets wide applications engineering science. problem, known accumulation point classical projected gradient (pg) method constant stepsize $1/l$ satisfies $l$-stationarity optimality condition introduced [3]. paper introduce new optimality condition stronger $l$-stationarity optimality condition. also propose nonmonotone projected gradient (npg) method problem incorporating support-changing coordintate-swapping strategies projected gradient method variable stepsizes. shown accumulation point npg satisfies new optimality condition moreover coordinatewise stationary point. suitable assumptions, show global local minimizer problem. numerical experiments conducted compare performance pg npg. computational results demonstrate npg substantially better solution quality pg, moreover, least comparable to, sometimes much faster pg terms speed.",10
"simultaneous feature learning hash coding deep neural networks. similarity-preserving hashing widely-used method nearest neighbour search large-scale image retrieval tasks. existing hashing methods, image first encoded vector hand-engineering visual features, followed another separate projection quantization step generates binary codes. however, visual feature vectors may optimally compatible coding process, thus producing sub-optimal hashing codes. paper, propose deep architecture supervised hashing, images mapped binary codes via carefully designed deep neural networks. pipeline proposed deep architecture consists three building blocks: 1) sub-network stack convolution layers produce effective intermediate image features; 2) divide-and-encode module divide intermediate image features multiple branches, encoded one hash bit; 3) triplet ranking loss designed characterize one image similar second image third one. extensive evaluations several benchmark image datasets show proposed simultaneous feature learning hash coding pipeline brings substantial improvements state-of-the-art supervised unsupervised hashing methods.",3
"approach temporal planning scheduling domains predictable exogenous events. treatment exogenous events planning practically important many real-world domains preconditions certain plan actions affected events. paper focus planning temporal domains exogenous events happen known times, imposing constraint certain actions plan must executed predefined time windows. actions durations, handling temporal constraints adds extra difficulty planning. propose approach planning domains integrates constraint-based temporal reasoning graph-based planning framework using local search. techniques implemented planner took part 4th international planning competition (ipc-4). statistical analysis results ipc-4 demonstrates effectiveness approach terms cpu-time plan quality. additional experiments show good performance temporal reasoning techniques integrated planner.",3
"ai arbitrary world. order build ai create program copes well arbitrary world. paper restrict attention one concrete world, represents game tick-tack-toe. world simple one sufficiently complicated task people cannot manage it. main difficulty world player cannot see entire internal state world build model order understand world. model offer consist final automata first order formulas.",3
"statistical machine learning approach yield curve forecasting. yield curve forecasting important problem finance. work explore use gaussian processes conjunction dynamic modeling strategy, much like kalman filter, model yield curve. gaussian processes successfully applied model functional data variety applications. gaussian process used model yield curve. hyper-parameters gaussian process model updated algorithm receives yield curve data. yield curve data typically available time series frequency one day. compare existing methods forecast yield curve proposed method. results study showed competing method (a multivariate time series method) performed well forecasting yields short term structure region yield curve, gaussian processes perform well medium long term structure regions yield curve. accuracy long term structure region yield curve important practical implications. gaussian process framework yields uncertainty probability estimates directly contrast competing methods. analysts frequently interested information. study proposed method applied yield curve forecasting, however applied model high frequency time series data data streams domains.",18
"sample complexity reinforcement learning generative model. consider problem learning optimal action-value function discounted-reward markov decision processes (mdps). prove new pac bound sample-complexity model-based value iteration algorithm presence generative model, indicates mdp n state-action pairs discount factor \gamma\in[0,1) o(n\log(n/\delta)/((1-\gamma)^3\epsilon^2)) samples required find \epsilon-optimal estimation action-value function probability 1-\delta. also prove matching lower bound \theta (n\log(n/\delta)/((1-\gamma)^3\epsilon^2)) sample complexity estimating optimal action-value function every rl algorithm. best knowledge, first matching result sample complexity estimating optimal (action-) value function upper bound matches lower bound rl terms n, \epsilon, \delta 1/(1-\gamma). also, lower bound upper bound significantly improve state-of-the-art terms 1/(1-\gamma).",3
"unsupervised adversarial learning 3d human pose 2d joint locations. task three-dimensional (3d) human pose estimation single image divided two parts: (1) two-dimensional (2d) human joint detection image (2) estimating 3d pose 2d joints. herein, focus second part, i.e., 3d pose estimation 2d joint locations. problem existing methods require either (1) 3d pose dataset (2) 2d joint locations consecutive frames taken video sequence. aim solve problems. first time, propose method learns 3d human pose without 3d datasets. method predict 3d pose 2d joint locations single image. system based generative adversarial networks, networks trained unsupervised manner. primary idea that, network predict 3d human pose correctly, 3d pose projected onto 2d plane collapse even rotated perpendicularly. evaluated performance method using human3.6m mpii dataset showed network predict 3d pose well even 3d dataset available training.",3
"joint learning distributed representations images texts. technical report provides extra details deep multimodal similarity model (dmsm) proposed (fang et al. 2015, arxiv:1411.4952). model trained via maximizing global semantic similarity images captions natural language using public microsoft coco database, consists large set images corresponding captions. learned representations attempt capture combination various visual concepts cues.",3
"towards adaptive training agent-based sparring partners fighter pilots. key requirement current generation artificial decision-makers adapt well changes unexpected situations. paper addresses situation ai aerial dog fighting, tunable parameters govern behavior, must optimize behavior respect objective function evaluated learned simulations. bayesian optimization gaussian process surrogate used method investigating objective function. one key benefit optimization, gaussian process learns global estimate true objective function, predicted outcomes statistical measure confidence areas investigated yet. model objective function important able understand possible outcomes decision space; example crucial training providing feedback human pilots. however, standard bayesian optimization perform consistently provide accurate gaussian process surrogate function highly volatile objective functions. treat problems introducing novel sampling technique called hybrid repeat/multi-point sampling. technique gives ai ability learn optimum behaviors highly uncertain environment. importantly, improves reliability optimization, also creates better model entire objective surface. improved model agent equipped accurately/efficiently predict performance unexplored scenarios.",18
"group-sparse block pca explained variance. paper addresses simultneous determination goup-sparse loadings block optimization, correlated problem defining explained variance set non orthogonal components. give cases comprehensive mathematical presentation problem, leads propose i) new formulation/algorithm group-sparse block pca ii) framework definition explained variance analysis five definitions. numerical results i) confirm superiority block optimization deflation determination group-sparse loadings, importance group information available, ii) show ranking algorithms according explained variance essentially independant definition explained variance. results lead propose new optimal variance definition choice explained variance.",18
"linear time active learning algorithm link classification -- full version --. present efficient active learning algorithms link classification signed networks. algorithms motivated stochastic model edge labels obtained perturbations initial sign assignment consistent two-clustering nodes. provide theoretical analysis within model, showing achieve optimal (to whithin constant factor) number mistakes graph g = (v,e) |e| = \omega(|v|^{3/2}) querying o(|v|^{3/2}) edge labels. generally, show algorithm achieves optimality within factor o(k) querying order |v| + (|v|/k)^{3/2} edge labels. running time algorithm order |e| + |v|\log|v|.",3
"deep neural network chinese zero pronoun resolution. existing approaches chinese zero pronoun resolution overlook semantic information. zero pronouns descriptive information, results difficulty explicitly capturing semantic similarities antecedents. moreover, dealing candidate antecedents, traditional systems simply take advantage local information single candidate antecedent failing consider underlying information provided candidates global perspective. address weaknesses, propose novel zero pronoun-specific neural network, capable representing zero pronouns utilizing contextual information semantic level. addition, dealing candidate antecedents, two-level candidate encoder employed explicitly capture local global information candidate antecedents. conduct experiments chinese portion ontonotes 5.0 corpus. experimental results show approach substantially outperforms state-of-the-art method various experimental settings.",3
"string gaussian process kernels. introduce new class nonstationary kernels, derive covariance functions novel family stochastic processes refer string gaussian processes (string gps). construct string gps allow multiple types local patterns data, ensuring mild global regularity condition. paper, illustrate efficacy approach using synthetic data demonstrate model outperforms competing approaches well studied, real-life datasets exhibit nonstationary features.",18
"longitudinal face aging wild - recent deep learning approaches. face aging raised considerable attentions interest computer vision community recent years. numerous approaches ranging purely image processing techniques deep learning structures proposed literature. paper, aim give review recent developments modern deep learning based approaches, i.e. deep generative models, face aging task. structures, formulation, learning algorithms well synthesized results also provided systematic discussions. moreover, aging databases used methods learn aging process also reviewed.",3
"discovering characteristic landmarks ancient coins using convolutional networks. paper, propose novel method find characteristic landmarks ancient roman imperial coins using deep convolutional neural network models (cnns). formulate optimization problem discover class-specific regions guaranteeing specific controlled loss accuracy. analysis visualization discovered region confirms proposed method successfully find set characteristic regions per class, also discovered region consistent human expert annotations. also propose new framework recognize roman coins exploits hierarchical structure ancient roman coins using state-of-the-art classification power cnns adopted new task coin classification. experimental results show proposed framework able effectively recognize ancient roman coins. research, collected new roman coin dataset coins annotated consist observe (head) reverse (tail) images.",3
"lasso based feature selection malaria risk exposure prediction. life sciences, experts generally use empirical knowledge recode variables, choose interactions perform selection classical approach. aim work perform automatic learning algorithm variables selection lead know experts help decision simply replaced machine improve knowledge results. lasso method detect optimal subset variables estimation prediction conditions. paper, propose novel approach uses automatically variables available interactions. double cross-validation combine lasso, select best subset variables glm simple cross-validation perform predictions. algorithm assures stability consistency estimators.",18
"regularized multi-task learning multi-dimensional log-density gradient estimation. log-density gradient estimation fundamental statistical problem possesses various practical applications clustering measuring non-gaussianity. naive two-step approach first estimating density taking log-gradient unreliable accurate density estimate necessarily lead accurate log-density gradient estimate. cope problem, method directly estimate log-density gradient without density estimation explored, demonstrated work much better two-step method. objective paper improve performance direct method multi-dimensional cases. idea regard problem log-density gradient estimation dimension task, apply regularized multi-task learning direct log-density gradient estimator. experimentally demonstrate usefulness proposed multi-task method log-density gradient estimation mode-seeking clustering.",18
"identifying reasoning patterns games. present algorithm identifies reasoning patterns agents game, iteratively examining graph structure multi-agent influence diagram (maid) representation. decision agent participates reasoning patterns, effectively ignore decision purpose calculating nash equilibrium game. cases, lead exponential time savings process equilibrium calculation. moreover, algorithm used enumerate reasoning patterns game, useful constructing effective computerized agents interacting humans.",3
"hierarchical implicit models likelihood-free variational inference. implicit probabilistic models flexible class models defined simulation process data. form basis theories encompass understanding physical world. despite fundamental nature, use implicit models remains limited due challenges specifying complex latent structure them, performing inferences models large data sets. paper, first introduce hierarchical implicit models (hims). hims combine idea implicit densities hierarchical bayesian modeling, thereby defining models via simulators data rich hidden structure. next, develop likelihood-free variational inference (lfvi), scalable variational inference algorithm hims. key lfvi specifying variational family also implicit. matches model's flexibility allows accurate approximation posterior. demonstrate diverse applications: large-scale physical simulator predator-prey populations ecology; bayesian generative adversarial network discrete data; deep implicit model text generation.",18
"gradient descent spiking neural networks. much studies neural computation based network models static neurons produce analog output, despite fact information processing brain predominantly carried dynamic neurons produce discrete pulses called spikes. research spike-based computation impeded lack efficient supervised learning algorithm spiking networks. here, present gradient descent method optimizing spiking network models introducing differentiable formulation spiking networks deriving exact gradient calculation. demonstration, trained recurrent spiking networks two dynamic tasks: one requires optimizing fast (~millisecond) spike-based interactions efficient encoding information, delayed memory xor task extended duration (~second). results show method indeed optimizes spiking network dynamics time scale individual spikes well behavioral time scales. conclusion, result offers general purpose supervised learning algorithm spiking neural networks, thus advancing investigations spike-based computation.",15
"coordinate descent methods symmetric nonnegative matrix factorization. given symmetric nonnegative matrix $a$, symmetric nonnegative matrix factorization (symnmf) problem finding nonnegative matrix $h$, usually much fewer columns $a$, $a \approx hh^t$. symnmf used data analysis particular various clustering tasks. paper, propose simple efficient coordinate descent schemes solve problem, handle large sparse input matrices. effectiveness methods illustrated synthetic real-world data sets, show perform favorably compared recent state-of-the-art methods.",3
"incentives privacy tradeoff community sensing. community sensing, fusing information populations privately-held sensors, presents great opportunity create efficient cost-effective sensing applications. yet, reasonable privacy concerns often limit access data streams. systems valuate negotiate access private information, example return monetary incentives? optimally choose participants large population strategic users privacy concerns, compensate information shared? paper, address questions present novel mechanism, seqtgreedy, budgeted recruitment participants community sensing. first show privacy tradeoffs community sensing cast adaptive submodular optimization problem. design budget feasible, incentive compatible (truthful) mechanism adaptive submodular maximization, achieves near-optimal utility large class sensing applications. mechanism general, independent interest. demonstrate effectiveness approach case study air quality monitoring, using data collected mechanical turk platform. compared state art, approach achieves 30% reduction cost order achieve desired level utility.",3
"unsupervised method estimating global horizontal irradiance photovoltaic power measurements. paper, present method determine global horizontal irradiance (ghi) power measurements one pv systems, located neighborhood. method completely unsupervised based physical model pv plant. precise assessment solar irradiance pivotal forecast electric power generated photovoltaic (pv) plants. however, on-ground measurements expensive generally performed small medium-sized pv plants. satellite-based services represent valid alternative site measurements, space-time resolution limited. results two case studies located switzerland presented. performance proposed method assessing ghi compared free commercial satellite services. results show presented method generally better satellite-based services, especially high temporal resolutions.",18
"robustness neural network. development neural networks based machine learning usage mission critical applications, voices rising \textit{black box} aspect neural networks becomes crucial understand limits capabilities. rise neuromorphic hardware, even critical understand neural network, distributed system, tolerates failures computing nodes, neurons, communication channels, synapses. experimentally assessing robustness neural networks involves quixotic venture testing possible failures, possible inputs, ultimately hits combinatorial explosion first, impossibility gather possible inputs second. paper, prove upper bound expected error output subset neurons crashes. bound involves dependencies network parameters seen pessimistic average case. involves polynomial dependency lipschitz coefficient neurons activation function, exponential dependency depth layer failure occurs. back theoretical results experiments illustrating extent prediction matches dependencies network parameters robustness. results show robustness neural networks average crash estimated without need neither test network failure configurations, access training set used train network, practically impossible requirements.",18
"smarttennistv: automatic indexing tennis videos. paper, demonstrate score based indexing approach tennis videos. given broadcast tennis video (btv), index video segments scores create navigable searchable match. approach temporally segments rallies video recognizes scores segments, refining scores using knowledge tennis scoring system. finally build interface effortlessly retrieve view relevant video segments also automatically tagging segmented rallies human accessible tags 'fault' 'deuce'. efficiency approach demonstrated btv's two major tennis tournaments.",3
"large scale, large margin classification using indefinite similarity measures. despite success popular kernelized support vector machines, two major limitations: restricted positive semi-definite (psd) kernels, training complexity scales least quadratically size data. many natural measures similarity pairs samples psd e.g. invariant kernels, implicitly explicitly defined latent variable models. paper, investigate scalable approaches using indefinite similarity measures large margin frameworks. particular show normalization similarity subset data points constitutes representation suitable linear classifiers. result classifier competitive kernelized svm terms accuracy, despite better training test time complexities. experimental results demonstrate cifar-10 dataset, model equipped similarity measures invariant rigid non-rigid deformations, made 5 times sparser accurate kernelized svm using rbf kernels.",3
"crisp - reliable uncertainty estimation medical image segmentation. accurate uncertainty estimation critical need medical imaging community. variety methods proposed, direct extensions classification uncertainty estimations techniques. independent pixel-wise uncertainty estimates, often based probabilistic interpretation neural networks, take account anatomical prior knowledge consequently provide sub-optimal results many segmentation tasks. reason, propose crisp contrastive image segmentation uncertainty prediction method. core, crisp implements contrastive method learn joint latent space encodes distribution valid segmentations corresponding images. use joint latent space compare predictions thousands latent vectors provide anatomically consistent uncertainty maps. comprehensive studies performed four medical image databases involving different modalities organs underlines superiority method compared state-of-the-art approaches.",5
"linear systems multiple points rational normal curve. give closed formula dimension linear systems $\mathbb{p}^n$ assigned multiplicity arbitrary collections points lying rational normal curve degree $n$. particular give purely geometric explanation speciality linear systems, due presence certain subvarieties base locus: linear spans points, secant varieties rational normal curve joins them.",10
"personalized situation-aware multimodal route recommendations: favour algorithm. route choice multimodal networks shows considerable variation different individuals well current situational context. personalization recommendation algorithms already common many areas, e.g., online retail. however, online routing applications still provide shortest distance shortest travel-time routes only, neglecting individual preferences well current situation. aspects particular importance multimodal setting attractivity transportation modes biking crucially depends personal characteristics exogenous factors like weather. paper introduces favourite route recommendation (favour) approach provide personalized, situation-aware route proposals based three steps: first, initialization stage, user provides limited information (home location, work place, mobility options, sociodemographics) used select one small number initial profiles. second, based information, stated preference survey designed order sharpen profile. step mass preference prior used encode prior knowledge preferences class identified step one. third, subsequently profile continuously updated usage routing services. last two steps use bayesian learning techniques order incorporate information contributing individuals. favour approach presented detail tested small number survey participants. experimental results real-world dataset show favour generates better-quality recommendations w.r.t. alternative learning algorithms literature. particular definition mass preference prior initialization step two shown provide better predictions number alternatives literature.",3
"loco: distributing ridge regression random projections. propose loco, algorithm large-scale ridge regression distributes features across workers cluster. important dependencies variables preserved using structured random projections cheap compute must communicated once. show loco obtains solution close exact ridge regression solution fixed design setting. verify experimentally simulation study well application climate prediction. furthermore, show loco achieves significant speedups compared state-of-the-art distributed algorithm large-scale regression problem.",18
"factoring shape, pose, layout 2d image 3d scene. goal paper take single 2d image scene recover 3d structure terms small set factors: layout representing enclosing surfaces well set objects represented terms shape pose. propose convolutional neural network-based approach predict representation benchmark large dataset indoor scenes. experiments evaluate number practical design questions, demonstrate infer representation, quantitatively qualitatively demonstrate merits compared alternate representations.",3
"online crowdsourcing. success modern internet based platform, amazon mechanical turk, normal collect large number hand labeled samples non-experts. dawid- skene algorithm, based expectation- maximization update, widely used inferring true labels noisy crowdsourced labels. however, dawid-skene scheme requires data perform em iteration, infeasible streaming data large scale data. paper, provide online version dawid- skene algorithm requires one data frame iteration. further, prove mild conditions, online dawid-skene scheme projection converges stationary point marginal log-likelihood observed data. experiments demonstrate online dawid- skene scheme achieves state art performance comparing methods based dawid- skene scheme.",3
"workshop notes 6th international workshop acquisition, representation reasoning context logic (arcoe-logic 2014). arcoe-logic 2014, 6th international workshop acquisition, representation reasoning context logic, held co-location 19th international conference knowledge engineering knowledge management (ekaw 2014) november 25, 2014 link\""oping, sweden. notes contain five papers accepted presented workshop.",3
"testing feature relevance: harvest algorithm. feature selection high-dimensional data small proportion relevant features poses severe challenge standard statistical methods. developed new approach (harvest) straightforward apply, albeit somewhat computer-intensive. algorithm used pre-screen large number features identify potentially useful. basic idea evaluate feature context many random subsets features. harvest predicated assumption irrelevant feature add real predictive value, regardless features included subset. motivated idea, derived simple statistical test feature relevance. empirical analyses simulations produced far indicate harvest algorithm highly effective predictive analytics, science business.",18
"image annotation using multi-layer sparse coding. automatic annotation images descriptive words challenging problem vast applications areas image search retrieval. problem viewed label-assignment problem classifier dealing large set labels, i.e., vocabulary set. propose novel annotation method employs two layers sparse coding performs coarse-to-fine labeling. themes extracted training data treated coarse labels. theme set training images share common subject visual textual contents. system extracts coarse labels training test images without requiring prior knowledge. vocabulary words fine labels associated images. annotation methods achieve low recall due large number available fine labels, i.e., vocabulary words. systems also tend achieve high precision highly frequent words relatively rare words important search retrieval purposes. system outperforms various previously proposed annotation systems, also achieves symmetric response terms precision recall. system scores maintains high precision words wide range frequencies. behavior achieved intelligently reducing number available fine labels words image based coarse labels assigned it.",3
"neuralecho: self-attentive recurrent neural network unified acoustic echo suppression speech enhancement. acoustic echo cancellation (aec) plays important role full-duplex speech communication well front-end speech enhancement recognition conditions loudspeaker plays back. paper, present all-deep-learning framework implicitly estimates second order statistics echo/noise target speech, jointly solves echo noise suppression attention based recurrent neural network. proposed model outperforms state-of-the-art joint echo cancellation speech enhancement method f-t-lstm terms objective speech quality metrics, speech recognition accuracy model complexity. show model work speaker embedding better target speech enhancement furthermore develop branch automatic gain control (agc) task form all-in-one front-end speech enhancement system.",5
"reformulating situation calculus event calculus general theory stable models answer set programming. circumscription logic programs stable model semantics two well-known nonmonotonic formalisms. former served basis classical logic based action formalisms, situation calculus, event calculus temporal action logics; latter served basis family action languages, language several descendants. based discovery circumscription stable model semantics coincide class canonical formulas, reformulate situation calculus event calculus general theory stable models. also present translation turns reformulations answer set programs, efficient answer set solvers applied compute situation calculus event calculus.",3
"learning mixed multinomial logit model ordinal data. motivated generating personalized recommendations using ordinal (or preference) data, study question learning mixture multinomial logit (mnl) model, parameterized class distributions permutations, partial ordinal preference data (e.g. pair-wise comparisons). despite long standing importance across disciplines including social choice, operations research revenue management, little known question. case single mnl models (no mixture), computationally statistically tractable learning pair-wise comparisons feasible. however, even learning mixture two mnl components infeasible general. given state affairs, seek conditions feasible learn mixture model computationally statistically efficient manner. present sufficient condition well efficient algorithm learning mixed mnl models partial preferences/comparisons data. particular, mixture $r$ mnl components $n$ objects learnt using samples whose size scales polynomially $n$ $r$ (concretely, $r^{3.5}n^3(log n)^4$, $r\ll n^{2/7}$ model parameters sufficiently incoherent). algorithm two phases: first, learn pair-wise marginals component using tensor decomposition; second, learn model parameters component using rank centrality introduced negahban et al. process proving results, obtain generalization existing analysis tensor decomposition realistic regime partial information sample available.",18
measurement amplitude moiré patterns digital autostereoscopic 3d display. article presents experimental measurements amplitude moir\'e patterns digital autostereoscopic barrier-type 3d display across wide angular range small increment. period orientation moir\'e patterns also measured functions angle. simultaneous branches observed analyzed. theoretical interpretation also given. results help preventing minimizing moir\'e effect displays.,5
"spatial pyramid context-aware moving object detection tracking full motion video wide aerial motion imagery. robust fast automatic moving object detection tracking system essential characterize target object extract spatial temporal information different functionalities including video surveillance systems, urban traffic monitoring navigation, robotic. dissertation, present collaborative spatial pyramid context-aware moving object detection tracking system. proposed visual tracker composed one master tracker usually relies visual object features two auxiliary trackers based object temporal motion information called dynamically assist master tracker. spct utilizes image spatial context different level make video tracking system resistant occlusion, background noise improve target localization accuracy robustness. chose pre-selected seven-channel complementary features including rgb color, intensity spatial pyramid hog encode object color, shape spatial layout information. exploit integral histogram building block meet demands real-time performance. novel fast algorithm presented accurately evaluate spatially weighted local histograms constant time complexity using extension integral histogram method. different techniques explored efficiently compute integral histogram gpu architecture applied fast spatio-temporal median computations 3d face reconstruction texturing. proposed multi-component framework based semantic fusion motion information projected building footprint map significantly reduce false alarm rate urban scenes many tall structures. experiments extensive votc2016 benchmark dataset aerial video confirm combining complementary tracking cues intelligent fusion framework enables persistent tracking full motion video wide aerial motion imagery.",3
"partitioned shape modeling on-the-fly sparse appearance learning anterior visual pathway segmentation. mri quantification cranial nerves anterior visual pathway (avp) mri challenging due thin small size, structural variation along path, adjacent anatomic structures. segmentation pathologically abnormal optic nerve (e.g. optic nerve glioma) poses additional challenges due changes shape unpredictable locations. work, propose partitioned joint statistical shape model approach sparse appearance learning segmentation healthy pathological avp. main contributions are: (1) optimally partitioned statistical shape models avp based regional shape variations greater local flexibility statistical shape model; (2) refinement model accommodate pathological regions well areas subtle variation training model on-the-fly using initial segmentation obtained (1); (3) hierarchical deformable framework incorporate scale information partitioned shape appearance models. method, entitled pascal (partitioned shape appearance learning), evaluated 21 mri scans (15 healthy + 6 glioma cases) pediatric patients (ages 2-17). experimental results show proposed localized shape sparse appearance-based learning approach significantly outperforms segmentation approaches analysis pathological data.",3
"analyzing robustness nearest neighbors adversarial examples. motivated safety-critical applications, test-time attacks classifiers via adversarial examples recently received great deal attention. however, general lack understanding adversarial examples arise; whether originate due inherent properties data due lack training samples remains ill-understood. work, introduce theoretical framework analogous bias-variance theory understanding effects. use framework analyze robustness canonical non-parametric classifier - k-nearest neighbors. analysis shows robustness properties depend critically value k - classifier may inherently non-robust small k, robustness approaches bayes optimal classifier fast-growing k. propose novel modified 1-nearest neighbor classifier, guarantee robustness large sample limit. experiments suggest classifier may good robustness properties even reasonable data set sizes.",18
"directional statistics permutations. distributions permutations arise applications ranging multi-object tracking ranking instances. difficulty dealing distributions caused size domain, factorial number considered entities ($n!$). makes direct definition multinomial distribution permutation space impractical small $n$. work propose embedding $n!$ permutations given $n$ surface hypersphere defined $\mathbbm{r}^{(n-1)^2}$. result embedding, acquire ability define continuous distributions hypersphere benefits directional statistics. provide polynomial time projections continuous hypersphere representation $n!$-element permutation space. framework provides way use continuous directional probability densities methods developed thereof establishing densities permutations. demonstration benefits framework derive inference procedure state-space model permutations. demonstrate approach applications.",18
"fdconfig: constraint-based interactive product configurator. present constraint-based approach interactive product configuration. configurator tool fdconfig based feature models representation product domain. models directly mapped constraint satisfaction problems dealt appropriate constraint solvers. interactive configuration process user generates new constraints result configuration decisions even may retract constraints posted earlier. discuss configuration process, explain underlying techniques show optimizations.",3
"fast hybrid primal heuristic multiband robust capacitated network design multiple time periods. investigate robust multiperiod network design problem, generalization capacitated network design problem (cndp) that, besides establishing flow routing network capacity installation canonical cndp, also considers planning horizon made multiple time periods protection fluctuations traffic volumes. remedy traffic volume uncertainty, propose robust optimization model based multiband robustness (b\""using d'andreagiovanni, 2012), refinement classical gamma-robustness bertsimas sim uses system multiple deviation bands. since resulting optimization problem may prove challenging even instances moderate size solved state-of-the-art optimization solver, propose hybrid primal heuristic combines randomized fixing strategy inspired ant colony optimization, exploits information coming linear relaxations problem, exact large neighbourhood search. computational experiments set realistic instances sndlib show original heuristic run fast produce solutions extremely high quality associated low optimality gaps.",10
"identification protein coding regions genomic dna using unsupervised fmaca based pattern classifier. genes carry instructions making proteins found cell specific sequence nucleotides found dna molecules. but, regions genes code proteins may occupy small region sequence. identifying coding regions play vital role understanding genes. paper propose unsupervised fuzzy multiple attractor cellular automata (fmca) based pattern classifier identify coding region dna sequence. propose distinct k-means algorithm designing fmaca classifier simple, efficient produces accurate classifier previously obtained range different sequence lengths. experimental results confirm scalability proposed unsupervised fca based classifier handle large volume datasets irrespective number classes, tuples attributes. good classification accuracy established.",3
"modified rindler geometry. following previous idea, curved geometry proposed valid accelerated systems, minkowski space. curvature turns generated source accelerated motion. exponential factor depending $\rho$ (the coordinate along acceleration) constant length introduced metric. source stress tensor appears represent imperfect fluid zero energy density nonzero tangential pressures depend newton's constant even $\rho>>l_{p}$, $l_{p}$ planck length. komar mass proportional constant acceleration $g$ depend choice value constant $k$ exponential factor. null timelike geodesics along $\rho$ direction investigated. slight change metric leads nonzero energy density pressure along acceleration direction, energy conditions satisfied far planck world.",14
"understand scene categories objects: semantic regularized scene classifier using convolutional neural networks. scene classification fundamental perception task environmental understanding today's robotics. paper, attempted exploit use popular machine learning technique deep learning enhance scene understanding, particularly robotics applications. scene images larger diversity iconic object images, challenging deep learning methods automatically learn features scene images less samples. inspired human scene understanding based object knowledge, address problem scene classification encouraging deep neural networks incorporate object-level information. implemented regularization semantic segmentation. 5 thousand training images, opposed 2.5 million images, show proposed deep architecture achieves superior scene classification results state-of-the-art publicly available sun rgb-d dataset. addition, performance semantic segmentation, regularizer, also reaches new record refinement derived predicted scene labels. finally, apply sun rgb-d dataset trained model mobile robot captured images classify scenes university demonstrating generalization ability proposed algorithm.",3
automorphisms $\mathcal p(λ)/[λ]^{<λ}$. investigate statement ``all automorphisms $\mathcal p(\lambda)/[\lambda]^{<\lambda}$ trivial''. show implies statement regular uncountable $\lambda<2^{\aleph_0}$; statement false measurable $\lambda$ $2^\lambda=\lambda^+$; ``densely trivial'' forced (together $2^\lambda=\lambda^{++}$) inaccessible $\lambda$.,10
"monoperfcap: human performance capture monocular video. present first marker-less approach temporally coherent 3d performance capture human general clothing monocular video. approach reconstructs articulated human skeleton motion well medium-scale non-rigid surface deformations general scenes. human performance capture challenging problem due large range articulation, potentially fast motion, considerable non-rigid deformations, even multi-view data. reconstruction monocular video alone drastically challenging, since strong occlusions inherent depth ambiguity lead highly ill-posed reconstruction problem. tackle challenges novel approach employs sparse 2d 3d human pose detections convolutional neural network using batch-based pose estimation strategy. joint recovery per-batch motion allows resolve ambiguities monocular reconstruction problem based low dimensional trajectory subspace. addition, propose refinement surface geometry based fully automatically extracted silhouettes enable medium-scale non-rigid alignment. demonstrate state-of-the-art performance capture results enable exciting applications video editing free viewpoint video, previously infeasible monocular video. qualitative quantitative evaluation demonstrates approach significantly outperforms previous monocular methods terms accuracy, robustness scene complexity handled.",3
"optimal data collection informative rankings expose well-connected graphs. given graph vertices represent alternatives arcs represent pairwise comparison data, statistical ranking problem find potential function, defined vertices, gradient potential function agrees pairwise comparisons. goal paper develop method collecting data least squares estimator ranking problem maximal fisher information. approach, based experimental design, view data collection bi-level optimization problem inner problem ranking problem outer problem identify data maximizes informativeness ranking. certain assumptions, data collection problem decouples, reducing problem finding multigraphs large algebraic connectivity. reduction data collection problem graph-theoretic questions one primary contributions work. application, study yahoo! movie user rating dataset demonstrate addition small number well-chosen pairwise comparisons significantly increase fisher informativeness ranking. another application, study 2011-12 ncaa football schedule propose schedules number games significantly informative. using spectral clustering methods identify highly-connected communities within division, argue ncaa could improve notoriously poor rankings simply scheduling out-of-conference games.",18
"flag n' flare: fast linearly-coupled adaptive gradient methods. consider first order gradient methods effectively optimizing composite objective form sum smooth and, potentially, non-smooth functions. present accelerated adaptive gradient methods, called flag flare, offer best worlds. achieve optimal convergence rate attaining optimal first-order oracle complexity smooth convex optimization. additionally, adaptively non-uniformly re-scale gradient direction adapt limited curvature available conform geometry domain. show theoretically empirically that, compounding effects acceleration adaptivity, flag flare highly effective many data fitting machine learning applications.",10
"bring sextic equilateral pentagons. consider equilateral pentagons $v_1,\ldots,v_5$ euclidean plane. identify pentagons differ translation, rotation, magnification, moduli space possible shapes get oft-studied polygon space: 2-manifold $e_5$ known topologically quadruple torus (genus 4). study $e_5$ geometrically, goal conformal map terrain possible shapes. differential geometry use due gauss, though much named student riemann. manifold $e_5$ inherits riemannian metric grassmannian approach hausmann knutson, metric $e_5$ $e_5$ 240 isometries: optional reflection combined permutation order five edge vectors $v_{k+1}-v_k$ get assembled pentagon. giving $e_5$ conformal structure imposed $e_5$ yields compact riemann surface genus 4 120 automorphisms: 120 isometries preserve orientation. one riemann surface properties: bring sextic. $(e_5, e_5)$ conformally embeds hyperbolic plane, like bring sextic, repeating pattern 240 triangles, vertex angles $\frac{\pi}{2}$, $\frac{\pi}{4}$, $\frac{\pi}{5}$. conformal map realizes goal. plot pentagons map, compute initial pair isothermal coordinates $e_5$ solving beltrami equation \`a la gauss. use conformal mapping convert one isothermal triangular regions poincar\'e projection $(\frac{\pi}{2},\frac{\pi}{4},\frac{\pi}{5})$ hyperbolic triangle.",10
"sequential principal curves analysis. work includes technical details sequential principal curves analysis (spca) single document. spca unsupervised nonlinear invertible feature extraction technique. identified curvilinear features interpreted set nonlinear sensors: response sensor projection onto corresponding feature. moreover, easily tuned different optimization criteria; e.g. infomax, error minimization, decorrelation; choosing right way measure distances along curvilinear feature. even though proposed [laparra et al. neural comp. 12] shown work multiple modalities [laparra malo frontiers hum. neuro. 15], spca framework original roots nonlinear ica algorithm [malo gutierrez network 06]. later on, spca philosophy nonlinear generalization pca originated substantially faster alternatives cost introducing different constraints model. namely, principal polynomial analysis (ppa) [laparra et al. ijns 14], dimensionality reduction via regression (drr) [laparra et al. ieee tgrs 15]. report illustrates reasons developed family appropriate technical companion missing details [laparra et al., neco 12, laparra malo, front.hum.neuro. 15]. see also data, code examples dedicated sites http://isp.uv.es/spca.html http://isp.uv.es/after effects.html",18
"near-optimal bayesian active learning noisy observations. tackle fundamental problem bayesian active learning noise, need adaptively select number expensive tests order identify unknown hypothesis sampled known prior distribution. case noise-free observations, greedy algorithm called generalized binary search (gbs) known perform near-optimally. show observations noisy, perhaps surprisingly, gbs perform poorly. develop ec2, novel, greedy active learning algorithm prove competitive optimal policy, thus obtaining first competitiveness guarantees bayesian active learning noisy observations. bounds rely recently discovered diminishing returns property called adaptive submodularity, generalizing classical notion submodular set functions adaptive policies. results hold even tests non-uniform cost noise correlated. also propose effecxtive, particularly fast approximation ec2, evaluate bayesian experimental design problem involving human subjects, intended tease apart competing economic theories people make decisions uncertainty.",3
"calibrated multivariate regression application neural semantic basis discovery. propose calibrated multivariate regression method named cmr fitting high dimensional multivariate regression models. compared existing methods, cmr calibrates regularization regression task respect noise level simultaneously attains improved finite-sample performance tuning insensitiveness. theoretically, provide sufficient conditions cmr achieves optimal rate convergence parameter estimation. computationally, propose efficient smoothed proximal gradient algorithm worst-case numerical rate convergence $\co(1/\epsilon)$, $\epsilon$ pre-specified accuracy objective function value. conduct thorough numerical simulations illustrate cmr consistently outperforms high dimensional multivariate regression methods. also apply cmr solve brain activity prediction problem find competitive handcrafted model created human experts. r package \texttt{camel} implementing proposed method available comprehensive r archive network \url{http://cran.r-project.org/web/packages/camel/}.",18
"long-term recurrent convolutional networks visual recognition description. models based deep convolutional networks dominated recent image interpretation tasks; investigate whether models also recurrent, ""temporally deep"", effective tasks involving sequences, visual otherwise. develop novel recurrent convolutional architecture suitable large-scale visual learning end-to-end trainable, demonstrate value models benchmark video recognition tasks, image description retrieval problems, video narration challenges. contrast current models assume fixed spatio-temporal receptive field simple temporal averaging sequential processing, recurrent convolutional models ""doubly deep""' compositional spatial temporal ""layers"". models may advantages target concepts complex and/or training data limited. learning long-term dependencies possible nonlinearities incorporated network state updates. long-term rnn models appealing directly map variable-length inputs (e.g., video frames) variable length outputs (e.g., natural language text) model complex temporal dynamics; yet optimized backpropagation. recurrent long-term models directly connected modern visual convnet models jointly trained simultaneously learn temporal dynamics convolutional perceptual representations. results show models distinct advantages state-of-the-art models recognition generation separately defined and/or optimized.",3
"gravitational clustering. downfall many supervised learning algorithms, neural networks, inherent need large amount training data. although lot buzz big data, still problem classification small dataset. methods support vector machines, although capable dealing samples, inherently binary classifiers, need learning strategies one vs case multi-classification. presence large number classes become problematic. paper present, novel approach supervised learning method clustering. unlike traditional methods k-means, gravitational clustering require initial number clusters, automatically builds clusters, individual samples arbitrarily weighted requires samples staying resilient over-fitting.",3
"multiclass classification cervical cancer tissues hidden markov model. paper, report hidden markov model based multiclass classification cervical cancer tissues. model validated directly time series generated medium refractive index fluctuations extracted differential interference contrast images healthy different stages cancer tissues. method shows promising results multiclass classification higher accuracy.",3
"zero-shot style transfer text using recurrent neural networks. zero-shot translation task translating language pair aligned data pair provided training. work employ model creates paraphrases written style another existing text. since provide model paired examples source style target style training, call task zero-shot style transfer. herein, identify high-quality source aligned, stylistically distinct text bible versions use data train encoder/decoder recurrent neural model. also train statistical machine translation system, moses, comparison. find neural network outperforms moses established bleu pinc metrics evaluating paraphrase quality. technique widely applied due broad definition style used. example, tasks like text simplification easily viewed style transfer. corpus highly parallel 33 distinct bible versions used, human-aligned due presence chapter verse numbers within text. makes data rich source study natural language tasks.",3
"on-device scalable image-based localization. present scalable design entire on-device system large-scale urban localization. proposed design integrates compact image retrieval 2d-3d correspondence search estimate camera pose city region extensive coverage. design gps agnostic require network connection. system explores use abundant dataset: google street view (gsv). order overcome resource constraints mobile devices, carefully optimize system design every stage: use state-of-the-art image retrieval quickly locate candidate regions limit candidate 3d points; propose new hashing-based approach fast computation 2d-3d correspondences new one-many ransac accurate pose estimation. experiments conducted benchmark datasets 2d-3d correspondence search database 227k google street view (gsv) images overall system. results show 2d-3d correspondence search achieves state-of-the-art performance benchmark datasets system accurately quickly localize mobile images; median error less 4 meters processing time averagely less 10s typical mobile device.",3
"merf: morphology-based entity relational entity extraction framework arabic. rule-based techniques tools extract entities relational entities documents allow users specify desired entities using natural language questions, finite state automata, regular expressions, structured query language statements, proprietary scripts. techniques tools require expertise linguistics programming lack support arabic morphological analysis key process arabic text. work, present merf; morphology-based entity relational entity extraction framework arabic text. merf provides user-friendly interface user, basic knowledge linguistic features regular expressions, defines tag types interactively associates regular expressions defined boolean formulae. boolean formulae range matches arabic morphological features, synonymity features. users define user defined relations tuples subexpression matches associate code actions subexpressions. merf computes feature matches, regular expression matches, constructs entities relational entities user defined relations. evaluated work several case studies compared existing application-specific techniques. results show merf requires shorter development time effort compared existing techniques produces reasonably accurate results within reasonable overhead run time.",3
"curriculum dropout. dropout effective way regularizing neural networks. stochastically ""dropping out"" units certain probability discourages over-specific co-adaptations feature detectors, preventing overfitting improving network generalization. besides, dropout interpreted approximate model aggregation technique, exponential number smaller networks averaged order get powerful ensemble. paper, show using fixed dropout probability training suboptimal choice. thus propose time scheduling probability retaining neurons network. induces adaptive regularization scheme smoothly increases difficulty optimization problem. idea ""starting easy"" adaptively increasing difficulty learning problem roots curriculum learning allows one train better models. indeed, prove optimization strategy implements general curriculum scheme, gradually adding noise input intermediate feature representations within network architecture. experiments seven image classification datasets different network architectures show method, named curriculum dropout, frequently yields better generalization and, worst, performs well standard dropout method.",3
"weakly supervised salient object detection using image labels. deep learning based salient object detection recently achieved great success performance greatly outperforms unsupervised methods. however, annotating per-pixel saliency masks tedious inefficient procedure. paper, note superior salient object detection obtained iteratively mining correcting labeling ambiguity saliency maps traditional unsupervised methods. propose use combination coarse salient object activation map classification network saliency maps generated unsupervised methods pixel-level annotation, develop simple yet effective algorithm train fully convolutional networks salient object detection supervised noisy annotations. algorithm based alternately exploiting graphical model training fully convolutional network model updating. graphical model corrects internal labeling ambiguity spatial consistency structure preserving fully convolutional network helps correct cross-image semantic ambiguity simultaneously update coarse activation map next iteration. experimental results demonstrate proposed method greatly outperforms state-of-the-art unsupervised saliency detection methods comparable current best strongly-supervised methods training thousands pixel-level saliency map annotations public benchmarks.",3
"computing stereo matching cost convolutional neural network. present method extracting depth information rectified image pair. train convolutional neural network predict well two image patches match use compute stereo matching cost. cost refined cross-based cost aggregation semiglobal matching, followed left-right consistency check eliminate errors occluded regions. stereo method achieves error rate 2.61 % kitti stereo dataset currently (august 2014) top performing method dataset.",3
"kemeny's constant graph bridges. paper, determine formula kemeny's constant graph multiple bridges, terms quantities inherent subgraphs obtained upon removal bridges computed independently. formula, consider several optimization problems kemeny's constant graphs bridges, remark computational benefit formula computation kemeny's constant. finally, discuss potential applications.",10
"gans created equal? large-scale study. generative adversarial networks (gan) powerful subclass generative models. despite rich research activity leading numerous interesting gan algorithms, still hard assess algorithm(s) perform better others. conduct neutral, multi-faceted large-scale empirical study state-of-the art models evaluation measures. find models reach similar scores enough hyperparameter optimization random restarts. suggests improvements arise higher computational budget tuning fundamental algorithmic changes. overcome limitations current metrics, also propose several data sets precision recall computed. experimental results suggest future gan research based systematic objective evaluation procedures. finally, find evidence tested algorithms consistently outperforms original one.",18
"generative convolutional networks latent fingerprint reconstruction. performance fingerprint recognition depends heavily extraction minutiae points. enhancement fingerprint ridge pattern thus essential pre-processing step noticeably reduces false positive negative detection rates. particularly challenging setting fingerprint images corrupted partially missing. work, apply generative convolutional networks denoise visible minutiae predict missing parts ridge pattern. proposed enhancement approach tested pre-processing step combination several standard feature extraction methods mindtct, followed biometric comparison using mcc bozorth3. evaluate method several publicly available latent fingerprint datasets captured using different sensors.",3
"unscented bayesian optimization safe robot grasping. address robot grasp optimization problem unknown objects considering uncertainty input space. grasping unknown objects achieved using trial error exploration strategy. bayesian optimization sample efficient optimization algorithm especially suitable setups actively reduces number trials learning function optimize. fact, active object exploration strategy infants learn optimal grasps. one problem arises learning grasping policies configurations grasp parameters may sensitive error relative pose object robot end-effector. call configurations unsafe small errors grasp execution may turn good grasps bad grasps. therefore, reduce risk grasp failure, grasps planned safe areas. propose new algorithm, unscented bayesian optimization able perform sample efficient optimization taking consideration input noise find safe optima. contribution unscented bayesian optimization twofold provides new decision process drives exploration safe regions new selection procedure chooses optimal terms safety without extra analysis computational cost. contributions rooted strong theory behind unscented transformation, popular nonlinear approximation method. show advantages respect classical bayesian optimization synthetic problems realistic robot grasp simulations. results highlights method achieves optimal robust grasping policies trials selected grasps remain safe regions.",3
"incorporating prior information compressive online robust principal component analysis. consider online version robust principle component analysis (pca), arises naturally time-varying source separations video foreground-background separation. paper proposes compressive online robust pca prior information recursively separating sequences frames sparse low-rank components small set measurements. contrast conventional batch-based pca, processes frames directly, proposed method processes measurements taken frame. moreover, method efficiently incorporate multiple prior information, namely previous reconstructed frames, improve separation thereafter, update prior information next frame. utilize multiple prior information solving $n\text{-}\ell_{1}$ minimization incorporating previous sparse components using incremental singular value decomposition ($\mathrm{svd}$) exploiting previous low-rank components. also establish theoretical bounds number measurements required guarantee successful separation assumptions static slowly-changing low-rank components. using numerical experiments, evaluate bounds performance proposed algorithm. addition, apply proposed algorithm online video foreground background separation compressive measurements. experimental results show proposed method outperforms existing methods.",3
"deep reinforcement learning framework autonomous driving. reinforcement learning considered strong ai paradigm used teach machines interaction environment learning mistakes. despite perceived utility, yet successfully applied automotive applications. motivated successful demonstrations learning atari games go google deepmind, propose framework autonomous driving using deep reinforcement learning. particular relevance difficult pose autonomous driving supervised learning problem due strong interactions environment including vehicles, pedestrians roadworks. relatively new area research autonomous driving, provide short overview deep reinforcement learning describe proposed framework. incorporates recurrent neural networks information integration, enabling car handle partially observable scenarios. also integrates recent work attention models focus relevant information, thereby reducing computational complexity deployment embedded hardware. framework tested open source 3d car racing simulator called torcs. simulation results demonstrate learning autonomous maneuvering scenario complex road curvatures simple interaction vehicles.",18
"probabilistic matrix factorization automated machine learning. order achieve state-of-the-art performance, modern machine learning techniques require careful data pre-processing hyperparameter tuning. moreover, given ever increasing number machine learning models developed, model selection becoming increasingly important. automating selection tuning machine learning pipelines consisting data pre-processing methods machine learning models, long one goals machine learning community. paper, tackle meta-learning task combining ideas collaborative filtering bayesian optimization. using probabilistic matrix factorization techniques acquisition functions bayesian optimization, exploit experiments performed hundreds different datasets guide exploration space possible pipelines. experiments, show approach quickly identifies high-performing pipelines across wide range datasets, significantly outperforming current state-of-the-art.",18
"visual semantic interpretability projections high dimensional data classification tasks. number visual quality measures introduced visual analytics literature order automatically select best views high dimensional data large number candidate data projections. methods generally concentrate interpretability visualization pay little attention interpretability projection axes. paper, argue interpretability visualizations feature transformation functions crucial visual exploration high dimensional labeled data. present two-part user study examine two related orthogonal aspects interpretability. first study humans judge quality 2d scatterplots various datasets varying number classes provide comparisons ten automated measures, including number visual quality measures related measures various machine learning fields. investigate user perception interpretability mathematical expressions relate various automated measures complexity used characterize data projection functions. conclude discussion automated measures visual semantic interpretability data projections used together exploratory analysis classification tasks.",3
"deep journey content collaborative filtering. recommender systems research, algorithms often characterized either collaborative filtering (cf) content based (cb). cf algorithms trained using dataset user explicit implicit preferences cb algorithms typically based item profiles. approaches harness different data sources hence resulting recommended items generally also different. paper presents novel model serves bridge items content cf representations. introduce multiple input deep regression model predict cf latent embedding vectors items based textual description metadata. showcase effectiveness proposed model predicting cf vectors movies apps based textual descriptions. finally, show model improved incorporating metadata movie release year tags contribute higher accuracy.",3
"simplex closing probabilities directed graphs. recent work mathematical neuroscience calculated directed graph homology directed simplicial complex given brains sparse adjacency graph, called connectome. biological connectomes show abundance high-dimensional directed simplices betti-numbers viable dimensions - contrast erd\h{o}s-r\'enyi-graphs comparable size density. analysis synthetically trained connectomes reveals similar findings, raising questions graphs comparability nature origin simplices. present new method capable delivering insight emergence simplices thus simplicial abundance. approach allows easily distinguish simplex-rich connectomes different origin. method relies novel concept almost-d-simplex, is, simplex missing exactly one edge, consequently almost-d-simplex closing probability dimension. also describe fast algorithm identify almost-d-simplices given graph. applying method biological artificial data allows us identify mechanism responsible simplex emergence, suggests mechanism responsible simplex signature excitatory subnetwork statistical reconstruction mouse primary visual cortex. highly optimised code new method publicly available.",10
"reduction behavior additive cellular automata groups. class additive cellular automata (aca) finite group defined index-group $\m g$ finite field $\m f_p$ prime modulus $p$ \cite{bul_arch_1}. paper deals mainly aca infinite commutative groups direct products non commutative $p$-groups. appears abelian groups, rules initial states finite supports define behaviors restricted infinite regular series time moments become significantly simplified. particular, free abelian groups $n$ generators states $v^{[t]}$ aca rule $r$ time moments $t=p^k,k>k_0,$ viewed $||r||$ copies initial state $v^{[0]}$ moving $n$-dimensional euclidean space. behavior similar gliders j.conway's automaton {\sl life}. special infinite series time moments automata states approximate self-similar structures approximation becomes better time. infinite class $\mathrm{dhc}(\mbf s,\theta)$ non-commutative $p$-groups described particular includes quaternion dihedral $p$-groups. shown simplification behaviors takes place well direct products non-commutative groups class $\mathrm{dhc}(\mbf s,\theta)$ commutative groups. finally, automaton non-commutative group constructed behavior time moments $2^k,k\ge2,$ similar glider gun. concluded aca non-commutative groups demonstrate diverse variety behaviors comparing aca commutative groups.",12
"fast accurate surface normal integration non-rectangular domains. integration surface normals purpose computing shape surface 3d space classic problem computer vision. however, even nowadays still challenging task devise method combines flexibility work non-trivial computational domains high accuracy, robustness computational efficiency. uniting classic approach surface normal integration modern computational techniques construct solver fulfils requirements. building upon poisson integration model propose use iterative krylov subspace solver core step tackling task. method efficient, may show full potential combined suitable numerical preconditioning problem-specific initialisation. perform thorough numerical study order identify appropriate preconditioner purpose. address issue suitable initialisation propose compute initial state via recently developed fast marching integrator. detailed numerical experiments illuminate benefits novel combination. addition, show real-world photometric stereo datasets developed numerical framework flexible enough tackle modern computer vision applications.",3
"generalization error invariant classifiers. paper studies generalization error invariant classifiers. particular, consider common scenario classification task invariant certain transformations input, classifier constructed (or learned) invariant transformations. approach relies factoring input space product base space set transformations. show whereas generalization error non-invariant classifier proportional complexity input space, generalization error invariant classifier proportional complexity base space. also derive set sufficient conditions geometry base space set transformations ensure complexity base space much smaller complexity input space. analysis applies general classifiers convolutional neural networks. demonstrate implications developed theory classifiers experiments mnist cifar-10 datasets.",18
"endomorphism algebras gelfand-graev representations ii. let $g$ connected reductive group defined finite field $\mathbb{f}_q$ characteristic $p$, deligne--lusztig dual $g^\ast$. show that, $\overline{\mathbb{z}}[1/pm]$ $m$ product bad primes $g$, endomorphism ring gelfand--graev representation $g(\mathbb{f}_q)$ isomorphic grothendieck ring category finite-dimensional $\overline{\mathbb{f}}_q$-representations $g^\ast(\mathbb{f}_q)$.",10
"abnormal event detection videos using spatiotemporal autoencoder. present efficient method detecting anomalies videos. recent applications convolutional neural networks shown promises convolutional layers object detection recognition, especially images. however, convolutional neural networks supervised require labels learning signals. propose spatiotemporal architecture anomaly detection videos including crowded scenes. architecture includes two main components, one spatial feature representation, one learning temporal evolution spatial features. experimental results avenue, subway ucsd benchmarks confirm detection accuracy method comparable state-of-the-art methods considerable speed 140 fps.",3
"artwork creation cognitive architecture integrating computational creativity dual process approaches. paper proposes novel cognitive architecture (ca) computational creativity based psi model mechanisms inspired dual process theories reasoning rationality. recent years, many cognitive models focused dual process theories better describe implement complex cognitive skills artificial agents, creativity approached descriptive level. previous works described various modules cognitive architecture allows robot execute creative paintings. means dual process theories refine relevant mechanisms obtain artworks, particular explain details resolution level ca dealing different strategies access long term memory (ltm) managing interaction s1 s2 processes dual process theory. creative process involves divergent convergent processes either implicit explicit manner. leads four activities (exploratory, reflective, tacit, analytic) that, triggered urges motivations, generate creative acts. creative acts exploit ltm wm order make novel substitutions perceived image properly mixing parts pictures coming different domains. paper highlights role interaction s1 s2 processes, modulated resolution level, focuses attention creative agent broadening narrowing exploration novel solutions, even drawing solution set already made associations. example artificial painter described experimentations using robotic platform.",3
"picasso, matisse, fake? automated analysis drawings stroke level attribution authentication. paper proposes computational approach analysis strokes line drawings artists. aim developing ai methodology facilitates attribution drawings unknown authors way easy deceived forged art. methodology used based quantifying characteristics individual strokes drawings. propose novel algorithm segmenting individual strokes. designed compared different hand-crafted learned features task quantifying stroke characteristics. also propose compare different classification methods drawing level. experimented dataset 300 digitized drawings 80 thousands strokes. collection mainly consisted drawings pablo picasso, henry matisse, egon schiele, besides small number representative works artists. experiments shows proposed methodology classify individual strokes accuracy 70%-90%, aggregate drawings accuracy 80%, robust deceived fakes (with accuracy 100% detecting fakes settings).",5
"deep reinforced model abstractive summarization. attentional, rnn-based encoder-decoder models abstractive summarization achieved good performance short input output sequences. longer documents summaries however models often include repetitive incoherent phrases. introduce neural network model novel intra-attention attends input continuously generated output separately, new training method combines standard supervised word prediction reinforcement learning (rl). models trained supervised learning often exhibit ""exposure bias"" - assume ground truth provided step training. however, standard word prediction combined global sequence prediction training rl resulting summaries become readable. evaluate model cnn/daily mail new york times datasets. model obtains 41.16 rouge-1 score cnn/daily mail dataset, improvement previous state-of-the-art models. human evaluation also shows model produces higher quality summaries.",3
"introducing variable importance tradeoffs cp-nets. ability make decisions assess potential courses action corner-stone many ai applications, usually requires explicit information decision-maker preferences. many applications, preference elicitation serious bottleneck.the user either time, knowledge, expert support required specify complex multi - attribute utility functions. cases, method based intuitive, yet expressive, preference statements required. paper suggest use tcp - nets, enhancement cp - nets, tool representing, reasoning qualitative preference statements.we present motivate framework, define semantics, show used perform constrained optimization.",3
"study actor action semantic retention video supervoxel segmentation. existing methods semantic computer vision community seem unable deal explosion richness modern, open-source social video content. although sophisticated methods object detection bag-of-words models well studied, typically operate low level features ultimately suffer either scalability issues lack semantic meaning. hand, video supervoxel segmentation recently established applied large scale data processing, potentially serves intermediate representation high level video semantic extraction. supervoxels rich decompositions video content: capture object shape motion well. however, yet known supervoxel segmentation retains semantics underlying video content. paper, conduct systematic study well actor action semantics retained video supervoxel segmentation. study human observers watching supervoxel segmentation videos trying discriminate actor (human animal) action (one eight everyday actions). gather analyze large set 640 human perceptions 96 videos 3 different supervoxel scales. furthermore, conduct machine recognition experiments feature defined supervoxel segmentation, called supervoxel shape context, inspired higher order processes human perception. ultimate findings suggest significant amount semantics well retained video supervoxel segmentation used video analysis.",3
"multitask deep learning model real-time deployment embedded systems. propose approach multitask learning (mtl) make deep learning models faster lighter applications multiple tasks need solved simultaneously, particularly useful embedded, real-time systems. develop multitask model object detection semantic segmentation analyze challenges appear training. multitask network 1.6x faster, lighter uses less memory deploying single-task models parallel. conclude mtl potential give superior performance exchange complex training process introduces challenges present single-task models.",3
"gaussian process decentralized data fusion meets transfer learning large-scale distributed cooperative perception. paper presents novel gaussian process decentralized data fusion algorithms exploiting notion agent-centric support sets distributed cooperative perception large-scale environmental phenomena. overcome limitations scale existing works, proposed algorithms allow every mobile sensing agent choose different support set dynamically switch another execution encapsulating data local summary that, perhaps surprisingly, still assimilated agents' local summaries (i.e., based current choices support sets) globally consistent summary used predicting phenomenon. achieve this, propose novel transfer learning mechanism team agents capable sharing transferring information encapsulated summary based support set utilizing different support set loss theoretically bounded analyzed. alleviate issue information loss accumulating multiple instances transfer learning, propose new information sharing mechanism incorporated algorithms order achieve memory-efficient lazy transfer learning. empirical evaluation real-world datasets show algorithms outperform state-of-the-art methods.",18
"manyopt: extensible tool mixed, non-linear optimization smt solving. optimization mixed-integer non-linear programming (minlp) supports important decisions applications chemical process engineering. current solvers limited ability deductive reasoning use domain-specific theories, management integrality constraints yet exploit automated reasoning tools smt solvers. seems limit scalability reach tools practice. therefore present tool, manyopt, minlp optimization enables experimentation reduction techniques transform minlp problem feasibility checking realized smt solver. manyopt similar sat solver manysat runs specified number reduction techniques parallel get strongest result given minlp problem. tool implemented layers, may see features reduction techniques feature vectors. features inspired known minlp techniques whereas others novel specific smt. experimental results standard benchmarks demonstrate benefits approach. tool supports variety smt solvers easily extensible new features, courtesy layered structure. example, logical formulas deductive reasoning easily added constrain optimization minlp problem interest.",3
"interplay disorder point-gap topology: chiral modes, localization non-hermitian anderson skin effect one dimension. symmetry-protected spectral topology non-hermitian systems interesting manifestations dynamically anomalous chiral currents skin effect. study interplay symmetries disorder paradigmatic model spectral topology - non-reciprocal su-schrieffer-heeger model. numerically study effect disorder on-site non-reciprocal hopping terms. using real-space winding number, investigate impact disorder spectral topology anomalous chiral modes periodic boundary conditions. discover remarkable robustness chiral current disorder. value chiral current retains clean system value, independent disorder strength tracked completely real-space winding number class symmetries, class aiii, sub-lattice symmetry. class $d^\dagger$, $pt$-symmetric on-site gain loss terms, find disorder-averaged current robust winding number robust. study localization physics using inverse participation ratio local density states. disorder strength increased, mobility-edge phase finite winding appears. abrupt vanishing winding number marks transition partially localized fully localized phase. open boundary conditions, observe series transitions skin effect-partial skin effect-no skin effect phases. further, study non-hermitian anderson skin effect (nhase) different symmetry classes, system without skin effect develops disorder-driven skin effect intermediate disorder values. remarkably nhase present different classes, real-space winding number shows direct correspondence symmetries broken.",2
"taut string interpretation one-dimensional rudin-osher-fatemi model: new proof, fundamental estimate applications. new proof equivalence taut string algorithm one-dimensional rudin-osher-fatemi model presented. based duality projection theorem hilbert space, proof strictly elementary. existence uniqueness solutions denoising models follow by-products. standard convergence properties denoised signal, regularizing parameter tends zero, recalled efficient proofs provided. moreover, new fundamental bound denoised signal derived. bound implies, among things, strong convergence (in space functions bounded variation) denoised signal insignal regularization parameter vanishes. methods developed paper modified cover interesting applications isotonic regression.",5
"survey sparse representation: algorithms applications. sparse representation attracted much attention researchers fields signal processing, image processing, computer vision pattern recognition. sparse representation also good reputation theoretical research practical applications. many different algorithms proposed sparse representation. main purpose article provide comprehensive study updated review sparse representation supply guidance researchers. taxonomy sparse representation methods studied various viewpoints. example, terms different norm minimizations used sparsity constraints, methods roughly categorized five groups: sparse representation $l_0$-norm minimization, sparse representation $l_p$-norm (0$<$p$<$1) minimization, sparse representation $l_1$-norm minimization sparse representation $l_{2,1}$-norm minimization. paper, comprehensive overview sparse representation provided. available sparse representation algorithms also empirically categorized four groups: greedy strategy approximation, constrained optimization, proximity algorithm-based optimization, homotopy algorithm-based sparse representation. rationales different algorithms category analyzed wide range sparse representation applications summarized, could sufficiently reveal potential nature sparse representation theory. specifically, experimentally comparative study sparse representation algorithms presented. matlab code used paper available at: http://www.yongxu.org/lunwen.html.",3
"translation memory retrieval methods. translation memory (tm) systems one widely used translation technologies. important part tm systems matching algorithm determines translations get retrieved bank available translations assist human translator. although detailed accounts matching algorithms used commercial systems can't found literature, widely believed edit distance algorithms used. paper investigates evaluates use several matching algorithms, including edit distance algorithm believed heart modern commercial tm systems. paper presents results showing well various matching algorithms correlate human judgments helpfulness (collected via crowdsourcing amazon's mechanical turk). new algorithm based weighted n-gram precision adjusted translator length preferences consistently returns translations judged helpful translators multiple domains language pairs.",3
"phrase-based image captioning. generating novel textual description image interesting problem connects computer vision natural language processing. paper, present simple model able generate descriptive sentences given sample image. model strong focus syntax descriptions. train purely bilinear model learns metric image representation (generated previously trained convolutional neural network) phrases used described them. system able infer phrases given image sample. based caption syntax statistics, propose simple language model produce relevant descriptions given test image using phrases inferred. approach, considerably simpler state-of-the-art models, achieves comparable results two popular datasets task: flickr30k recently proposed microsoft coco.",3
"matrix waring problem -- ii. prove integers $k \geq 1$, exists constant $c_k$ depending $k$ $q > c_k$ $n \geq 1$ every matrix $m_n(\mathbb f_q)$ sum two $k$th powers.",10
"convolutional neural networks analyzed via convolutional sparse coding. convolutional neural networks (cnn) led many state-of-the-art results spanning various fields. however, clear profound theoretical understanding forward pass, core algorithm cnn, still lacking. parallel, within wide field sparse approximation, convolutional sparse coding (csc) gained increasing attention recent years. theoretical study model recently conducted, establishing reliable stable alternative commonly practiced patch-based processing. herein, propose novel multi-layer model, ml-csc, signals assumed emerge cascade csc layers. shown tightly connected cnn, much forward pass cnn fact thresholding pursuit serving ml-csc model. connection brings fresh view cnn, able attribute architecture theoretical claims uniqueness representations throughout network, stable estimation, guaranteed simple local sparsity conditions. lastly, identifying weaknesses pursuit scheme, propose alternative forward pass, connected deconvolutional, recurrent residual networks, better theoretical guarantees.",18
"deep probabilistic programming. propose edward, turing-complete probabilistic programming language. edward defines two compositional representations---random variables inference. treating inference first class citizen, par modeling, show probabilistic programming flexible computationally efficient traditional deep learning. flexibility, edward makes easy fit model using variety composable inference methods, ranging point estimation variational inference mcmc. addition, edward reuse modeling representation part inference, facilitating design rich variational models generative adversarial networks. efficiency, edward integrated tensorflow, providing significant speedups existing probabilistic systems. example, show benchmark logistic regression task edward least 35x faster stan 6x faster pymc3. further, edward incurs runtime overhead: fast handwritten tensorflow.",18
"consistency principle digitization procedure. algorithm building normal digital spaces continuous n-dimensional objects. paper considers conditions, allow preserve important topological geometric properties process digitization. purpose, introduce triplet {c,m,d} consisting continuous object c, intermediate model m, collection subregions whose union c, digital model d, intersection graph m, apply consistency principle criteria similarity order make mathematical structure consistent natural structure d. specifically, paper introduces locally centered lump collection subregions shows locally centered lump cover n-dimensional continuous manifold, digital model manifold digital normal n-dimensional space. addition, give examples locally centered lump tilings two-manifolds. propose algorithm constructing normal digital models continuous objects.",3
"ladar-based mover detection moving vehicles. detecting moving vehicles people crucial safe operation ugvs challenging cluttered, real world environments. propose registration technique enables objects robustly matched tracked, hence movers detected even high clutter. range data acquired using 2d scanning ladar moving platform. automatically clustered objects modeled using surface density function. bhattacharya similarity optimized register subsequent views object enabling good discrimination tracking, hence mover detection.",3
"knowledge-based decision model construction hierarchical diagnosis: preliminary report. numerous methods probabilistic reasoning large, complex belief decision networks currently developed. little research automating dynamic, incremental construction decision models. uniform value-driven method decision model construction proposed hierarchical complete diagnosis. hierarchical complete diagnostic reasoning formulated stochastic process modeled using influence diagrams. given observations, method creates decision models order obtain best actions sequentially locating repairing fault minimum cost. method construct decision models incrementally, interleaving probe actions model construction evaluation. method treats meta-level baselevel tasks uniformly. is, method takes decision-theoretic look control search causal pathways structural hierarchies.",3
"proceedings fourteenth conference uncertainty artificial intelligence (1998). proceedings fourteenth conference uncertainty artificial intelligence, held madison, wi, july 24-26, 1998",3
"prediction, expectation, surprise: methods, designs, study deployed traffic forecasting service. present research developing models forecast traffic flow congestion greater seattle area. research led deployment service named jambayes, actively used 2,500 users via smartphones desktop versions system. review modeling effort describe experiments probing predictive accuracy models. finally, present research building models identify current future surprises, via efforts modeling forecasting unexpected situations.",3
"comparative evaluation approximate probabilistic simulation deep neural networks accounts human physical scene understanding. humans demonstrate remarkable abilities predict physical events complex scenes. two classes models physical scene understanding recently proposed: ""intuitive physics engines"", ipes, posit people make predictions running approximate probabilistic simulations causal mental models similar nature video-game physics engines, memory-based models, make judgments based analogies stored experiences previously encountered scenes physical outcomes. versions latter recently instantiated convolutional neural network (cnn) architectures. report four experiments that, knowledge, first rigorous comparisons simulation-based cnn-based models, approaches concretely instantiated algorithms run raw image inputs produce outputs physical judgments whether stack blocks fall. approaches achieve super-human accuracy levels quantitatively predict human judgments similar degree, simulation-based models generalize novel situations ways people do, qualitatively consistent systematic perceptual illusions judgment asymmetries people show.",3
"matters transferable neural network model relation classification biomedical domain?. lack sufficient labeled data often limits applicability advanced machine learning algorithms real life problems. however efficient use transfer learning (tl) shown useful across domains. tl utilizes valuable knowledge learned one task (source task), sufficient data available, task interest (target task). biomedical clinical domain, quite common lack sufficient training data allow fully exploit machine learning models. work, present two unified recurrent neural models leading three transfer learning frameworks relation classification tasks. systematically investigate effectiveness proposed frameworks transferring knowledge multiple aspects related source target tasks, as, similarity relatedness source target tasks, size training data source task. empirical results show proposed frameworks general improve model performance, however improvements depend aspects related source target tasks. dependence finally determine choice particular tl framework.",3
"demystifying inferential models: fiducial perspective. inferential models recently gained popularity valid uncertainty quantification. paper, investigate inferential models exploring relationships inferential models, fiducial inference, confidence curves. short, argue certain point view, inferential models viewed fiducial distribution based confidence curves. show probabilistic uncertainty quantification inferential models based collection sets name principle sets principle assertions.",18
"sparse multivariate factor regression. consider problem multivariate regression setting relevant predictors could shared among different responses. propose algorithm decomposes coefficient matrix product long matrix wide matrix, elastic net penalty former $\ell_1$ penalty latter. first matrix linearly transforms predictors set latent factors, second one regresses responses factors. algorithm simultaneously performs dimension reduction coefficient estimation automatically estimates number latent factors data. formulation results non-convex optimization problem, despite flexibility impose effective low-dimensional structure, difficult, even impossible, solve exactly reasonable time. specify optimization algorithm based alternating minimization three different sets updates solve non-convex problem provide theoretical results convergence optimality. finally, demonstrate effectiveness algorithm via experiments simulated real data.",18
"$\ell_1$-k-svd: robust dictionary learning algorithm simultaneous update. develop dictionary learning algorithm minimizing $\ell_1$ distortion metric data term, known robust non-gaussian noise contamination. proposed algorithm exploits idea iterative minimization weighted $\ell_2$ error. refer algorithm $\ell_1$-k-svd, dictionary atoms corresponding sparse coefficients simultaneously updated minimize $\ell_1$ objective, resulting noise-robustness. demonstrate experiments $\ell_1$-k-svd algorithm results higher atom recovery rate compared k-svd robust dictionary learning (rdl) algorithm proposed lu et al., gaussian non-gaussian noise conditions. also show that, fixed values sparsity, number dictionary atoms, data-dimension, $\ell_1$-k-svd algorithm outperforms k-svd rdl algorithms training set available small. apply proposed algorithm denoising natural images corrupted additive gaussian laplacian noise. images denoised using $\ell_1$-k-svd observed slightly higher peak signal-to-noise ratio (psnr) k-svd laplacian noise, improvement structural similarity index (ssim) significant (approximately $0.1$) lower values input psnr, indicating efficacy $\ell_1$ metric.",3
"identifying spatial relations images using convolutional neural networks. traditional approaches building large scale knowledge graph usually relied extracting information (entities, properties, relations them) unstructured text (e.g. dbpedia). recent advances convolutional neural networks (cnn) allow us shift focus learning entities relations images, build robust models require little pre-processing images. paper, present approach identify extract spatial relations (e.g., girl standing behind table) images using cnns. research addresses two specific challenges: providing insight spatial relations learned network parts image used predict relations. use pre-trained network vggnet extract features image train multi-layer perceptron (mlp) set synthetic images sun09 dataset extract spatial relations. mlp predicts spatial relations without bounding box around objects space image depicting relation. understand spatial relations represented network, heatmap overlayed image show regions deemed important network. also, analyze mlp show relationship activation consistent groups nodes prediction spatial relation. show loss groups affects networks ability identify relations.",3
"joint learning sentence embeddings relevance entailment. consider problem recognizing textual entailment within information retrieval context, must simultaneously determine relevancy well degree entailment individual pieces evidence determine yes/no answer binary natural language question. compare several variants neural networks sentence embeddings setting decision-making based evidence varying relevance. propose basic model integrate evidence entailment, show joint training sentence embeddings model relevance entailment feasible even explicit per-evidence supervision, show importance evaluating strong baselines. also demonstrate benefit carrying text comprehension model trained unrelated task small datasets. research motivated primarily new open dataset introduce, consisting binary questions news-based evidence snippets. also apply proposed relevance-entailment model similar task ranking multiple-choice test answers, evaluating preliminary dataset school test questions well standard mctest dataset, improve neural model state-of-art.",3
"collaborative deep learning fixed topology networks. significant recent interest parallelize deep learning algorithms order handle enormous growth data model sizes. advances focus model parallelization engaging multiple computing agents via using central parameter server, aspect data parallelization along decentralized computation explored sufficiently. context, paper presents new consensus-based distributed sgd (cdsgd) (and momentum variant, cdmsgd) algorithm collaborative deep learning fixed topology networks enables data parallelization well decentralized computation. framework extremely useful learning agents access local/private data communication constrained environment. analyze convergence properties proposed algorithm strongly convex nonconvex objective functions fixed diminishing step sizes using concepts lyapunov function construction. demonstrate efficacy algorithms comparison baseline centralized sgd recently proposed federated averaging algorithm (that also enables data parallelism) based benchmark datasets mnist, cifar-10 cifar-100.",18
"asymptotic behavior minimal-exploration allocation policies: almost sure, arbitrarily slow growing regret. purpose paper provide understanding structure sequential allocation (""stochastic multi-armed bandit"", mab) problem establishing probability one finite horizon bounds convergence rates sample (or ""pseudo"") regret associated two simple classes allocation policies $\pi$. slowly increasing function $g$, subject mild regularity constraints, construct two policies (the $g$-forcing, $g$-inflated sample mean) achieve measure regret order $ o(g(n))$ almost surely $n \to \infty$, bound below. additionally, almost sure upper lower bounds remainder term established. constructions herein, function $g$ effectively controls ""exploration"" classical ""exploration/exploitation"" tradeoff.",18
"crowd behavior analysis: review physics meets biology. although traits emerged mass gathering often non-deliberative, act mass impulse may lead irre- vocable crowd disasters. two-fold increase carnage crowd since past two decades spurred significant advances field computer vision, towards effective proactive crowd surveillance. computer vision stud- ies related crowd observed resonate understanding emergent behavior physics (complex systems) biology (animal swarm). studies, inspired biology physics, share surprisingly common insights, interesting contradictions. however, aspect discussion fully explored. therefore, survey provides readers review state-of-the-art methods crowd behavior analysis physics biologically inspired perspectives. provide insights comprehensive discussions broader understanding underlying prospect blending physics biology studies computer vision.",3
"convex network flow optimization structured sparsity. consider class learning problems regularized structured sparsity-inducing norm defined sum l_2- l_infinity-norms groups variables. whereas much effort put developing fast optimization techniques groups disjoint embedded hierarchy, address case general overlapping groups. end, present two different strategies: one hand, show proximal operator associated sum l_infinity-norms computed exactly polynomial time solving quadratic min-cost flow problem, allowing use accelerated proximal gradient methods. hand, use proximal splitting techniques, address equivalent formulation non-overlapping groups, higher dimension additional constraints. propose efficient scalable algorithms exploiting two strategies, significantly faster alternative approaches. illustrate methods several problems cur matrix factorization, multi-task learning tree-structured dictionaries, background subtraction video sequences, image denoising wavelets, topographic dictionary learning natural image patches.",10
"deslib: dynamic ensemble selection library python. deslib open-source python library providing implementation several dynamic selection techniques. library divided three modules: (i) dcs, containing implementation dynamic classifier selection methods (dcs); (ii) des, containing implementation dynamic ensemble selection methods (des); (iii) static, implementation static ensemble techniques. library fully documented (documentation available online read docs), high test coverage (cover.io) well code quality (landscape). documentation, code examples found github page: https://github.com/menelau/deslib.",3
"brain tumor mri image classification feature selection extraction using linear discriminant analysis. feature extraction method capturing visual content image. feature extraction process represent raw image reduced form facilitate decision making pattern classification. tried address problem classification mri brain images creating robust accurate classifier act expert assistant medical practitioners. objective paper present novel method feature selection extraction. approach combines intensity, texture, shape based features classifies tumor white matter, gray matter, csf, abnormal normal area. experiment performed 140 tumor contained brain mr images internet brain segmentation repository. proposed technique carried larger database compare previous work robust effective. pca linear discriminant analysis (lda) applied training sets. support vector machine (svm) classifier served comparison nonlinear techniques vs linear ones. pca lda methods used reduce number features used. feature selection using proposed technique beneficial analyses data according grouping class variable gives reduced feature set high classification accuracy.",3
"structural health monitoring using neural network based vibrational system identification. composite fabrication technologies provide means producing high-strength, low-weight panels, plates, spars structural components use embedded fiber optic sensors piezoelectric transducers. materials, often referred smart structures, make possible sense internal characteristics, delaminations structural degradation. effort use neural network based techniques modeling analyzing dynamic structural information recognizing structural defects. yields adaptable system gives measure structural integrity composite structures.",3
"deep generative adversarial compression artifact removal. compression artifacts arise images whenever lossy compression algorithm applied. artifacts eliminate details present original image, add noise small structures; effects make images less pleasant human eye, may also lead decreased performance computer vision algorithms object detectors. eliminate artifacts, decompressing image, required recover original image disturbed version. end, present feed-forward fully convolutional residual network model trained using generative adversarial framework. provide baseline, show model also trained optimizing structural similarity (ssim), better loss respect simpler mean squared error (mse). gan able produce images photorealistic details mse ssim based networks. moreover show approach used pre-processing step object detection case images degraded compression point state-of-the art detectors fail. task, gan method obtains better performance mse ssim trained networks.",3
"ray: distributed framework emerging ai applications. next generation ai applications continuously interact environment learn interactions. applications impose new demanding systems requirements, terms performance flexibility. paper, consider requirements present ray---a distributed system address them. ray implements dynamic task graph computation model supports task-parallel actor programming models. meet performance requirements ai applications, propose architecture logically centralizes system's control state using sharded storage system novel bottom-up distributed scheduler. experiments, demonstrate sub-millisecond remote task latencies linear throughput scaling beyond 1.8 million tasks per second. empirically validate ray speeds challenging benchmarks serves natural performant fit emerging class reinforcement learning applications algorithms.",3
"lagrangian gauss-newton-krylov solver mass- intensity-preserving diffeomorphic image registration. present efficient solver diffeomorphic image registration problems framework large deformations diffeomorphic metric mappings (lddmm). use optimal control formulation, velocity field hyperbolic pde needs found distance final state system (the transformed/transported template image) observation (the reference image) minimized. solver supports stationary non-stationary (i.e., transient time-dependent) velocity fields. transformation models, consider transport equation (assuming intensities preserved deformation) continuity equation (assuming mass-preservation). consider reduced form optimal control problem solve resulting unconstrained optimization problem using discretize-then-optimize approach. key contribution elimination pde constraint using lagrangian hyperbolic pde solver. lagrangian methods rely concept characteristic curves approximate using fourth-order runge-kutta method. also present efficient algorithm computing derivatives final state system respect velocity field. allows us use fast gauss-newton based methods. present quickly converging iterative linear solvers using spectral preconditioners render overall optimization efficient scalable. method embedded image registration framework fair and, thus, supports commonly used similarity measures regularization functionals. demonstrate potential new approach using several synthetic real world test problems 14.7 million degrees freedom.",3
mob-esp improvements probability estimation. key prerequisite optimal reasoning uncertainty intelligent systems start good class probability estimates. paper improves current best probability estimation trees (bagged-pets) also presents new ensemble-based algorithm (mob-esp). comparisons made using several benchmark datasets multiple metrics. experiments show mob-esp outputs significantly accurate class probabilities either baseline bpets algorithm enhanced version presented (eb-pets). results based metrics closely associated average accuracy predictions. mob-esp also provides much better probability rankings b-pets. paper suggests estimation techniques applied concert broader category classifiers.,3
"word network topic model: simple general solution short imbalanced texts. short text prevalent format information internet recent decades, especially development online social media, whose millions users generate vast number short messages everyday. although sophisticated signals delivered short text make promising source topic modeling, extreme sparsity imbalance brings unprecedented challenges conventional topic models like lda variants. aiming presenting simple general solution topic modeling short texts, present word co-occurrence network based model named wntm tackle sparsity imbalance simultaneously. different previous approaches, wntm models distribution topics word instead learning topics document, successfully enhance semantic density data space without importing much time space complexity. meanwhile, rich contextual information preserved word-word space also guarantees sensitivity identifying rare topics convincing quality. furthermore, employing gibbs sampling lda makes wntm easily extended various application scenarios. extensive validations short normal texts testify outperformance wntm compared baseline methods. finally also demonstrate potential precisely discovering newly emerging topics unexpected events weibo pretty early stages.",3
"clustering signed networks geometric mean laplacians. signed networks allow model positive negative relationships. analyze existing extensions spectral clustering signed networks. turns existing approaches recover ground truth clustering several situations either positive negative network structures contain noise. analysis shows problems arise existing approaches take form arithmetic mean laplacians positive negative part. solution propose use geometric mean laplacians positive negative part show outperforms existing approaches. geometric mean matrices computationally expensive, show eigenvectors geometric mean computed efficiently, leading numerical scheme sparse matrices independent interest.",18
"hybrid forecasting chaotic processes: using machine learning conjunction knowledge-based model. model-based approach forecasting chaotic dynamical systems utilizes knowledge physical processes governing dynamics build approximate mathematical model system. contrast, machine learning techniques demonstrated promising results forecasting chaotic systems purely past time series measurements system state variables (training data), without prior knowledge system dynamics. motivation paper potential machine learning filling gaps underlying mechanistic knowledge cause widely-used knowledge-based models inaccurate. thus propose general method leverages advantages two approaches combining knowledge-based model machine learning technique build hybrid forecasting scheme. potential applications approach numerous (e.g., improving weather forecasting). demonstrate test utility approach using particular illustrative version machine learning known reservoir computing, apply resulting hybrid forecaster low-dimensional chaotic system, well high-dimensional spatiotemporal chaotic system. tests yield extremely promising results hybrid technique able accurately predict much longer period time either machine-learning component model-based component alone.",3
"fastventricle: cardiac segmentation enet. cardiac magnetic resonance (cmr) imaging commonly used assess cardiac structure function. one disadvantage cmr post-processing exams tedious. without automation, precise assessment cardiac function via cmr typically requires annotator spend tens minutes per case manually contouring ventricular structures. automatic contouring lower required time per patient generating contour suggestions lightly modified annotator. fully convolutional networks (fcns), variant convolutional neural networks, used rapidly advance state-of-the-art automated segmentation, makes fcns natural choice ventricular segmentation. however, fcns limited computational cost, increases monetary cost degrades user experience production systems. combat shortcoming, developed fastventricle architecture, fcn architecture ventricular segmentation based recently developed enet architecture. fastventricle 4x faster runs 6x less memory previous state-of-the-art ventricular segmentation architecture still maintaining excellent clinical accuracy.",3
"database querying changing preferences. present formal foundation iterative incremental approach constructing evaluating preference queries. main focus query modification: query transformation approach works revising preference relation query. provide detailed analysis cases order-theoretic properties preference relation preserved revision. consider number different revision operators: union, prioritized pareto composition. also formulate algebraic laws enable incremental evaluation preference queries. finally, consider two variations basic framework: finite restrictions preference relations weak-order extensions strict partial order preference relations.",3
"flower categorization using deep convolutional neural networks. developed deep learning network classification different flowers. this, used visual geometry group's 102 category flower dataset 8189 images 102 different flowers university oxford. method basically divided two parts; image segmentation classification. compared performance two different convolutional neural network architectures googlenet alexnet classification purpose. keeping hyper parameters architectures, found top 1 top 5 accuracies googlenet 47.15% 69.17% respectively whereas top 1 top 5 accuracies alexnet 43.39% 68.68% respectively. results extremely good compared random classification accuracy 0.98%. method classification flowers implemented real time applications used help botanists research well camping enthusiasts.",3
"comprehensive evaluation machine learning techniques cancer class prediction based microarray data. prostate cancer among common cancer males heterogeneity well known. early detection helps making therapeutic decision. standard technique procedure yet full-proof predicting cancer class. genomic level changes detected gene expression data changes may serve standard model random cancer data class prediction. various techniques implied prostate cancer data set order accurately predict cancer class including machine learning techniques. huge number attributes number sample microarray data leads poor machine learning, therefore challenging part attribute reduction non significant gene reduction. work compared several machine learning techniques accuracy predicting cancer class. machine learning effective number attributes (genes) larger number samples rarely possible gene expression data. attribute reduction gene filtering absolutely required order make data meaningful genes participate tumor development irrelevant cancer prediction. applied combination statistical techniques inter-quartile range t-test, effective filtering significant genes minimizing noise data. done comprehensive evaluation ten state-of-the-art machine learning techniques accuracy class prediction prostate cancer. techniques, bayes network performed accuracy 94.11% followed navie bayes accuracy 91.17%. cross validate results, modified training dataset six different way found average sensitivity, specificity, precision accuracy bayes network highest among techniques used.",3
"boosting dictionary learning error codes. conventional sparse representations based dictionary learning algorithms, initial dictionaries generally assumed proper representatives system hand. however, may case, especially systems restricted random initializations. therefore, supposedly optimal state-update based improper model might lead undesired effects conveyed successive iterations. paper, propose dictionary learning method includes general feedback process codes intermediate error left less intensive initial learning attempt, adjusts sparse codes accordingly. experimental observations show additional step vastly improves rates convergence high-dimensional cases, also results better converged states case random initializations. improvements also scale lenient sparsity constraints.",3
"nuclear transient 2017gge: tidal disruption event dusty gas-rich environment awakening dormant smbh. present results dense multi-wavelength (optical/uv, ir, x-ray) follow-up campaign nuclear transient 2017gge, covering total 1698 days transient's discovery. bolometric light-curve, black body temperature radius, well broad h $\lambda$5876 emission lines evolution time, consistent tde nature. soft x-ray flare detected delay ~200 days respect optical/uv peak rapidly followed emergence broad ii $\lambda$4686 number long-lasting high ionization coronal emission lines. ir echo, resulting dust re-radiation optical/uv tde light observed x-ray flare associated near-ir spectra show transient broad feature correspondence $\lambda$10830. data well explained scenario tde occurs gas dust rich environment optical/uv, soft x-ray, ir emission different origins locations. optical emission may produced stellar debris stream collisions prior accretion disk formation, instead responsible soft x-ray flare, emitted end circularization process.",0
"dynamic poisson factorization. models recommender systems use latent factors explain preferences behaviors users respect set items (e.g., movies, books, academic papers). typically, latent factors assumed static and, given factors, observed preferences behaviors users assumed generated without order. assumptions limit explorative predictive capabilities models, since users' interests item popularity may evolve time. address this, propose dpf, dynamic matrix factorization model based recent poisson factorization model recommendations. dpf models time evolving latent factors kalman filter actions poisson distributions. derive scalable variational inference algorithm infer latent factors. finally, demonstrate dpf 10 years user click data arxiv.org, one largest repository scientific papers formidable source information behavior scientists. empirically show performance improvement static and, recently proposed, dynamic recommendation models. also provide thorough exploration inferred posteriors latent variables.",3
"latent sequence decompositions. present latent sequence decompositions (lsd) framework. lsd decomposes sequences variable lengthed output units function input sequence output sequence. present training algorithm samples valid extensions approximate decoding algorithm. experiment wall street journal speech recognition task. lsd model achieves 12.9% wer compared character baseline 14.8% wer. combined convolutional network encoder, achieve 9.6% wer.",18
"distilling knowledge neural network. simple way improve performance almost machine learning algorithm train many different models data average predictions. unfortunately, making predictions using whole ensemble models cumbersome may computationally expensive allow deployment large number users, especially individual models large neural nets. caruana collaborators shown possible compress knowledge ensemble single model much easier deploy develop approach using different compression technique. achieve surprising results mnist show significantly improve acoustic model heavily used commercial system distilling knowledge ensemble models single model. also introduce new type ensemble composed one full models many specialist models learn distinguish fine-grained classes full models confuse. unlike mixture experts, specialist models trained rapidly parallel.",18
"variational information maximisation intrinsically motivated reinforcement learning. mutual information core statistical quantity applications areas machine learning, whether training density models multiple data modalities, maximising efficiency noisy transmission channels, learning behaviour policies exploration artificial agents. learning algorithms involve optimisation mutual information rely blahut-arimoto algorithm --- enumerative algorithm exponential complexity suitable modern machine learning applications. paper provides new approach scalable optimisation mutual information merging techniques variational inference deep learning. develop approach focusing problem intrinsically-motivated learning, mutual information forms definition well-known internal drive known empowerment. using variational lower bound mutual information, combined convolutional networks handling visual input streams, develop stochastic optimisation algorithm allows scalable information maximisation empowerment-based reasoning directly pixels actions.",18
"regularized spectral clustering degree-corrected stochastic blockmodel. spectral clustering fast popular algorithm finding clusters networks. recently, chaudhuri et al. (2012) amini et al.(2012) proposed inspired variations algorithm artificially inflate node degrees improved statistical performance. current paper extends previous statistical estimation results canonical spectral clustering algorithm way removes assumption minimum degree provides guidance choice tuning parameter. moreover, results show ""star shape"" eigenvectors--a common feature empirical networks--can explained degree-corrected stochastic blockmodel extended planted partition model, two statistical models allow highly heterogeneous degrees. throughout, paper characterizes justifies several variations spectral clustering algorithm terms models.",18
"adversarial spheres. state art computer vision models shown vulnerable small adversarial perturbations input. words, images data distribution correctly classified model close visually similar misclassified image. despite substantial research interest, cause phenomenon still poorly understood remains unsolved. hypothesize counter intuitive behavior naturally occurring result high dimensional geometry data manifold. first step towards exploring hypothesis, study simple synthetic dataset classifying two concentric high dimensional spheres. dataset show fundamental tradeoff amount test error average distance nearest error. particular, prove model misclassifies small constant fraction sphere vulnerable adversarial perturbations size $o(1/\sqrt{d})$. surprisingly, train several different architectures dataset, error sets naturally approach theoretical bound. result theory, vulnerability neural networks small adversarial perturbations logical consequence amount test error observed. hope theoretical analysis simple case point way forward explore geometry complex real-world data sets leads adversarial examples.",3
"general quantum hilbert space modeling scheme entanglement. work classification scheme quantum modeling hilbert space kind composite entity violating bell's inequalities exhibiting entanglement. theoretical framework includes situations entangled states product measurements ('customary quantum situation'), also situations entangled states entangled measurements ('nonlocal box situation', 'nonlocal non-marginal box situation'). show entanglement structurally joint property states measurements. furthermore, entangled measurements enable quantum modeling situations usually believed 'beyond quantum'. results also extended pure states quantum mixtures.",17
"factorized multi-modal topic model. multi-modal data collections, corpora paired images text snippets, require analysis methods beyond single-view component topic models. continuous observations current dominant approach based extensions canonical correlation analysis, factorizing variation components shared different modalities private them. count data, multiple variants topic models attempting tie modalities together presented. these, however, lack ability learn components private one modality, consequently try force dependencies even minimally correlating modalities. work combine two approaches presenting novel hdp-based topic model automatically learns shared private topics. model shown especially useful querying contents one domain given samples other.",3
"algorithmic chaining role partial feedback online nonparametric learning. investigate contextual online learning nonparametric (lipschitz) comparison classes different assumptions losses feedback information. full information feedback lipschitz losses, design first explicit algorithm achieving minimax regret rate (up log factors). partial feedback model motivated second-price auctions, obtain algorithms lipschitz semi-lipschitz losses regret bounds improving known bounds standard bandit feedback. analysis combines novel results contextual second-price auctions novel algorithmic approach based chaining. context space euclidean, chaining approach efficient delivers even better regret bound.",18
"galactic component mapping galaxy ugc 2885 machine learning classification. automating classification galaxy components important understanding formation evolution galaxies. traditionally, larger galaxy structures spiral arms, bulge, disc classified. use machine learning (ml) pixel-by-pixel classification automatically classify galaxy components within digital imagery massive spiral galaxy ugc 2885. galaxy components include young stellar population, old stellar population, dust lanes, galaxy center, outer disc, celestial background. test three ml models: maximum likelihood classifier (mlc), random forest (rf), support vector machine (svm). use high-resolution hubble space telescope (hst) digital imagery along textural features derived hst imagery, band ratios derived hst imagery, distance layers. textural features typically used remote sensing studies useful identifying patterns within digital imagery. run ml classification models different combinations hst digital imagery, textural features, band ratios, distance layers determine useful information galaxy component classification. textural features distance layers useful galaxy component identification, svm rf models performing best. mlc model performs worse overall comparable performance svm rf circumstances. overall, models best classifying spectrally unique galaxy components including galaxy center, outer disc, celestial background. confusion occurs young stellar population, old stellar population, dust lanes. suggest experimentation textural features astronomical research small-scale galactic structures.",0
"application verification algorithm learning based neural network. paper withdrawn author due crucial accuracy error fig. 5. precise performance albnn please refer yoon et al.'s work following article. yoon, h., park, c. s., kim, j. s., & baek, j. g. (2013). algorithm learning based neural network integrating feature selection classification. expert systems applications, 40(1), 231-241. http://www.sciencedirect.com/science/article/pii/s0957417412008731",3
"kernel-based distance metric learning output space. paper present two related, kernel-based distance metric learning (dml) methods. respective models non-linearly map data original space output space, subsequent distance measurements performed output space via mahalanobis metric. dimensionality output space directly controlled facilitate learning low-rank metric. methods allow simultaneous inference associated metric mapping output space, used visualize data, output space 2- 3-dimensional. experimental results collection classification tasks illustrate advantages proposed methods traditional kernel-based dml approaches.",3
"learning compare: relation network few-shot learning. present conceptually simple, flexible, general framework few-shot learning, classifier must learn recognise new classes given examples each. method, called relation network (rn), trained end-to-end scratch. meta-learning, learns learn deep distance metric compare small number images within episodes, designed simulate few-shot setting. trained, rn able classify images new classes computing relation scores query images examples new class without updating network. besides providing improved performance few-shot learning, framework easily extended zero-shot learning. extensive experiments four datasets demonstrate simple approach provides unified effective approach two tasks.",3
"learning sets separating kernels. consider problem learning set random samples. show relevant geometric topological properties set studied analytically using concepts theory reproducing kernel hilbert spaces. new kind reproducing kernel, call separating kernel, plays crucial role study analyzed detail. prove new analytic characterization support distribution, naturally leads family provably consistent regularized learning algorithms discuss stability methods respect random sampling. numerical experiments show approach competitive, often better, state art techniques.",18
"sub-optimal multi-phase path planning: method solving rubik's revenge. rubik's revenge, 4x4x4 variant rubik's puzzles, remains date unsolved puzzle. say, method successful categorization optimally solve every one approximately $7.401 \times 10^{45}$ possible configurations. rubik's cube, rubik's revenge's predecessor (3x3x3), approximately $4.33 \times 10^{19}$ possible configurations, recently completely solved rokicki et. al, finding configuration requires 20 moves. sheer dimension rubik's revenge total configuration space, brute-force method finding optimal solutions would vain. similar methods used rokicki et. al rubik's cube, paper develop method solving arbitrary configurations rubik's revenge phases, using combination powerful algorithm known ida* useful definition distance cube space. time-series results successfully gathered, shown method far outweighs current human-solving methods used determine loose upper bounds cube space. discussion suggest method also applied puzzles proper transformations.",10
"spectral convergence rate graph laplacian. laplacian eigenvectors graph constructed data set used many spectral manifold learning algorithms diffusion maps spectral clustering. given graph constructed random sample $d$-dimensional compact submanifold $m$ $\mathbb{r}^d$, establish spectral convergence rate graph laplacian. implies consistency spectral clustering algorithm via standard perturbation argument. simple numerical study indicates necessity denoising step applying spectral algorithms.",18
"building machines learn think themselves: commentary lake et al., behavioral brain sciences, 2017. agree lake colleagues list key ingredients building humanlike intelligence, including idea model-based reasoning essential. however, favor approach centers one additional ingredient: autonomy. particular, aim toward agents build exploit internal models, minimal human hand-engineering. believe approach centered autonomous learning greatest chance success scale toward real-world complexity, tackling domains ready-made formal models available. survey several important examples progress made toward building autonomous agents humanlike abilities, highlight outstanding challenges.",3
"translating np-spec asp. np-spec language specifying problems np declarative way. despite fact semantics language given referring datalog circumscription, close asp, far existing implementations means eclipse prolog via boolean satisfiability solvers. paper, present translations np-spec various forms asp analyze them. also argue might useful incorporate certain language constructs np-spec mainstream asp.",3
"logdet rank minimization application subspace clustering. low-rank matrix desired many machine learning computer vision problems. recent studies use nuclear norm convex surrogate rank operator. however, singular values simply added together nuclear norm, thus rank may well approximated practical problems. paper, propose use log-determinant (logdet) function smooth closer, though non-convex, approximation rank obtaining low-rank representation subspace clustering. augmented lagrange multipliers strategy applied iteratively optimize logdet-based non-convex objective function potentially large-scale data. making use angular information principal directions resultant low-rank representation, affinity graph matrix constructed spectral clustering. experimental results motion segmentation face clustering data demonstrate proposed method often outperforms state-of-the-art subspace clustering algorithms.",3
"interaction matters: note non-asymptotic local convergence generative adversarial networks. motivated pursuit systematic computational algorithmic understanding generative adversarial networks (gans), present simple yet unified non-asymptotic local convergence theory smooth two-player games, subsumes several discrete-time gradient-based saddle point dynamics. analysis reveals surprising nature off-diagonal interaction term blessing curse. one hand, interaction term explains origin slow-down effect convergence simultaneous gradient ascent (sga) stable nash equilibria. hand, unstable equilibria, exponential convergence proved thanks interaction term, three modified dynamics proposed stabilize gan training: optimistic mirror descent (omd), consensus optimization (co) predictive method (pm). analysis uncovers intimate connections among stabilizing techniques, provides detailed characterization choice learning rate.",18
"decentralized gradient methods time-varying uncoordinated stepsizes: convergence analysis privacy design. decentralized optimization enables network agents cooperatively optimize overall objective function without central coordinator gaining increased attention domains diverse control, sensor networks, data mining, robotics. however, information sharing among agents decentralized optimization also discloses agents' information, undesirable even unacceptable involved data sensitive. paper proposes two gradient based decentralized optimization algorithms protect participating agents' privacy without compromising optimization accuracy incurring heavy communication/computational overhead. distinct difference differential privacy based approaches trade optimization accuracy privacy, encryption based approaches incur heavy communication computational overhead. algorithms leverage judiciously designed mixing matrix time-varying uncoordinated stepsizes enable privacy, one using diminishing stepsizes using non-diminishing stepsizes. algorithms require participating agent share one message neighboring agent iteration reach convergence exact optimal solution, contrast gradient-tracking based algorithms requiring every agent share two messages (an optimization variable auxiliary gradient-tracking variable) non-diminishing stepsizes. furthermore, algorithms guarantee privacy participating agent even information shared agent accessible adversary, scenario existing accuracy-maintaining privacy approaches fail protect privacy. simulation results confirm effectiveness proposed algorithms.",10
"new characterization probabilities bayesian networks. characterize probabilities bayesian networks terms algebraic expressions called quasi-probabilities. arrived casting bayesian networks noisy and-or-not networks, viewing subnetworks lead node arguments node. quasi-probabilities sense ""natural"" algebra bayesian networks: easily compute marginal quasi-probability node recursively, compact form; obtain joint quasi-probability set nodes multiplying marginals (using idempotent product operator). quasi-probabilities easily manipulated improve efficiency probabilistic inference. also turn representable square-wave pulse trains, joint marginal distributions computed multiplication complementation pulse trains.",3
"computational characteristics random field ising model long-range interaction. ising model widely studied class models quantum computation. paper investigate computational characteristics random field ising model (rfim) long-range interactions decays inverse polynomial distance, achieved current ion trap system. prove rfim long-range interaction embedded 2-dimensional plane, solving ground state np-complete diminishing exponent, prove 1-dimensional rfim long-range interaction efficiently approximated interaction decays fast enough.",17
"combining semantic syntactic structure language modeling. structured language models speech recognition shown remedy weaknesses n-gram models. current structured language models are, however, limited take account dependencies non-headwords. show non-headword dependencies contribute significantly improved word error rate, data-oriented parsing model trained semantically syntactically annotated data exploit dependencies. paper also contains first dop model trained means maximum likelihood reestimation procedure, solves theoretical shortcomings previous dop models.",3
"machine learning based data mining milky way filamentary structures reconstruction. present innovative method called filexsec (filaments extraction, selection classification), data mining tool developed investigate possibility refine optimize shape reconstruction filamentary structures detected consolidated method based flux derivative analysis, column-density maps computed herschel infrared galactic plane survey (hi-gal) observations galactic plane. present methodology based feature extraction module followed machine learning model (random forest) dedicated select features classify pixels input images. tests simulations real observations method appears reliable robust respect variability shape distribution filaments. cases highly defined filament structures, presented method able bridge gaps among detected fragments, thus improving shape reconstruction. preliminary ""a posteriori"" analysis derived filament physical parameters, method appears potentially able add sufficient contribution complete refine filament reconstruction.",0
"model multi-agent consensus vague uncertain beliefs. consensus formation investigated multi-agent systems agents' beliefs vague uncertain. vagueness represented third truth state meaning \emph{borderline}. combined probabilistic model uncertainty. belief combination operator proposed exploits borderline truth values enable agents conflicting beliefs reach compromise. number simulation experiments carried agents apply operator pairwise interactions, bounded confidence restriction two agents' beliefs must sufficiently consistent agreement reached. well studying consensus operator isolation also investigate scenarios agents influenced either directly indirectly state world. former conduct simulations combine consensus formation belief updating based evidence. latter investigate effect assuming closer agent's beliefs truth visible consensus building process. cases applying consensus operators results population converging single shared belief crisp certain. furthermore, simulations combine consensus formation evidential updating converge faster shared opinion closer actual state world beliefs changed result directly receiving new evidence. finally, agent interactions guided belief quality measured similarity true state world, applying consensus operator alone results population converging high quality shared belief.",3
"permutation polynomials compositional inverses. paper, prove every pp agw-pp. also extend result wan lidl permutation polynomials finite fields determine group structure. moreover, provide new general method find compositional inverses pps, new pps compositional inverses given.",10
"sparsity averaging radio-interferometric imaging. propose novel regularization method compressive imaging context compressed sensing (cs) theory coherent redundant dictionaries. natural images often complicated several types structures present once. well known piecewise smooth images exhibit gradient sparsity, images extended structures better encapsulated wavelet frames. therefore, conjecture promoting average sparsity compressibility multiple frames rather single frames extremely powerful regularization prior.",0
"accurate inference adaptive linear models. estimators computed adaptively collected data behave like non-adaptive brethren. rather, sequential dependence collection policy lead severe distributional biases persist even infinite data limit. develop general method decorrelation procedure -- w-decorrelation -- transforming bias adaptive linear regression estimators variance. method uses coarse-grained information data collection policy need access propensity scores exact knowledge policy. bound finite-sample bias variance w-estimator develop asymptotically correct confidence intervals based novel martingale central limit theorem. demonstrate empirical benefits generic w-decorrelation procedure two different adaptive data settings: multi-armed bandits autoregressive time series models.",18
"symmetry learning function approximation reinforcement learning. paper explore methods exploit symmetries ensuring sample efficiency reinforcement learning (rl), problem deserves ever increasing attention recent advances use deep networks complex rl tasks require large amount training data. introduce novel method detect symmetries using reward trails observed episodic experience prove completeness. also provide framework incorporate discovered symmetries functional approximation. finally show use potential based reward shaping especially effective symmetry exploitation mechanism. experiments various classical problems show method improves learning performance significantly utilizing symmetry information.",18
"quantification without adjustments work?. classification task predicting class labels objects based observation features. contrast, quantification defined task determining prevalences different sorts class labels target dataset. simplest approach quantification classify & count classifier optimised classification training set applied target dataset prediction class labels. case binary quantification, number predicted positive labels used estimate prevalence positive class target dataset. since performance classify & count quantification known inferior results typically subject adjustments. however, researchers recently suggested classify & count might actually work without adjustments based classifer specifically trained quantification. discuss theoretical foundation claim explore potential limitations numerical example based binormal model equal variances. order identify optimal quantifier binormal setting, introduce concept local bayes optimality. side remark, present complete proof theorem ye et al. (2012).",18
kernel diff-hash. paper presents kernel formulation recently introduced diff-hash algorithm construction similarity-sensitive hash functions. kernel diff-hash algorithm shows superior performance problem image feature descriptor matching.,3
"initialization affects stability k-means algorithm. investigate role initialization stability k-means clustering algorithm. opposed papers, consider actual k-means algorithm ignore property getting stuck local optima. interested actual clustering, costs solution. analyze different initializations lead local optimum, lead different local optima. enables us prove reasonable select number clusters based stability scores.",18
"modelling protagonist goals desires first-person narrative. many genres natural language text narratively structured, testament predilection organizing experiences narratives. broad consensus understanding narrative requires identifying tracking goals desires characters narrative outcomes. however, date, limited work computational models problem. introduce new dataset, desiredb, includes gold-standard labels identifying statements desire, textual evidence desire fulfillment, annotations whether stated desire fulfilled given evidence narrative context. report experiments tracking desire fulfillment using different methods, show lstm skip-thought model achieves f-measure 0.7 corpus.",3
"use machine translation-based approaches vietnamese diacritic restoration. paper presents empirical study two machine translation-based approaches vietnamese diacritic restoration problem, including phrase-based neural-based machine translation models. first work applies neural-based machine translation method problem gives thorough comparison phrase-based machine translation method current state-of-the-art method problem. large dataset, phrase-based approach accuracy 97.32% neural-based approach 96.15%. neural-based method slightly lower accuracy, twice faster phrase-based method terms inference speed. moreover, neural-based machine translation method much room future improvement incorporating pre-trained word embeddings collecting training data.",3
"image processing-based domain-matching simulation heteroepitaxy. heteroepitaxy functional thin films single-crystal substrates one general themes electronic materials research. here, propose algorithm based image processing rapid simulation heteroepitaxial relationships. superposition rotation various lattice plane images film substrate, automatically generated crystal structure, rapidly verified domain matching patterns. furthermore, comprehensive validation allowed us discuss domain matching multiple perspectives, mismatch, matching period, density matching lattice points. therefore, proposed algorithm contribute materials informatics, streamlining wide range materials research functional thin films.",2
"information, privacy stability adaptive data analysis. traditional statistical theory assumes analysis performed given data set selected independently data themselves. assumption breaks downs data re-used across analyses analysis performed given stage depends results earlier stages. dependency arise data used several scientific studies, single analysis consists multiple stages. draw statistically valid conclusions data re-used? focus recent active line work. high level, results show limiting information revealed earlier stages analysis controls bias introduced later stages adaptivity. review known results area highlight role information-theoretic concepts, notably several one-shot notions mutual information.",3
"global performance poor metric verifying convergence multi-agent learning. experimental verification method choice verifying stability multi-agent reinforcement learning (marl) algorithm number agents grows theoretical analysis becomes prohibitively complex. cooperative agents, ultimate goal optimize global metric, stability usually verified observing evolution global performance metric time. global metric improves eventually stabilizes, considered reasonable verification system's stability. main contribution note establishing need better experimental frameworks measures assess stability large-scale adaptive cooperative systems. show experimental case study stability global performance metric rather deceiving, hiding underlying instability system later leads significant drop performance. propose alternative metric relies agents' local policies show, experimentally, proposed metric effective (than traditional global performance metric) exposing instability marl algorithms.",3
"framework learning agents quantum environments. paper provide broad framework describing learning agents general quantum environments. analyze types classically specified environments allow quantum enhancements learning, contrasting environments quantum oracles. show whether quantum improvements possible depends internal structure quantum environment. environments constructed internal structure appropriately chosen, agent limited capacities influence internal states environment, show improvements learning times possible broad range scenarios. scenarios call luck-favoring settings. case constructed environments particularly relevant class model-based learning agents, results imply near-generic improvement.",17
"bridging gap probabilistic deterministic models: simulation study variational bayes predictive coding recurrent neural network model. current paper proposes novel variational bayes predictive coding rnn model, learn generate fluctuated temporal patterns exemplars. model learns maximize lower bound weighted sum regularization reconstruction error terms. examined weighting affect development different types information processing learning fluctuated temporal patterns. simulation results show strong weighting reconstruction term causes development deterministic chaos imitating randomness observed target sequences, strong weighting regularization term causes development stochastic dynamics imitating probabilistic processes observed targets. moreover, results indicate generalized learning emerges two extremes. paper concludes implications terms underlying neuronal mechanisms autism spectrum disorder free action.",3
"stance detection online discussions. paper describes system created detect stance online discussions. goal identify whether author comment favor given target against. approach based maximum entropy classifier, uses surface-level, sentiment domain-specific features. system originally developed detect stance english tweets. adapted process czech news commentaries.",3
learning fuse 2d 3d image cues monocular body pose estimation. recent approaches monocular 3d human pose estimation rely deep learning. typically involve regressing image either 3d joint coordinates directly 2d joint locations 3d coordinates inferred. approaches strengths weaknesses therefore propose novel architecture designed deliver best worlds performing simultaneously fusing information along way. heart framework trainable fusion scheme learns fuse information optimally instead hand-designed. yields significant improvements upon state-of-the-art standard 3d human pose estimation benchmarks.,3
"variable selection individualized treatment rules discrete outcomes. individualized treatment rule (itr) decision rule aims improve individual patients health outcomes recommending optimal treatments according patients specific information. observational studies, collected data may contain many variables irrelevant making treatment decisions. including available variables statistical model itr could yield loss efficiency unnecessarily complicated treatment rule, difficult physicians interpret implement. thus, data-driven approach select important tailoring variables aim improving estimated decision rules crucial. growing body literature selecting variables itrs continuous outcomes, relatively methods exist discrete outcomes, pose additional computational challenges even absence variable selection. paper, propose variable selection method itrs discrete outcomes. show theoretically empirically approach double robustness property, compares favorably competing approaches. illustrate proposed method data study adaptive web-based stress management tool identify variables relevant tailoring treatment.",18
"sprite: response model multiple choice testing. item response theory (irt) models categorical response data widely used analysis educational data, computerized adaptive testing, psychological surveys. however, irt models rely assumption categories strictly ordered assumption ordering known priori. assumptions impractical many real-world scenarios, multiple-choice exams levels incorrectness distractor categories often unknown. number results exist irt models unordered categorical data, tend restrictive modeling assumptions lead poor data fitting performance practice. furthermore, existing unordered categorical models parameters difficult interpret. work, propose novel methodology unordered categorical irt call sprite (short stochastic polytomous response item model) that: (i) analyzes ordered unordered categories, (ii) offers interpretable outputs, (iii) provides improved data fitting compared existing models. compare sprite existing item response models demonstrate efficacy synthetic real-world educational datasets.",18
"distinct word length frequencies: distributions symbol entropies. distribution frequency counts distinct words length language's vocabulary analyzed using two methods. first, look empirical distributions several languages derive distribution reasonably explains number distinct words function length. able derive frequency count, mean word length, variance word length based marginal probability letters spaces. second, based information theory, demonstrate conditional entropies also used estimate frequency distinct words given length language. addition, shown techniques also applied estimate higher order entropies using vocabulary word length.",3
"interactive learning acyclic conditional preference networks. learning user preferences, represented by, example, conditional preference networks (cp-nets), become core issue ai research. recent studies investigate learning cp-nets randomly chosen examples membership equivalence queries. assess optimality learning algorithms well better understand combinatorial structure classes cp-nets, helpful calculate certain learning-theoretic information complexity parameters. paper determines bounds exact values central information complexity parameters, namely vc dimension, (recursive) teaching dimension, self-directed learning complexity, optimal mistake bound, classes acyclic cp-nets. provide algorithm learns tree-structured cp-nets membership queries. using results complexity parameters, assess optimality algorithm well another query learning algorithm acyclic cp-nets presented literature. algorithm near-optimal, can, certain assumptions adapted case membership oracle faulty.",3
"preferred well-founded semantics logic programming alternating fixpoints: preliminary report. analyze problem defining well-founded semantics ordered logic programs within general framework based alternating fixpoint theory. start showing generalizations existing answer set approaches preference weak setting well-founded semantics. specify informal yet intuitive criteria propose semantical framework preference handling suitable defining well-founded semantics ordered logic programs. suitability new approach convinced fact many attractive properties satisfied semantics. particular, semantics still correct respect various existing answer sets semantics successfully overcomes weakness generalization well-founded semantics. finally, indicate existing preferred well-founded semantics captured within semantical framework.",3
"automatic lexical semantic classification nouns. work present addresses cue-based noun classification english spanish. main objective automatically acquire lexical semantic information classifying nouns previously known noun lexical classes. achieved using particular aspects linguistic contexts cues identify specific lexical class. concentrate task identifying cues theoretical background allows assessment complexity task. results show that, despite a-priori complexity task, cue-based classification useful tool automatic acquisition lexical semantic classes.",3
"smoothed analysis tensor decompositions. low rank tensor decompositions powerful tool learning generative models, uniqueness results give significant advantage matrix decomposition methods. however, tensors pose significant algorithmic challenges tensors analogs much matrix algebra toolkit unlikely exist hardness results. efficient decomposition overcomplete case (where rank exceeds dimension) particularly challenging. introduce smoothed analysis model studying questions develop efficient algorithm tensor decomposition highly overcomplete case (rank polynomial dimension). setting, show algorithm robust inverse polynomial error -- crucial property applications learning since allowed polynomial number samples. algorithms known exact tensor decomposition overcomplete settings, main contribution analyzing stability framework smoothed analysis. main technical contribution show tensor products perturbed vectors linearly independent robust sense (i.e. associated matrix singular values least inverse polynomial). key result paves way applying tensor methods learning problems smoothed setting. particular, use obtain results learning multi-view models mixtures axis-aligned gaussians many ""components"" dimensions. assumption model adversarially chosen, formalized perturbation model parameters. believe appealing way analyze realistic instances learning problems, since framework allows us overcome many usual limitations using tensor methods.",3
"projectile motion medium quadratic drag constant horizontal wind. classic problem motion projectile thrown angle horizon studied. air resistance force taken account use quadratic resistance law. force wind also taken account, assumed constant inclined constant angle horizon. equation hodograph projectile velocity obtained taking account wind force. comparatively simple analytical approximations proposed main variables motion (cartesian projectile coordinates time). obtained formulas contain elementary functions. proposed formulas universal, is, used initial conditions throwing. addition, acceptable accuracy wide range change parameters. motion baseball, tennis ball shuttlecock badminton presented examples. calculation results show good agreement proposed analytical solutions numerical solutions. proposed analytical formulas useful researchers classical problem",14
"bet independence. study problem nonparametric dependence detection. many existing methods suffer severe power loss due non-uniform consistency, illustrate paradox. avoid power loss, approach nonparametric test independence new framework binary expansion statistics (bestat) binary expansion testing (bet), examine dependence novel binary expansion filtration approximation copula. hadamard-walsh transform, find cross interactions binary variables filtration complete sufficient statistics dependence. interactions also uncorrelated null. utilizing interactions, bet avoids problem non-uniform consistency improves upon wide class commonly used methods (a) achieving minimax rate sample size requirement specified power (b) providing clear interpretations global local relationships upon rejection independence. binary expansion approach also connects test statistics current computing system facilitate efficient bitwise implementation. illustrate bet study distribution stars night sky exploratory data analysis tcga breast cancer data.",10
"column generation interaction coverage combinatorial software testing. paper proposes novel column generation framework combinatorial software testing. particular, combines mathematical programming constraint programming hybrid decomposition generate covering arrays. approach allows generating parameterized test cases coverage guarantees parameter interactions given application. compared exhaustive testing, combinatorial test case generation reduces number tests run significantly. column generation algorithm generic accommodate mixed coverage arrays heterogeneous alphabets. algorithm realized practice cloud service recognized one five winners company-wide cloud application challenge oracle. service currently helping software developers range different product teams testing efforts exposing declarative constraint models hybrid optimization techniques broader audience.",3
"multi-scale local shape analysis feature selection machine learning applications. introduce method called multi-scale local shape analysis, mlsa, extracting features describe local structure points within dataset. method uses geometric topological features multiple levels granularity capture diverse types local information subsequent machine learning algorithms operating dataset. using synthetic real dataset examples, demonstrate significant performance improvement classification algorithms constructed datasets correspondingly augmented features.",3
"zonotope hit-and-run efficient sampling projection dpps. determinantal point processes (dpps) distributions sets items model diversity using kernels. applications machine learning include summary extraction recommendation systems. yet, cost sampling dpp prohibitive large-scale applications, triggered effort towards efficient approximate samplers. build novel mcmc sampler combines ideas combinatorial geometry, linear programming, monte carlo methods sample dpps fixed sample cardinality, also called projection dpps. sampler leverages ability hit-and-run mcmc kernel efficiently move across convex bodies. previous theoretical results yield fast mixing time chain targeting distribution close projection dpp, dpp general. empirical results demonstrate extends sampling projection dpps, i.e., sampler sample-efficient previous approaches turn translates faster convergence dealing costly-to-evaluate functions, summary extraction experiments.",18
"codebook design method noise robust speaker identification based genetic algorithm. paper, novel method designing codebook noise robust speaker identification purpose utilizing genetic algorithm proposed. wiener filter used remove background noises source speech utterances. speech features extracted using standard speech parameterization method lpc, lpcc, rcc, mfcc, (delta)mfcc (delta)(delta) mfcc. techniques, performance proposed system compared. codebook design method, genetic algorithm capability getting global optimal result hence improves quality codebook. comparing noizeous speech database, experimental result shows 79.62 percent accuracy achieved.",3
"investigating well contextual features captured bi-directional recurrent neural network models. learning algorithms natural language processing (nlp) tasks traditionally rely manually defined relevant contextual features. hand, neural network models using distributional representation words successfully applied several nlp tasks. models learn features automatically avoid explicit feature engineering. across several domains, neural models become natural choice specifically limited characteristics data known. however, flexibility comes cost interpretability. paper, define three different methods investigate ability bi-directional recurrent neural networks (rnns) capturing contextual features. particular, analyze rnns sequence tagging tasks. perform comprehensive analysis general well biomedical domain datasets. experiments focus important contextual words features, easily extended analyze various feature types. also investigate positional effects context words show developed methods used error analysis.",3
lorentzian vacuum transitions generalized uncertainty principle. vacuum transition probabilities minima scalar field potential presence gravity studied using wentzel-kramers-brillouin approximation. first propose method compute transisition probabilities solving wheeler-dewitt equation semi-classical approach model superspace contains terms squared well linear momenta hamiltonian constraint generalizing way previous results. apply method compute transition probabilities friedmann-lemaitre-robertson-walker metric positive null curvature bianchi iii metric coordinates minisuperspace obey standard uncertainty principle generalized uncertainty principle taken account. cases compare results found effect considering generalized uncertainty principle probability enhanced first decays faster correspoding scale factor big enough probability reduced. also compare effect anisotropy comparing result bianchi iii metric flat flrw metric corresponds isotropy limit comment differences previous work.,6
"toward geometric deep slam. present point tracking system powered two deep convolutional neural networks. first network, magicpoint, operates single images extracts salient 2d points. extracted points ""slam-ready"" design isolated well-distributed throughout image. compare network classical point detectors discover significant performance gap presence image noise. transformation estimation simple detected points geometrically stable, designed second network, magicwarp, operates pairs point images (outputs magicpoint), estimates homography relates inputs. transformation engine differs traditional approaches use local point descriptors, point locations. networks trained simple synthetic data, alleviating requirement expensive external camera ground truthing advanced graphics rendering pipelines. system fast lean, easily running 30+ fps single cpu.",3
"new type-ii fuzzy logic based controller non-linear dynamical systems application 3-psp parallel robot. concept uncertainty posed almost complex system including parallel robots outstanding instance dynamical robotics systems. suggested name, uncertainty, missing information beyond knowledge human thus may tend handle properly minimize side-effects control process. type-ii fuzzy logic shown superiority traditional fuzzy logic dealing uncertainty. type-ii fuzzy logic controllers however newer promising approaches recently applied various fields due significant contribution especially noise (as important instance uncertainty) emerges. design type-i fuzzy logic systems, presume almost certain fuzzy membership functions true many cases. thus t2fls realistic approach dealing practical applications might lot offer. type-ii fuzzy logic takes account higher level uncertainty, words, membership grade type-ii fuzzy variable longer crisp number rather type-i linguistic term. thesis effects uncertainty dynamic control parallel robot considered. specifically, intended incorporate type-ii fuzzy logic paradigm model based controller, so-called computed torque control method, apply result 3 degrees freedom parallel manipulator. ...",3
"high-q resonances governed quasi-bound states continuum all-dielectric metasurfaces. realization high-q resonances silicon metasurface various broken-symmetry blocks reported. theoretical analysis reveals sharp resonances metasurfaces originate symmetry-protected bound states continuum (bic) magnetic dipole dominates peculiar states. smaller size defect broken-symmetry block gives rise resonance larger q factor. importantly, relationship tuned changing structural parameter, resulting modulation topological configuration bics. consequently, q factor 3,000 easily achieved optimizing dimensions nanostructure. sharp resonance, intensity third harmonic generation signal patterned structure 368 times larger flat silicon film. proposed strategy underlying theory open new avenues realize ultrasharp resonances, may promote development potential meta-devices nonlinearity, lasing action, sensing.",14
"nested hierarchical dirichlet processes. develop nested hierarchical dirichlet process (nhdp) hierarchical topic modeling. nhdp generalization nested chinese restaurant process (ncrp) allows word follow path topic node according document-specific distribution shared tree. alleviates rigid, single-path formulation ncrp, allowing document easily express thematic borrowings random effect. derive stochastic variational inference algorithm model, addition greedy subtree selection method document, allows efficient inference using massive collections text documents. demonstrate algorithm 1.8 million documents new york times 3.3 million documents wikipedia.",18
"simulating perceptron quantum computer. perceptrons basic computational unit artificial neural networks, model activation mechanism output neuron due incoming signals neighbours. linear classifiers, play important role foundations machine learning. context emerging field quantum machine learning, several attempts made develop corresponding unit using quantum information theory. based quantum phase estimation algorithm, paper introduces quantum perceptron model imitating step-activation function classical perceptron. scheme requires resources $\mathcal{o}(n)$ (where $n$ size input) promises efficient applications complex structures trainable quantum neural networks.",17
"discriminative parameter estimation random walks segmentation. random walks (rw) algorithm one e - cient easy-to-use probabilistic segmentation methods. combining contrast terms prior terms, provides accurate segmentations medical images fully automated manner. however, one main drawbacks using rw algorithm parameters hand-tuned. propose novel discriminative learning framework estimates parameters using training dataset. main challenge face training samples fully supervised. speci cally, provide hard segmentation images, instead proba- bilistic segmentation. overcome challenge treating opti- mal probabilistic segmentation compatible given hard segmentation latent variable. allows us employ latent support vector machine formulation parameter estimation. show approach signi cantly outperforms baseline methods challenging dataset consisting real clinical 3d mri volumes skeletal muscles.",3
"random walk initialization training deep feedforward networks. training deep networks important open problem machine learning. one many difficulties norm back-propagated error gradient grow decay exponentially. show training deep feed-forward networks (ffns) difficult previously thought. unlike back-propagation applied recurrent network, application ffn amounts multiplying error gradient different random matrix layer. show successive application correctly scaled random matrices initial vector results random walk log norm resulting vectors, compute scaling makes walk unbiased. variance random walk grows linearly network depth inversely proportional size layer. practically, implies gradient whose log-norm scales square root network depth shows vanishing gradient problem mitigated increasing width layers. mathematical analyses experimental results using stochastic gradient descent optimize tasks related mnist timit datasets provided support claims. equations optimal matrix scaling provided linear relu cases.",3
"nonparametric latent tree graphical models: inference, estimation, structure learning. tree structured graphical models powerful expressing long range hierarchical dependency among many variables, widely applied different areas computer science statistics. however, existing methods parameter estimation, inference, structure learning mainly rely gaussian discrete assumptions, restrictive many applications. paper, propose new nonparametric methods based reproducing kernel hilbert space embeddings distributions recover latent tree structures, estimate parameters, perform inference high dimensional continuous non-gaussian variables. usefulness proposed methods illustrated thorough numerical results.",18
"fast clustering scalable statistical analysis structured images. use brain images markers diseases behavioral differences challenged small effects size ensuing lack power, issue incited researchers rely systematically large cohorts. coupled resolution increases, leads large datasets. striking example case brain imaging human connectome project: 20 terabytes data growing. resulting data deluge poses severe challenges regarding tractability processing steps (discriminant analysis, multivariate models) due memory demands posed data. work, revisit dimension reduction approaches, random projections, aim replacing costly function evaluations cheaper ones decreasing memory requirements. specifically, investigate use alternate schemes, based fast clustering, well suited signals exhibiting strong spatial structure, anatomical functional brain images. contribution twofold: i) propose linear-time clustering scheme bypasses percolation issues inherent algorithms thus provides compressions nearly good traditional quadratic-complexity variance-minimizing clustering schemes, ii) show cluster-based compression virtuous effect removing high-frequency noise, actually improving subsequent estimations steps. consequence, proposed approach yields accurate models several large-scale problems yet impressive gains computational efficiency, making possible analyze large datasets.",18
"feedback detection live predictors. predictor deployed live production system may perturb features uses make predictions. feedback loop occur, example, model predicts certain type behavior ends causing behavior predicts, thus creating self-fulfilling prophecy. paper analyze predictor feedback detection causal inference problem, introduce local randomization scheme used detect non-linear feedback real-world problems. conduct pilot study proposed methodology using predictive system currently deployed part search engine.",18
"learning stochastically revealed preference. study learning problem revealed preference stochastic setting: learner observes utility-maximizing actions set agents whose utility follows unknown distribution, learner aims infer distribution observations actions. problem viewed single-constraint special case inverse linear optimization problem. existing works assume agents share one common utility easily violated practical contexts. paper, consider two settings underlying utility distribution: gaussian setting customer utility follows von mises-fisher distribution, $\delta$-corruption setting customer utility distribution concentrates one fixed vector high probability arbitrarily corrupted otherwise. devise bayesian approaches parameter estimation develop theoretical guarantees recovery true parameter. illustrate algorithm performance numerical experiments.",10
"proceedings answer set programming computing paradigms (aspocp 2013), 6th international workshop, august 25, 2013, istanbul, turkey. volume contains papers presented sixth workshop answer set programming computing paradigms (aspocp 2013) held august 25th, 2013 istanbul, co-located 29th international conference logic programming (iclp 2013). thus continues series previous events co-located iclp, aiming facilitating discussion crossing boundaries current asp techniques theory, solving, applications, combination inspired computing paradigms.",3
"language, emotions, cultures: emotional sapir-whorf hypothesis. emotional version sapir-whorf hypothesis suggests differences language emotionalities influence differences among cultures less conceptual differences. conceptual contents languages cultures significant extent determined words semantic differences; could borrowed among languages exchanged among cultures. emotional differences, suggested paper, related grammar mostly cannot borrowed. conceptual emotional mechanisms languages considered along functions mind cultural evolution. fundamental contradiction human mind considered: language evolution requires reduced emotionality, ""too low"" emotionality makes language ""irrelevant life,"" disconnected sensory-motor experience. neural mechanisms processes suggested well mathematical models: knowledge instinct, language instinct, dual model connecting language cognition, dynamic logic, neural modeling fields. mathematical results related cognitive science, linguistics, psychology. experimental evidence theoretical arguments discussed. approximate equations evolution human minds cultures obtained. solutions identify three types cultures: ""conceptual""-pragmatic cultures, emotionality language reduced differentiation overtakes synthesis resulting fast evolution price uncertainty values, self doubts, internal crises; ""traditional-emotional"" cultures differentiation lags behind synthesis, resulting cultural stability price stagnation; ""multi-cultural"" societies combining fast cultural evolution stability. unsolved problems future theoretical experimental directions discussed.",3
"new descriptor glomerulus detection kidney microscopy image. glomerulus detection key step histopathological evaluation microscopy images kidneys. however, task automatic detection glomeruli poses challenges due disparity sizes shapes glomeruli renal sections. moreover, extensive variations intensities due heterogeneity immunohistochemistry staining also encountered. despite widely recognized powerful descriptor general object detection, rectangular histogram oriented gradients (rectangular hog) suffers many false positives due aforementioned difficulties context glomerulus detection. new descriptor referred segmental hog developed perform comprehensive detection hundreds glomeruli images whole kidney sections. new descriptor possesses flexible blocks adaptively fitted input images acquire robustness deformations glomeruli. moreover, novel segmentation technique employed herewith generates high quality segmentation outputs algorithm assured converge optimal solution. consequently, experiments using real world image data reveal segmental hog achieves significant improvements detection performance compared rectangular hog. proposed descriptor method glomeruli detection present promising results expected useful pathological evaluation.",3
"definition confounder. causal inference literature provided clear formal definition confounding expressed terms counterfactual independence. literature not, however, come consensus formal definition confounder, given priority concept confounding confounder. consider number candidate definitions arising various informal statements made literature. consider properties satisfied candidate definition, principally focusing (i) whether candidate definition control ""confounders"" suffices control ""confounding"" (ii) whether confounder context helps eliminate reduce confounding bias. several candidate definitions two properties. one candidate definition considered satisfies properties. propose ""confounder"" defined pre-exposure covariate c exists set covariates x effect exposure outcome unconfounded conditional (x,c) proper subset (x,c) effect exposure outcome unconfounded given subset. also provide conditional analogue definition; propose variable helps reduce bias eliminate bias referred ""surrogate confounder."" definitions closely related given robins morgenstern [comput. math. appl. 14 (1987) 869-916]. implications hold among various candidate definitions discussed.",18
"boolean dimension dim-boundedness: planar cover graph zero. 1989, ne\v{s}et\v{r}il pudl\'ak posed following challenging question: planar posets bounded boolean dimension? show every poset planar cover graph unique minimal element boolean dimension 13. consequence, able show reachability labeling scheme labels consisting $o(\log n)$ bits planar digraphs single source. best known scheme general planar digraphs uses labels $o(\log^2 n)$ bits [thorup jacm 2004], remains open determine whether scheme using labels $o(\log n)$ bits exists. boolean dimension result proved tandem second result showing dimension poset planar cover graph unique minimal element bounded linear function standard example number. however, one major challenges dimension theory determine whether dimension bounded terms standard example number posets planar cover graphs.",10
"improving capacity solving large-scale wireless network design problems genetic algorithms. last decade, wireless networks experienced impressive growth play main role many telecommunications systems. consequence, scarce radio resources, frequencies, became congested need effective efficient assignment methods arose. work, present genetic algorithm solving large instances power, frequency modulation assignment problem, arising design wireless networks. best knowledge, first genetic algorithm proposed problem. compared previous works, approach allows wider exploration set power solutions, eliminating sources numerical problems. performance algorithm assessed tests set large realistic instances fixed wimax network.",10
imagination holographic processor text animation. imagination critical point developing realistic artificial intelligence (ai) systems. one way approach imagination would simulation properties operations. developed two models: ai-brain network hierarchy languages semantical holographic calculus well simulation system scriptwriter emulate process imagination automatic animation english texts. purpose paper demonstrate model present scriptwriter system http://nvo.sdsc.edu/nvo/jcsg/get_srb_mime_file2.cgi//home/tamara.sdsc/test/demo.zip?f=/home/tamara.sdsc/test/demo.zip&m=application/x-gtar simulation imagination.,3
"statistical analysis loopy belief propagation random fields. loopy belief propagation (lbp), equivalent bethe approximation statistical mechanics, message-passing-type inference method widely used analyze systems based markov random fields (mrfs). paper, propose message-passing-type method analytically evaluate quenched average lbp random fields using replica cluster variation method. proposed analytical method applicable general pair-wise mrfs random fields whose distributions differ give quenched averages bethe free energies random fields, consistent numerical results. order computational cost equivalent standard lbp. latter part paper, describe application proposed method bayesian image restoration, observed theoretical results good agreement numerical results natural images.",18
"learning approach natural language understanding. paper propose learning paradigm problem understanding spoken language. basis work formalization understanding problem communication problem. results definition stochastic model production speech text starting meaning sentence. resulting understanding algorithm consists viterbi maximization procedure, analogous commonly used recognizing speech. algorithm implemented building",1
"spectral descriptors graph matching. paper, consider weighted graph matching problem. recently, approaches problem based spectral methods gained significant attention. propose two graph spectral descriptors based graph laplacian, namely laplacian family signature (lfs) nodes, pairwise heat kernel distance edges. show stability descriptors small perturbation edges nodes. addition, show pairwise heat kernel distance noise-tolerant approximation classical adjacency matrix-based second order compatibility function. nice properties suggest descriptor-based matching scheme, set integer quadratic problem (iqp) apply approximate solver find near optimal solution. tested matching method set randomly generated graphs, widely-used cmu house sequence set real images. experiments show superior performance selected node signatures edge descriptors graph matching, compared existing signature-based matchings adjacency matrix-based matchings.",3
"support vector comparison machines. ranking problems, goal learn ranking function labeled pairs input points. paper, consider related comparison problem, label indicates element pair better, significant difference. cast learning problem margin maximization, show solved converting standard svm. use simulated nonlinear patterns, real learning rank sushi data set, chess data set show proposed svmcompare algorithm outperforms svmrank equality pairs.",18
"modeling epistemological principles bias mitigation ai systems: illustration hiring decisions. artificial intelligence (ai) used extensively automatic decision making broad variety scenarios, ranging credit ratings loans recommendations movies. traditional design guidelines ai models focus essentially accuracy maximization, recent work shown economically irrational socially unacceptable scenarios discrimination unfairness likely arise unless issues explicitly addressed. undesirable behavior several possible sources, biased datasets used training may detected black-box models. pointing connections bias ai problem induction, focus popper's contributions hume's, offer logical theory preferences. ai model preferred others purely rational grounds one attempts refutation based accuracy fairness. inspired epistemological principles, paper proposes structured approach mitigate discrimination unfairness caused bias ai systems. proposed computational framework, models selected enhanced attempts refutation. illustrate discussion, focus hiring decision scenarios ai system filters job applicants go interview phase.",3
"estimating number connected components graph via subgraph sampling. learning properties large graphs samples important problem statistical network analysis since early work goodman \cite{goodman1949} frank \cite{frank1978}. revisit problem formulated frank \cite{frank1978} estimating number connected components large graph based subgraph sampling model, randomly sample subset vertices observe induced subgraph. key question whether accurate estimation achievable \emph{sublinear} regime vanishing fraction vertices sampled. show impossible parent graph allowed contain high-degree vertices long induced cycles. class chordal graphs, induced cycles length four forbidden, characterize optimal sample complexity within constant factors construct linear-time estimators provably achieve bounds. significantly expands scope previous results focused unbiased estimators special classes graphs forests cliques. construction analysis proposed methodology rely combinatorial properties chordal graphs identities induced subgraph counts. they, turn, also play key role proving minimax lower bounds based construction random instances graphs matching structures small subgraphs.",10
"probabilistic evaluation candidates symptom clustering multidisorder diagnosis. paper derives formula computing conditional probability set candidates, candidate set disorders explain given set positive findings. candidate sets produced recent method multidisorder diagnosis called symptom clustering. symptom clustering represents set candidates compactly cartesian product differential diagnoses. evaluating probability candidate set, then, large set candidates validated pruned simultaneously. probability candidate set specialized obtain probability single candidate. unlike earlier results, equation derived allows specification positive, negative, unknown symptoms make assumptions disorders candidate.",3
"convergence em algorithm gaussian mixtures unbalanced mixing coefficients. speed convergence expectation maximization (em) algorithm gaussian mixture model fitting known dependent amount overlap among mixture components. paper, study impact mixing coefficients convergence em. show mixture components exhibit overlap, convergence em becomes slower dynamic range among mixing coefficients increases. propose deterministic anti-annealing algorithm, significantly improves speed convergence em mixtures unbalanced mixing coefficients. proposed algorithm compared standard optimization techniques like bfgs, conjugate gradient, traditional em algorithm. finally, propose similar deterministic anti-annealing based algorithm dirichlet process mixture model demonstrate advantages conventional variational bayesian approach.",3
"deformably registering annotating whole clarity brains atlas via masked lddmm. clarity method renders brains optically transparent enable high-resolution imaging structurally intact brain. anatomically annotating clarity brains necessary discovering regions contain signals interest. manually annotating whole-brain, terabyte clarity images difficult, time-consuming, subjective, error-prone. automatically registering clarity images pre-annotated brain atlas offers solution, difficult several reasons. removal brain skull subsequent storage processing cause variable non-rigid deformations, thus compounding inter-subject anatomical variability. additionally, signal clarity images arises various biochemical contrast agents sparsely label brain structures. sparse labeling challenges commonly used registration algorithms need match image histogram statistics densely labeled histological brain atlases. standard method multiscale mutual information b-spline algorithm dynamically generates average template intermediate registration target. determined method performs poorly registering clarity brains allen institute's mouse reference atlas (ara), image histogram statistics poorly matched. therefore, developed method (mask-lddmm) registering clarity images, automatically find brain boundary learns optimal deformation brain atlas masks. using mask-lddmm without average template provided better results standard approach registering clarity brains ara. lddmm pipelines developed provide fast automated way anatomically annotate clarity images. code available open source software http://neurodata.io.",15
"generalization equivariance convolution neural networks action compact groups. convolutional neural networks extremely successful image recognition domain ensure equivariance translations. many recent attempts generalize framework domains, including graphs data lying manifolds. paper give rigorous, theoretical treatment convolution equivariance neural networks respect translations, action compact group. main result prove (given natural constraints) convolutional structure sufficient, also necessary condition equivariance action compact group. exposition makes use concepts representation theory noncommutative harmonic analysis derives new generalized convolution formulae.",18
"tell look: guided attention inference network. weakly supervised learning coarse labels obtain visual explanations deep neural network attention maps back-propagating gradients. attention maps available priors tasks object localization semantic segmentation. one common framework address three shortcomings previous approaches modeling attention maps: (1) first time make attention maps explicit natural component end-to-end training, (2) provide self-guidance directly maps exploring supervision form network improve them, (3) seamlessly bridge gap using weak extra supervision available. despite simplicity, experiments semantic segmentation task demonstrate effectiveness methods. clearly surpass state-of-the-art pascal voc 2012 val. test set. besides, proposed framework provides way explaining focus learner also feeding back direct guidance towards specific tasks. mild assumptions method also understood plug-in existing weakly supervised learners improve generalization performance.",3
"neural network approach context-sensitive generation conversational responses. present novel response generation system trained end end large quantities unstructured twitter conversations. neural network architecture used address sparsity issues arise integrating contextual information classic statistical models, allowing system take account previous dialog utterances. dynamic-context generative models show consistent gains context-sensitive non-context-sensitive machine translation information retrieval baselines.",3
"generation near-optimal solutions using ilp-guided sampling. interest paper optimisation problems intractable solve direct numerical optimisation, nevertheless significant amounts relevant domain-specific knowledge. category heuristic search techniques known estimation distribution algorithms (edas) seek incrementally sample probability distributions optimal (or near-optimal) solutions increasingly higher probabilities. use domain knowledge assist estimation distributions? answer affirmative, need: (a)a general-purpose technique incorporation domain knowledge constructing models optimal values; (b)a way using models generate new data samples. investigate combination use inductive logic programming (ilp) (a), standard logic-programming machinery generate new samples (b). specifically, iteration distribution estimation, ilp engine used construct model good solutions. resulting theory used guide generation new data instances, restricted derivable using ilp model conjunction background knowledge). demonstrate approach two optimisation problems (predicting optimal depth-of-win krk endgame, job-shop scheduling). results promising: (a)on iteration distribution estimation, samples obtained ilp theory substantially greater proportion good solutions samples without theory; (b)on termination distribution estimation, samples obtained ilp theory contain near-optimal samples samples without theory. taken together, results suggest use ilp-constructed theories could useful technique incorporating complex domain-knowledge estimation distribution procedures.",3
"thermovoltage heat dissipation triangle quantum dot junction. numerically investigate thermoelectric properties triangle quantum dot connected metallic electrodes using non-equilibrium green's function method anderson model. exploiting equation motion method coulomb-blockade regime, thermovoltage, thermocurrent heat dissipation calculated. results show thermovoltage thermocurrent nonlinear behavior, magnitude sign controlled site energy coupling strength quantum dots. moreover, find heat current nonlinear asymmetric respect sign bias voltage site energies quantum dots. analyses show heat current positive negative site energies becomes zero nonzero voltages. results useful determine performance nanoscale electronic devices control heat dissipations.",2
"monte carlo tree search heuristic evaluations using implicit minimax backups. monte carlo tree search (mcts) improved performance game engines domains go, hex, general game playing. mcts shown outperform classic alpha-beta search games good heuristic evaluations difficult obtain. recent years, combining ideas traditional minimax search mcts shown advantageous domains, lines action, amazons, breakthrough. paper, propose new way use heuristic evaluations guide mcts search storing two sources information, estimated win rates heuristic evaluations, separately. rather using heuristic evaluations replace playouts, technique backs implicitly mcts simulations. minimax values used guide future simulations. show using implicit minimax backups leads stronger play performance kalah, breakthrough, lines action.",3
"sparse nonlinear regression: parameter estimation asymptotic inference. study parameter estimation asymptotic inference sparse nonlinear regression. specifically, assume data given $y = f( x^\top \beta^* ) + \epsilon$, $f$ nonlinear. recover $\beta^*$, propose $\ell_1$-regularized least-squares estimator. unlike classical linear regression, corresponding optimization problem nonconvex nonlinearity $f$. spite nonconvexity, prove mild conditions, every stationary point objective enjoys optimal statistical rate convergence. addition, provide efficient algorithm provably converges stationary point. also access uncertainty obtained estimator. specifically, based stationary point objective, construct valid hypothesis tests confidence intervals low dimensional components high-dimensional parameter $\beta^*$. detailed numerical results provided back theory.",18
"using n-grams models visual semantic place recognition. aim paper present new method visual place recognition. system combines global image characterization visual words, allows use efficient bayesian filtering methods integrate several images. precisely, extend classical hmm model techniques inspired field natural language processing. paper presents system bayesian filtering algorithm. performance system influence main parameters evaluated standard database. discussion highlights interest using models proposes improvements.",18
"non-convex one-pass framework generalized factorization machine rank-one matrix sensing. develop efficient alternating framework learning generalized version factorization machine (gfm) steaming data provable guarantees. instances sampled $d$ dimensional random gaussian vectors target second order coefficient matrix gfm rank $k$, algorithm converges linearly, achieves $o(\epsilon)$ recovery error retrieving $o(k^{3}d\log(1/\epsilon))$ training instances, consumes $o(kd)$ memory one-pass dataset requires matrix-vector product operations iteration. key ingredient framework construction estimation sequence endowed so-called conditionally independent rip condition (ci-rip). special cases gfm, framework applied symmetric asymmetric rank-one matrix sensing problems, inductive matrix completion phase retrieval.",18
"fishing clickbaits social images texts linguistically-infused neural network models. paper presents results conclusions participation clickbait challenge 2017 automatic clickbait detection social media. first describe linguistically-infused neural network models identify informative representations predict level clickbaiting present twitter posts. models allow answer question whether post clickbait not, extent clickbait post e.g., all, slightly, considerably, heavily clickbaity using score ranging 0 1. evaluate predictive power models trained varied text image representations extracted tweets. best performing model relies tweet text linguistic markers biased language extracted tweet corresponding page yields mean squared error (mse) 0.04, mean absolute error (mae) 0.16 r2 0.43 held-out test data. binary classification setup (clickbait vs. non-clickbait), model achieved f1 score 0.69. found image representations combined text yield significant performance improvement yet. nevertheless, work first present preliminary analysis objects extracted using google tensorflow object detection api images clickbait vs. non-clickbait twitter posts. finally, outline several steps improve model performance part future work.",3
"latticeopt: heuristic topology optimization framework thin-walled, 2d extruded lattices. paper introduces heuristic topology optimization framework thin-walled, 2d extruded lattice structures subject complex high-speed loading. proposed framework optimizes wall thickness distribution lattice cross section different thickness update schemes, inspired idea equalization absorbed energy density across lattice walls. proposed framework ubiquitous used explicit dynamic simulations, primary numerical method used crashworthiness studies. information material tangent stiffness matrix required, complex material behaviors complex loading conditions handled. three numerical examples presented demonstrate framework capabilities: (1) optimization long, slender column axial compression maximize specific energy absorption, (2) optimization lattice-filled sandwich panel off-center blast loading minimize material damage, (3) generation periodic lattice core design blast loading. results show framework effectively increase specific energy absorption minimize material damage 25 finite element simulations optimization iterations.",10
"optical flow requires multiple strategies (but one network). show matching problem underlies optical flow requires multiple strategies, depending amount image motion factors. study implications observation training deep neural network representing image patches context descriptor based optical flow. propose metric learning method, selects suitable negative samples based nature true match. type training produces network displays multiple strategies depending input leads state art results kitti 2012 kitti 2015 optical flow benchmarks.",3
"model reward-modulated motor learning parallelcortical basal ganglia pathways. many recent studies motor system divided two distinct approaches: investigate motor responses encoded cortical neurons' firing rate dynamics study learning rules mammals songbirds develop reliable motor responses. computationally, first approach encapsulated reservoir computing models, learn intricate motor tasks produce internal dynamics strikingly similar motor cortical neurons, rely biologically unrealistic learning rules. realistic learning rules developed second approach often derived simplified, discrete tasks contrast intricate dynamics characterize real motor responses. bridge two approaches develop biologically realistic learning rule reservoir computing. algorithm learns simulated motor tasks previous reservoir computing algorithms fail, reproduces experimental findings including relate motor learning parkinson's disease treatment.",15
"model spatial thinking computational intelligence. trying effective (no matter exactly field) person face problem inevitably destroys attempts easily get desired goal. problem existence insuperable barriers mind, anotherwords barriers principles thinking. clue main reason research. investigate barriers features exposing nature mental process. start special structures reflect ways define relations objects. came realizing material mind uses build thoughts, make conclusions, understand, form reasoning, etc. called mental dynamics. nature mental barriers required level abstraction well ways pass became clear. begin understand thinking flows way, specifics limitations observe reality. help us optimal. final step start understand, ma-thematical models applied picture. start express thoughts language mathematics, developing apparatus spatial theory mind, suitable represent processes infrastructure thinking. use abstract algebra stay invariant relation nature objects.",3
"gan vae optimal transport point view. short article revisits ideas introduced arxiv:1701.07875 arxiv:1705.07642 simple setup. sheds lights connexions variational autoencoders (vae), generative adversarial networks (gan) minimum kantorovitch estimators (mke).",18
"long time behavior focusing energy-critical nls $\mathbb{r}^d\times\mathbb{t}$ via semivirial-vanishing geometry. study focusing energy-critical nls \begin{align}\label{nls_abstract} i\partial_t u+\delta_{x,y} u=-|u|^{\frac{4}{d-1}} u\tag{nls} \end{align} waveguide manifold $\mathbb{r}_x^d\times\mathbb{t}_y$ $d\geq 2$. reveal somewhat counterintuitive phenomenon despite energy-criticality nonlinear potential, long time dynamics \eqref{nls_abstract} purely determined semivirial-vanishing geometry possesses energy-subcritical characteristic. starting point, consider minimization problem $m_c$ defined semivirial-vanishing manifold prescribed mass $c$. prove sufficiently large mass variational problem $m_c$ unique optimizer $u_c$ satisfying $\partial_y u_c=0$, sufficiently small mass, optimizer $m_c$ must non-trivial $y$-dependence. afterwards, prove $m_c$ characterizes sharp threshold bifurcation finite time blow-up ($d=2,3$) globally scattering ($d=3$) solutions \eqref{nls_abstract} dependence sign semivirial. author's knowledge, paper also gives first large data scattering result focusing energy-critical nls product spaces.",10
amino acid interaction network prediction using multi-objective optimization. protein represented amino acid interaction network. network graph whose vertices proteins amino acids whose edges interactions them. interaction network first step proteins three-dimensional structure prediction. paper present multi-objective evolutionary algorithm interaction prediction ant colony probabilistic optimization algorithm used confirm interaction.,3
"large-scale machine learning metagenomics sequence classification. metagenomics characterizes taxonomic diversity microbial communities sequencing dna directly environmental sample. one main challenges metagenomics data analysis binning step, sequenced read assigned taxonomic clade. due large volume metagenomics datasets, binning methods need fast accurate algorithms operate reasonable computing requirements. standard alignment-based methods provide state-of-the-art performance, compositional approaches assign taxonomic class dna read based k-mers contains potential provide faster solutions. work, investigate potential modern, large-scale machine learning implementations taxonomic affectation next-generation sequencing reads based k-mers profile. show machine learning-based compositional approaches benefit increasing number fragments sampled reference genome tune parameters, coverage 10, increasing k-mer size 12. tuning models involves training machine learning model 10 8 samples 10 7 dimensions, reach standard soft-wares done efficiently modern implementations large-scale machine learning. resulting models competitive terms accuracy well-established alignment tools problems involving small moderate number candidate species, reasonable amounts sequencing errors. show, however, compositional approaches still limited ability deal problems involving greater number species, sensitive sequencing errors. finally confirm compositional approach achieve faster prediction times, gain 3 15 times respect bwa-mem short read mapper, depending number candidate species level sequencing noise.",15
"perspective: energy landscapes machine learning. machine learning techniques increasingly used flexible non-linear fitting prediction tools physical sciences. fitting functions exhibit multiple solutions local minima analysed terms corresponding machine learning landscape. methods explore visualise molecular potential energy landscapes applied machine learning landscapes gain new insight solution space involved training nature corresponding predictions. particular, define quantities analogous molecular structure, thermodynamics, kinetics, relate emergent properties structure underlying landscape. perspective aims describe analogies examples recent applications, suggest avenues new interdisciplinary research.",18
"deep autoregressive approach topic modeling multimodal data. topic modeling based latent dirichlet allocation (lda) framework choice deal multimodal data, image annotation tasks. another popular approach model multimodal data deep neural networks, deep boltzmann machine (dbm). recently, new type topic model called document neural autoregressive distribution estimator (docnade) proposed demonstrated state-of-the-art performance text document modeling. work, show successfully apply extend model multimodal data, simultaneous image classification annotation. first, propose supdocnade, supervised extension docnade, increases discriminative power learned hidden topic features show employ learn joint representation image visual words, annotation words class label information. test model labelme uiuc-sports data sets show compares favorably topic models. second, propose deep extension model provide efficient way training deep model. experimental results show deep model outperforms shallow version reaches state-of-the-art performance multimedia information retrieval (mir) flickr data set.",3
"consistent optimization ams logistic loss minimization. paper, theoretically justify approach popular among participants higgs boson machine learning challenge optimize approximate median significance (ams). approach based following two-stage procedure. first, real-valued function learned minimizing surrogate loss binary classification, logistic loss, training sample. then, threshold tuned separate validation sample, direct optimization ams. show regret resulting (thresholded) classifier measured respect squared ams, upperbounded regret underlying real-valued function measured respect logistic loss. hence, prove minimizing logistic surrogate consistent method optimizing ams.",3
"eliciting worker preference task completion. current crowdsourcing platforms provide little support worker feedback. workers sometimes invited post free text describing experience preferences completing tasks. also use forums turker nation1 exchange preferences tasks requesters. fact, crowdsourcing platforms rely heavily observing workers inferring preferences implicitly. work, believe asking workers indicate preferences explicitly improve experience task completion hence, quality contributions. explicit elicitation indeed help build accurate worker models task completion captures evolving nature worker preferences. design worker model whose accuracy improved iteratively requesting preferences task factors required skills, task payment, task relevance. propose generic framework, develop efficient solutions realistic scenarios, run extensive experiments show benefit explicit preference elicitation implicit ones statistical significance.",3
"kernel approach tractable bayesian nonparametrics. inference popular nonparametric bayesian models typically relies sampling approximations. paper presents general methodology constructing novel tractable nonparametric bayesian methods applying kernel trick inference parametric bayesian model. example, gaussian process regression derived way bayesian linear regression. despite success gaussian process framework, kernel trick rarely explicitly considered bayesian literature. paper, aim fill gap demonstrate potential applying kernel trick tractable bayesian parametric models wider context regression. example, present intuitive bayesian kernel machine density estimation obtained applying kernel trick gaussian generative model feature space.",18
"laplacian support vector machines trained primal. last years, due growing ubiquity unlabeled data, much effort spent machine learning community develop better understanding improve quality classifiers exploiting unlabeled data. following manifold regularization approach, laplacian support vector machines (lapsvms) shown state art performance semi--supervised classification. paper present two strategies solve primal lapsvm problem, order overcome issues original dual formulation. whereas training lapsvm dual requires two steps, using primal form allows us collapse training single step. moreover, computational complexity training algorithm reduced o(n^3) o(n^2) using preconditioned conjugate gradient, n combined number labeled unlabeled examples. speed training using early stopping strategy based prediction unlabeled data or, available, labeled validation examples. allows algorithm quickly compute approximate solutions roughly classification accuracy optimal ones, considerably reducing training time. due simplicity, training lapsvm primal starting point additional enhancements original lapsvm formulation, dealing large datasets. present extensive experimental evaluation real world data showing benefits proposed approach.",18
"towards social pattern characterization egocentric photo-streams. following increasingly popular trend social interaction analysis egocentric vision, manuscript presents comprehensive study automatic social pattern characterization wearable photo-camera user, relying visual analysis egocentric photo-streams. proposed framework consists three major steps. first step detect social interactions user impact several social signals task explored. detected social events inspected second step categorization different social meetings. two steps act event-level potential social event modeled multi-dimensional time-series, whose dimensions correspond set relevant features task, lstm employed classify time-series. last step framework characterize social patterns, essentially infer diversity frequency social relations user discovery recurrences people across whole set social events user. experimental evaluation dataset acquired 9 users demonstrates promising results task social pattern characterization egocentric photo-streams.",3
"vertico ii: effects hi-identified environmental mechanisms molecular gas. vertico early science paper explore detail environmental mechanisms, identified hi, affect resolved properties molecular gas reservoirs cluster galaxies. molecular gas probed using alma aca (+tp) observations 12co(2-1) 51 spiral galaxies virgo cluster (of 49 detected), included viva hi survey. sample spans stellar mass range 9 < log m*/msol < 11. study molecular gas radial profiles, isodensity radii, surface densities function galaxy hi deficiency morphology. weak correlation global hi h2 deficiencies, resolved properties molecular gas correlate hi deficiency: galaxies large hi deficiencies relatively steep truncated molecular gas radial profiles, due removal low-surface density molecular gas outskirts. therefore, environmental mechanisms observed hi also affect molecular gas reservoirs, moderate reduction total amount molecular gas.",0
"topicrnn: recurrent neural network long-range semantic dependency. paper, propose topicrnn, recurrent neural network (rnn)-based language model designed directly capture global semantic meaning relating words document via latent topics. sequential nature, rnns good capturing local structure word sequence - semantic syntactic - might face difficulty remembering long-range dependencies. intuitively, long-range dependencies semantic nature. contrast, latent topic models able capture global underlying semantic structure document account word ordering. proposed topicrnn model integrates merits rnns latent topic models: captures local (syntactic) dependencies using rnn global (semantic) dependencies using latent topics. unlike previous work contextual rnn language modeling, model learned end-to-end. empirical results word prediction show topicrnn outperforms existing contextual rnn baselines. addition, topicrnn used unsupervised feature extractor documents. sentiment analysis imdb movie review dataset report error rate $6.28\%$. comparable state-of-the-art $5.91\%$ resulting semi-supervised approach. finally, topicrnn also yields sensible topics, making useful alternative document models latent dirichlet allocation.",3
"greta - novel global recursive tracking algorithm three dimensions. tracking multiple moving targets allows quantitative measure dynamic behavior systems diverse animal groups biology, turbulence fluid dynamics crowd traffic control. three dimensions, tracking several targets becomes increasingly hard since optical occlusions likely, i.e. two featureless targets frequently overlap several frames. occlusions particularly frequent biological groups bird flocks, fish schools, insect swarms, fact severely limited collective animal behavior field studies past. paper presents 3d tracking method robust case severe occlusions. ensure robustness, adopt global optimization approach works objects frames once. achieve practicality scalability, employ divide conquer formulation, thanks computational complexity problem reduced orders magnitude. tested algorithm synthetic data, experimental data bird flocks insect swarms public benchmark datasets, show system yields high quality trajectories hundreds moving targets severe overlap. results obtained heterogeneous data show potential applicability method diverse experimental situations.",15
"adopting level set theory based algorithms segment human ear. human identification always topic interested researchers around world. biometric methods found effective much easier users traditional identification methods like keys, smart cards passwords. unlike traditional methods, biometric methods data acquisition times passive, means users take active part data acquisition. data acquisition performed using cameras, scanners sensors. human physiological biometrics face, eye ear good candidates uniquely identifying individual. however, human ear scores face eye certain advantages face. challenging phase human identification based ear biometric segmentation ear image captured image may contain many unwanted details. work, pde based image processing techniques used segment ear image. level set theory based image processing employed obtain contour ear image. level set algorithms compared efficiency segmenting test ear images.",3
"solving multistage influence diagrams using branch-and-bound search. branch-and-bound approach solving influ- ence diagrams previously proposed literature, appears never implemented evaluated - apparently due difficulties computing effective bounds branch-and-bound search. paper, describe efficiently compute effective bounds, develop practical implementa- tion depth-first branch-and-bound search influence diagram evaluation outperforms existing methods solving influence diagrams multiple stages.",3
"recurrent convolutional neural networks scene parsing. scene parsing technique consist giving label pixels image according class belong to. ensure good visual coherence high class accuracy, essential scene parser capture image long range dependencies. feed-forward architecture, simply achieved considering sufficiently large input context patch, around pixel labeled. propose approach consisting recurrent convolutional neural network allows us consider large input context, limiting capacity model. contrary standard approaches, method rely segmentation methods, task-specific features. system trained end-to-end manner raw pixels, models complex spatial dependencies low inference cost. context size increases built-in recurrence, system identifies corrects errors. approach yields state-of-the-art performance stanford background dataset sift flow dataset, remaining fast test time.",3
"strides: automated uniform models 30 quadruply imaged quasars. gravitational time delays provide powerful one step measurement $h_0$, independent probes. one key ingredient time delay cosmography high accuracy lens models. currently expensive obtain, both, terms computing investigator time (10$^{5-6}$ cpu hours $\sim$ 0.5-1 year, respectively). major improvements modeling speed therefore necessary exploit large number lenses forecast discovered current decade. order bypass roadblock, building work shajib et al. (2019), develop automated modeling pipeline apply sample 30 quadruply imaged quasars one lensed compact galaxy, observed hubble space telescope multiple bands. automated pipeline derive models 30/31 lenses hours human time <100 cpu hours computing time typical system. lens, provide measurements key parameters predictions magnification well time delays multiple images. characterize cosmography-readiness models using stability differences fermat potential (proportional time delay) w.r.t. modeling choices. find 10/30 lenses models cosmography nearly cosmography grade (<3% 3-5% variations). 6/30 lenses models close cosmography grade (5-10%). results based informative priors need confirmed analysis. however, also likely improve extending pipeline modeling sequence options. conclusion, show uniform cosmography grade modeling large strong lens samples within reach.",0
"land use classification remote sensing images convolutional neural networks. explore use convolutional neural networks semantic classification remote sensing scenes. two recently proposed architectures, caffenet googlenet, adopted, three different learning modalities. besides conventional training scratch, resort pre-trained networks fine-tuned target data, avoid overfitting problems reduce design time. experiments two remote sensing datasets, markedly different characteristics, testify effectiveness wide applicability proposed solution, guarantees significant performance improvement state-of-the-art references.",3
"causal discovery linear acyclic models arbitrary distributions. important task data analysis discovery causal relationships observed variables. continuous-valued data, linear acyclic causal models commonly used model data-generating process, inference models well-studied problem. however, existing methods significant limitations. methods based conditional independencies (spirtes et al. 1993; pearl 2000) cannot distinguish independence-equivalent models, whereas approaches purely based independent component analysis (shimizu et al. 2006) inapplicable data partially gaussian. paper, generalize combine two approaches, yield method able learn model structure many cases previous methods provide answers either incorrect informative possible. give exact graphical conditions two distinct models represent family distributions, empirically demonstrate power method thorough simulations.",18
"synchrotron-based near-field photothermal microspectroscopy: development quantitative nanohistology set-up expansion infrared capability 2. purpose two-fold: explore capability diamond synchrotron infra-red include near-field photothermal microspectroscopy (ptms); toward quantitative nanohistology - investigation scleroderma using synchrotron radiation (mu- ftir). recent advances afm, integration ir temperature-based system ir beamline still promising unique approach. preliminary tests ptms stand-alone system diamond multimode infrared imaging microspectroscopy beamline (miriam) successful one beamtime. knowledge, still among studies worldwide managed interface afm-resistive probe measure infrared spectrum using synchrotron radiation (sr) source. alignment successful, managed obtain good spectra samples strong infrared absorbers. however, alignment could guaranteed time, signal-to-noise ratio became low record meaningful spectra samples deemed scientific interest proposal. clarified possible improvements ptms instrumentation data acquisition needed ensure collect sufficient thermal signals thin samples.",14
"bag-of-words equivalent recurrent neural network action recognition. traditional bag-of-words approach found wide range applications computer vision. standard pipeline consists generation visual vocabulary, quantization features histograms visual words, classification step usually support vector machine combination non-linear kernel used. given large amounts data, however, model suffers lack discriminative power. applies particularly action recognition, vast amount video features needs subsampled unsupervised visual vocabulary generation. moreover, kernel computation expensive large datasets. work, propose recurrent neural network equivalent traditional bag-of-words approach enables application discriminative training. model allows incorporate kernel computation neural network directly, solving complexity issue allowing represent complete classification system within single network. evaluate method four recent action recognition benchmarks show conventional model well sparse coding methods outperformed.",3
"learning graphical models images, videos spatial transformations. mixtures gaussians, factor analyzers (probabilistic pca) hidden markov models staples static dynamic data modeling image video modeling particular. show topographic transformations input, translation shearing images, accounted models including discrete transformation variable. resulting models perform clustering, dimensionality reduction time-series analysis way invariant transformations input. using em algorithm, transformation-invariant models fit static data time series. give results filtering microscopy images, face facial pose clustering, handwritten digit modeling recognition, video clustering, object tracking, removal distractions video sequences.",3
"efficient scene text localization recognition local character refinement. unconstrained end-to-end text localization recognition method presented. method detects initial text hypothesis single pass efficient region-based method subsequently refines text hypothesis using robust local text model, deviates common assumption region-based methods characters detected connected components. additionally, novel feature based character stroke area estimation introduced. feature efficiently computed region distance map, invariant scaling rotations allows efficiently detect text regions regardless portion text capture. method runs real time achieves state-of-the-art text localization recognition results icdar 2013 robust reading dataset.",3
"influence-optimistic local values multiagent planning --- extended version. recent years seen development methods multiagent planning uncertainty scale tens even hundreds agents. however, methods either make restrictive assumptions problem domain, provide approximate solutions without guarantees quality. methods former category typically build heuristic search using upper bounds value function. unfortunately, techniques exist compute upper bounds problems non-factored value functions. allow meaningful benchmarking measurable quality guarantees general class problems, paper introduces family influence-optimistic upper bounds factored decentralized partially observable markov decision processes (dec-pomdps) factored value functions. intuitively, derive bounds large multiagent planning problems subdividing sub-problems, sub-problems making optimistic assumptions respect influence exerted rest system. numerically compare different upper bounds demonstrate achieve non-trivial guarantee heuristic solution problems hundreds agents close optimal. furthermore, provide evidence upper bounds may improve effectiveness heuristic influence search, discuss potential applications multiagent planning.",3
"towards automation data quality system cern cms experiment. daily operation large-scale experiment challenging task, particularly perspectives routine monitoring quality data taken. describe approach uses machine learning automated system monitor data quality, based partial use data qualified manually detector experts. system automatically classifies marginal cases: good bad data, use human expert decision classify remaining ""grey area"" cases. study uses collision data collected cms experiment lhc 2010. demonstrate proposed workflow able automatically process least 20\% samples without noticeable degradation result.",14
"structured variational auto-encoder learning deep hierarchies sparse features. note present generative model natural images consisting deep hierarchy layers latent random variables, follows new type distribution call rectified gaussian. rectified gaussian units allow spike-and-slab type sparsity, retaining differentiability necessary efficient stochastic gradient variational inference. learn parameters new model, approximate posterior latent variables variational auto-encoder. rather making usual mean-field assumption however, encoder parameterizes new type structured variational approximation retains prior dependencies generative model. using structured posterior approximation, able perform joint training deep models many layers latent random variables, without resort stacking layerwise training procedures.",18
"asymptotic risks viterbi segmentation. consider maximum likelihood (viterbi) alignment hidden markov model (hmm). hmm, underlying markov chain usually hidden viterbi alignment often used estimate it. approach referred viterbi segmentation. goodness viterbi segmentation measured several risks. paper, prove existence asymptotic risks. independent data, asymptotic risks considered characteristics model illustrate long-run behavior viterbi segmentation.",10
"novel approach cardiac disease prediction classification using intelligent agents. goal develop novel approach cardiac disease prediction diagnosis using intelligent agents. initially symptoms preprocessed using filter wrapper based agents. filter removes missing irrelevant symptoms. wrapper used extract data data set according threshold limits. dependency symptom identified using dependency checker agent. classification based prior posterior probability symptoms evidence value. finally symptoms classified five classes namely absence, starting, mild, moderate serious. using cooperative approach cardiac problem solved verified.",3
"reflexive regular equivalence bipartite data. bipartite data common data engineering brings unique challenges, particularly comes clustering tasks impose strong structural assumptions. work presents unsupervised method assessing similarity bipartite data. similar co-clustering methods, method based regular equivalence graphs. algorithm uses spectral properties bipartite adjacency matrix estimate similarity dimensions. method reflexive similarity one dimension used inform similarity other. reflexive regular equivalence also use structure transitivities -- network sense -- contribution controlled algorithm's free-parameter, $\alpha$. method completely unsupervised used validate assumptions co-similarity, required often untested, co-clustering analyses. three variants method different normalizations tested synthetic data. method found robust noise well-suited asymmetric co-similar structure, making particularly informative cluster analysis recommendation bipartite data unknown structure. experiments, convergence speed algorithm found stable different levels noise. real-world data network malaria genes analyzed, similarity produced reflexive method shown out-perform measures' ability correctly classify genes.",3
"gaussian processes limiting linear models. gaussian processes retain linear model either special case, limit. show relationship exploited data least partially linear. however perspective bayesian posterior, gaussian processes encode linear model either probability nearly zero otherwise unattainable without explicit construction prior limiting linear model mind. develop prior, show practical benefits extend well beyond computational conceptual simplicity linear model. example, linearity extracted per-dimension basis, combined treed partition models yield highly efficient nonstationary model. approach demonstrated synthetic real datasets varying linearity dimensionality.",18
"reconstruction photospheric velocity fields highly corrupted data. analysis photospheric velocity field essential understanding plasma turbulence solar surface, may responsible driving processes magnetic reconnection, flares, wave propagation, particle acceleration, coronal heating. currently, available methods estimate velocities solar photosphere transverse observer's line sight infer flows differences image structure successive observations. due data noise, algorithms local correlation tracking (lct) may lead vector field wide gaps velocity vectors provided. letter, novel method image inpainting highly corrupted data proposed applied restoration horizontal velocity fields solar photosphere. restored velocity field preserves vector field components present original field. method shows robustness applied simulated observational data.",0
improvement negative sentences translation english-to-korean machine translation. paper describes algorithm translating english negative sentences korean english-korean machine translation (ekmt). proposed algorithm based comparative study english korean negative sentences. earlier translation software cannot translate english negative sentences accurate korean equivalents. established new algorithm negative sentence translation evaluated it.,3
"boosting image forgery detection using resampling features copy-move analysis. realistic image forgeries involve combination splicing, resampling, cloning, region removal methods. resampling detection algorithms effective detecting splicing resampling, copy-move detection algorithms excel detecting cloning region removal. paper, combine complementary approaches way boosts overall accuracy image manipulation detection. use copy-move detection method pre-filtering step pass images classified untampered deep learning based resampling detection framework. experimental results various datasets including 2017 nist nimble challenge evaluation dataset comprising nearly 10,000 pristine tampered images shows consistent increase 8%-10% detection rates, copy-move algorithm combined different resampling detection algorithms.",3
"proximal scope distributed sparse learning: better data partition implies faster convergence rate. distributed sparse learning cluster multiple machines attracted much attention machine learning, especially large-scale applications high-dimensional data. one popular way implement sparse learning use $l_1$ regularization. paper, propose novel method, called proximal \mbox{scope}~(\mbox{pscope}), distributed sparse learning $l_1$ regularization. pscope based \underline{c}ooperative \underline{a}utonomous \underline{l}ocal \underline{l}earning~(\mbox{call}) framework. \mbox{call} framework \mbox{pscope}, find data partition affects convergence learning procedure, subsequently define metric measure goodness data partition. based defined metric, theoretically prove pscope convergent linear convergence rate data partition good enough. also prove better data partition implies faster convergence rate. furthermore, pscope also communication efficient. experimental results real data sets show pscope outperform state-of-the-art distributed methods sparse learning.",18
"pde-foam - probability-density estimation method using self-adapting phase-space binning. probability density estimation (pde) multivariate discrimination technique based sampling signal background densities defined event samples data monte-carlo (mc) simulations multi-dimensional phase space. paper, present modification pde method uses self-adapting binning method divide multi-dimensional phase space finite number hyper-rectangles (cells). binning algorithm adjusts size position predefined number cells inside multi-dimensional phase space, minimising variance signal background densities inside cells. implementation binning algorithm pde-foam based mc event-generation package foam. present performance results representative examples (toy models) discuss dependence obtained results choice parameters. new pde-foam shows improved classification capability small training samples reduced classification time compared original pde method based range searching.",14
"modèle à processus latent et algorithme em pour la régression non linéaire. non linear regression approach consists specific regression model incorporating latent process, allowing various polynomial regression models activated preferentially smoothly, introduced paper. model parameters estimated maximum likelihood performed via dedicated expecation-maximization (em) algorithm. experimental study using simulated real data sets reveals good performances proposed approach.",10
"knowledge representation analysis graph mining. many problems, especially composite structure, naturally expressed higher order logic. kr perspective modeling problems intuitive way challenging task. paper study graph mining problem example higher order problem. short, problem asks us find graph frequently occurs subgraph among set example graphs. start problem's mathematical definition solve three state-of-the-art specification systems. idp asp, native support higher order logic, propose use encoding techniques disjoint union technique saturation technique. prob benefits higher order support sets. compare performance three approaches get idea overhead higher order support. propose higher-order language extensions idp-like specification languages discuss kind solver support needed. native higher order shifts burden rewriting specifications using encoding techniques user solver itself.",3
"collaborative multi-sensor classification via sparsity-based representation. paper, propose general collaborative sparse representation framework multi-sensor classification, takes account correlations well complementary information heterogeneous sensors simultaneously considering joint sparsity within sensor's observations. also robustify models deal presence sparse noise low-rank interference signals. specifically, demonstrate incorporating noise interference signal low-rank component models essential multi-sensor classification problem multiple co-located sources/sensors simultaneously record physical event. extend frameworks kernelized models rely sparsely representing test sample terms training samples feature space induced kernel function. fast efficient algorithm based alternative direction method proposed convergence optimal solution guaranteed. extensive experiments conducted several real multi-sensor data sets results compared conventional classifiers verify effectiveness proposed methods.",3
"model free artificial entities. impression free feeling according choices neither imposed inside outside. sense ultimate cause acts. direct opposition universal determinism, existence free continues discussed. paper, free linked decisional mechanism: agent provided free performed predictable choice cp, immediately perform another choice cr random way. intangible feeling free replaced decision-making process including predictable decision-making process immediately followed unpredictable decisional one.",3
"linear convergence sdca statistical estimation. paper, consider stochastic dual coordinate (sdca) {\em without} strongly convex assumption convex assumption. show sdca converges linearly mild conditions termed restricted strong convexity. covers wide array popular statistical models including lasso, group lasso, logistic regression $\ell_1$ regularization, corrected lasso linear regression scad regularizer. significantly improves previous convergence results sdca problems strongly convex. product, derive dual free form sdca handle general regularization term, interest itself.",18
"unsupervised algorithm learning lie group transformations. present several theoretical contributions allow lie groups fit high dimensional datasets. transformation operators represented eigen-basis, reducing computational complexity parameter estimation training linear transformation model. transformation specific ""blurring"" operator introduced allows inference escape local minima via smoothing transformation space. penalty traversed manifold distance added encourages discovery sparse, minimal distance, transformations states. learning inference demonstrated using methods full set affine transformations natural image patches. transformation operators trained natural video sequences. shown learned video transformations provide better description inter-frame differences standard motion model based rigid translation.",3
"differential evolution quantum robust control: algorithm, applications experiments. robust control design quantum systems recognized key task quantum information technology, molecular chemistry atomic physics. paper, improved differential evolution algorithm msms_de proposed search robust fields various quantum control problems. msms_de, multiple samples used fitness evaluation mixed strategy employed mutation operation. particular, msms_de algorithm applied control problem open inhomogeneous quantum ensembles consensus problem quantum network uncertainties. numerical results presented demonstrate excellent performance improved de algorithm two classes quantum robust control problems. furthermore, msms_de experimentally implemented femtosecond laser control systems generate good signals two photon absorption control fragmentation halomethane molecules ch2bri. experimental results demonstrate excellent performance msms_de searching effective femtosecond laser pulses various tasks.",17
"learning optimization viewpoint. dissertation study statistical online learning problems optimization viewpoint.the dissertation divided two parts : i. first consider question learnability statistical learning problems general learning setting. question learnability well studied fully characterized binary classification real valued supervised learning problems using theory uniform convergence. however show general learning setting uniform convergence theory fails characterize learnability. fill void use stability learning algorithms fully characterize statistical learnability general setting. next consider problem online learning. unlike statistical learning framework dearth generic tools used establish learnability rates online learning problems general. provide online analogs classical tools statistical learning theory like rademacher complexity, covering numbers, etc. use tools fully characterize learnability online supervised learning problems. ii. second part, general classes convex learning problems, provide appropriate mirror descent (md) updates online statistical learning problems. further, show md near optimal online convex learning cases, also near optimal statistical convex learning. next consider problem convex optimization show oracle complexity lower bounded called fat-shattering dimension associated linear class. thus establish strong connection offline convex optimization problems statistical learning problems. also show large class high dimensional optimization problems, md fact near optimal even convex optimization.",3
"joint reconstruction multi-view compressed images. distributed representation correlated multi-view images important problem arise vision sensor networks. paper concentrates joint reconstruction problem distributively compressed correlated images jointly decoded order improve reconstruction quality compressed images. consider scenario images captured different viewpoints encoded independently using common coding solutions (e.g., jpeg, h.264 intra) balanced rate distribution among different cameras. central decoder first estimates underlying correlation model independently compressed images used joint signal recovery. joint reconstruction cast constrained convex optimization problem reconstructs total-variation (tv) smooth images comply estimated correlation model. time, add constraints force reconstructed images consistent compressed versions. show experiments proposed joint reconstruction scheme outperforms independent reconstruction terms image quality, given target bit rate. addition, decoding performance proposed algorithm compares advantageously state-of-the-art distributed coding schemes based disparity learning discover.",3
"physarum-inspired approach optimal supply chain network design minimum total cost demand satisfaction. supply chain system moves products supplier customers. supply chains ubiquitous. play key role economic activities. inspired biological principles nutrients' distribution protoplasmic networks slime mould physarum polycephalum propose novel algorithm supply chain design. algorithm handles supply networks capacity investments product flows variables. networks constrained need satisfy product demands. two features slime mould adopted algorithm. first continuity flux iterative process, used real-time update costs associated supply links. second feature adaptivity. supply chain converge equilibrium state costs changed. practicality flexibility algorithm illustrated numerical examples.",3
"driven translocation semiflexible polymer conical channel presence attractive surface interactions. study translocation semiflexible polymer conical channel attractive surface interactions driving force varies spatially inside channel. using results translocation dynamics flexible polymer extended channel control, first show asymmetric shape channel gives rise non-monotonic features total translocation time function apex angle channel. waiting time distributions individual monomer beads inside channel show unique features strongly dependent driving force surface interactions. polymer stiffness results longer translocation times angles channel. further, non-monotonic features translocation time function channel angle changes substantially polymer becomes stiffer, reflected changing features waiting time distributions. construct free energy description system incorporating entropic energetic contributions low force regime explain simulation results.",2
"recognizing descriptive wikipedia categories historical figures. wikipedia useful knowledge source benefits many applications language processing knowledge representation. important feature wikipedia categories. wikipedia pages assigned different categories according contents human-annotated labels used information retrieval, ad hoc search improvements, entity ranking tag recommendations. however, important pages usually assigned many categories, makes difficult recognize important ones give best descriptions. paper, propose approach recognize descriptive wikipedia categories. observe historical figures precise category presumably mutually similar categorical coherence could evaluated via texts wikipedia links corresponding members category. rank descriptive level wikipedia categories according coherence ranking yield overall agreement 88.27% compared human wisdom.",3
"photometric catalogue quasars point sources sloan digital sky survey. present catalogue 6 million unresolved photometric detections sloan digital sky survey seventh data release classifying stars, galaxies quasars. use machine learning classifier trained subset spectroscopically confirmed objects 14th 22nd magnitude sdss {\it i}-band. catalogue consists 2,430,625 quasars, 3,544,036 stars 63,586 unresolved galaxies 14th 24th magnitude sdss {\it i}-band. algorithm recovers 99.96% spectroscopically confirmed quasars 99.51% stars $\sim$21.3 colour window study. level contamination due data artefacts objects beyond $i=21.3$ highly uncertain mention completeness contamination paper valid objects brighter magnitude. however, comparison predicted number quasars theoretical number counts shows reasonable agreement.",0
"learning diagnose scratch exploiting dependencies among labels. field medical diagnostics contains wealth challenges closely resemble classical machine learning problems; practical constraints, however, complicate translation endpoints naively classical architectures. many tasks radiology, example, largely problems multi-label classification wherein medical images interpreted indicate multiple present suspected pathologies. clinical settings drive necessity high accuracy simultaneously across multitude pathological outcomes greatly limit utility tools consider subset. issue exacerbated general scarcity training data maximizes need extract clinically relevant features available samples -- ideally without use pre-trained models may carry forward undesirable biases tangentially related tasks. present evaluate partial solution constraints using lstms leverage interdependencies among target labels predicting 14 pathologic patterns chest x-rays establish state art results largest publicly available chest x-ray dataset nih without pre-training. furthermore, propose discuss alternative evaluation metrics relevance clinical practice.",3
"stochastic deep learning memristive networks. study performance stochastically trained deep neural networks (dnns) whose synaptic weights implemented using emerging memristive devices exhibit limited dynamic range, resolution, variability programming characteristics. show key device parameter optimize learning efficiency dnns variability programming characteristics. dnns memristive synapses, even dynamic range low $15$ $32$ discrete levels, trained based stochastic updates suffer less $3\%$ loss accuracy compared floating point software baseline. also study performance stochastic memristive dnns used inference engines noise corrupted data find device variability minimized, relative degradation performance stochastic dnn better software baseline. hence, study presents new optimization corner memristive devices building large noise-immune deep learning systems.",18
"audio-replay attack detection countermeasures. paper presents speech technology center (stc) replay attack detection systems proposed automatic speaker verification spoofing countermeasures challenge 2017. study focused comparison different spoofing detection approaches. gmm based methods, high level features extraction simple classifier deep learning frameworks. experiments performed development evaluation parts challenge dataset demonstrated stable efficiency deep learning approaches case changing acoustic conditions. time svm classifier high level features provided substantial input efficiency resulting stc systems according fusion systems results.",3
"cheshire: online algorithm activity maximization social networks. user engagement social networks depends critically number online actions users take network. design algorithm finds incentivize users take actions maximize overall activity social network? paper, model number online actions time using multidimensional hawkes processes, derive alternate representation processes based stochastic differential equations (sdes) jumps and, exploiting alternate representation, address question perspective stochastic optimal control sdes jumps. find optimal level incentivized actions depends linearly current level overall actions. moreover, coefficients linear relationship found solving matrix riccati differential equation, solved efficiently, first order differential equation, closed form solution. result, able design efficient online algorithm, cheshire, sample optimal times users' incentivized actions. experiments synthetic real data gathered twitter show algorithm able consistently maximize number online actions effectively state art.",18
"biometric authentication using nonparametric methods. physiological behavioral trait employed develop biometric authentication systems. proposed work deals authentication iris signature based minimum variance criteria. iris patterns preprocessed based area connected components. segmented image used authentication consists region large variations gray level values. image region split quadtree components. components minimum variance determined training samples. hu moments applied components. summation moment values corresponding minimum variance components provided input vector k-means fuzzy kmeans classifiers. best performance obtained mmu database consisting 45 subjects. number subjects zero false rejection rate [frr] 44 number subjects zero false acceptance rate [far] 45. paper addresses computational load reduction off-line signature verification based minimal features using k-means, fuzzy k-means, k-nn, fuzzy k-nn novel average-max approaches. frr 8.13% far 10% achieved using k-nn classifier. signature biometric, variations genuine case, natural expectation. genuine signature, certain parts signature vary one instance another. system aims provide simple, fast robust system using less number features compared state art works.",3
"quantum autoencoders via quantum adders genetic algorithms. quantum autoencoder recent paradigm field quantum machine learning, may enable enhanced use resources quantum technologies. end, quantum neural networks less nodes inner outer layers considered. here, propose useful connection approximate quantum adders quantum autoencoders. specifically, link allows us employ optimized approximate quantum adders, obtained genetic algorithms, implementation quantum autoencoders variety initial states. furthermore, also directly optimize quantum autoencoders via genetic algorithms. approach opens different path design quantum autoencoders controllable quantum platforms.",17
"snaptogrid: statistical interpretable models biomedical information extraction. propose approach biomedical information extraction marries advantages machine learning models, e.g., learning directly data, benefits rule-based approaches, e.g., interpretability. approach starts training feature-based statistical model, converts model rule-based variant converting features rules, ""snapping grid"" feature weights discrete votes. so, proposal takes advantage large body work machine learning, produces interpretable model, directly edited experts. evaluate approach bionlp 2009 event extraction task. results show small performance penalty converting statistical model rules, gain interpretability compensates that: minimal effort, human experts improve model similar performance statistical model served starting point.",3
"deep galaxy: classification galaxies based deep convolutional neural networks. paper, deep convolutional neural network architecture galaxies classification presented. galaxy classified based features main three categories elliptical, spiral, irregular. proposed deep galaxies architecture consists 8 layers, one main convolutional layer features extraction 96 filters, followed two principles fully connected layers classification. trained 1356 images achieved 97.272% testing accuracy. comparative result made testing accuracy compared related works. proposed architecture outperformed related works terms testing accuracy.",3
"inertia-constrained pixel-by-pixel nonnegative matrix factorisation: hyperspectral unmixing method dealing intra-class variability. blind source separation common processing tool analyse constitution pixels hyperspectral images. methods usually suppose pure pixel spectra (endmembers) image class materials. framework remote sensing, assumption valid presence intra-class variabilities due illumination conditions, weathering, slight variations pure materials, etc... paper, first describe results investigations highlighting intra-class variability measured real images. considering results, new formulation linear mixing model presented leading two new methods. unconstrained pixel-by-pixel nmf (up-nmf) new blind source separation method based assumption linear mixing model, deal intra-class variability. overcome up-nmf limitations extended method proposed, named inertia-constrained pixel-by-pixel nmf (ip-nmf). sensed spectrum, extended versions nmf extract corresponding set source spectra. constraint set limit spreading source's estimates ip-nmf. methods tested semi-synthetic data set built spectra extracted real hyperspectral image numerically mixed. thus demonstrate interest methods realistic source variabilities. finally, ip-nmf tested real data set shown yield better performance state art methods.",18
"fixed-point iterations several dissimilarity measure barycenters gaussian case. target tracking sensor fusion contexts unusual deal large number gaussian densities encode available information (multiple hypotheses), applications many sensors, affected clutter multimodal noise, take measurements scene. cases reduction procedures must implemented, purpose limiting computational load. situations required fuse available information single hypothesis, usually done computing barycenter set. however, computation strongly depends chosen dissimilarity measure, often must performed making use numerical methods, since cases barycenter computed analytically. issues, like constraint covariance, must symmetric positive definite, make hard numerical computation barycenter set gaussians. work, fixed-point iterations (fpi) presented computation barycenters according several dissimilarity measures, making useful toolbox fusion/reduction gaussian sets applications specific dissimilarity measures required.",18
"bionic humans using eap artificial muscles reality challenges. many years, idea human bionic muscles immediately conjures science fiction images tv series superhuman character implanted bionic muscles portrayed strength speed far superior normal human. fantastic idea may seem, recent developments electroactive polymers (eap) may one day make bionics possible. polymers exhibit large displacement response stimulation electrical signal known many years. initially, eap received relatively little attention due limited actuation capability. however, recent years, view eap materials changed due introduction effective new materials significantly surpassed capability widely used piezoelectric polymer, pvdf. technology continues evolve, novel mechanisms biologically inspired expected emerge. eap materials potentially provide actuation lifelike response flexible configurations. improvements performance robustness still needed, already several reported successes. recognition need cooperation multidisciplinary field, author initiated organized series international forums leading growing number research development projects great advances field. 1999, challenged worldwide science engineering community eap experts develop robotic arm actuated artificial muscles win wrestling match human opponent. paper, field eap artificial muscles reviewed covering state art, challenges vision progress future years.",3
"differentiable scheduled sampling credit assignment. demonstrate continuous relaxation argmax operation used create differentiable approximation greedy decoding sequence-to-sequence (seq2seq) models. incorporating approximation scheduled sampling training procedure (bengio et al., 2015)--a well-known technique correcting exposure bias--we introduce new training objective continuous differentiable everywhere provide informative gradients near points previous decoding decisions change value. addition, using related approximation, demonstrate similar approach sampled-based training. finally, show approach outperforms cross-entropy training scheduled sampling procedures two sequence prediction tasks: named entity recognition machine translation.",3
"aggregated sparse attention steering angle prediction. paper, apply attention mechanism autonomous driving steering angle prediction. propose first model, applying recently introduced sparse attention mechanism visual domain, well aggregated extension model. show improvement proposed method, comparing attention well different types attention.",3
"gaussian probabilities expectation propagation. gaussian probability densities omnipresent applied mathematics, gaussian cumulative probabilities hard calculate univariate case. study utility expectation propagation (ep) approximate integration method problem. rectangular integration regions, approximation highly accurate. also extend derivations general case polyhedral integration regions. however, find polyhedral case, ep's answer, though often accurate, almost arbitrarily wrong. consider unexpected results empirically theoretically, problem gaussian probabilities ep generally. results elucidate interesting non-obvious feature ep yet studied detail.",18
"solving set constraint satisfaction problems using robdds. paper present new approach modeling finite set domain constraint problems using reduced ordered binary decision diagrams (robdds). show possible construct efficient set domain propagator compactly represents many set domains set constraints using robdds. demonstrate robdd-based approach provides unprecedented flexibility modeling constraint satisfaction problems, leading performance improvements. also show robdd-based modeling approach extended modeling integer multiset constraint problems straightforward manner. since domain propagation always practical, also show incorporate less strict consistency notions robdd framework, set bounds, cardinality bounds lexicographic bounds consistency. finally, present experimental results demonstrate robdd-based solver performs better various conventional constraint solvers several standard set constraint problems.",3
"semi-supervised data representation via affinity graph learning. consider general problem utilizing labeled unlabeled data improve data representation performance. new semi-supervised learning framework proposed combing manifold regularization data representation methods non negative matrix factorization sparse coding. adopt unsupervised data representation methods learning machines depend labeled data, improve machine's generation ability much possible. proposed framework forms laplacian regularizer learning affinity graph. incorporate new laplacian regularizer unsupervised data representation smooth low dimensional representation data make use label information. experimental results several real benchmark datasets indicate semi-supervised learning framework achieves encouraging results compared state-of-art methods.",3
"performance network parallel training artificial neural networks. artificial neural networks (anns) received increasing attention recent years applications span wide range disciplines including vital domains medicine, network security autonomous transportation. however, neural network architectures becoming increasingly complex increasing need obtain real-time results models, become pivotal use parallelization mechanism speeding network training deployment. work propose implementation network parallel training cannon's algorithm matrix multiplication. show increasing number processes speeds training point process communication costs become prohibitive; point varies network complexity. also show empirical efficiency calculations speedup obtained superlinear.",3
"globally optimal object tracking fully convolutional networks. tracking one important still difficult tasks computer vision pattern recognition. main difficulties tracking field appearance variation occlusion. traditional tracking methods set parameters templates track target objects advance modified accordingly. thus, propose new robust tracking method using fully convolutional network (fcn) obtain object probability map dynamic programming (dp) seek globally optimal path frames video. proposed method solves object appearance variation problem use fcn deals occlusion dp. show method effective tracking various single objects video frames.",3
"fixing error caponnetto de vito (2007). seminal paper caponnetto de vito (2007) provides minimax-optimal rates kernel ridge regression general setting. proof, however, contains error bound effective dimensionality. note, explain mistake, provide correct bound, show main theorem remains true.",18
"effective extensible feature extraction method using genetic algorithm-based frequency-domain feature search epileptic eeg multi-classification. paper, genetic algorithm-based frequency-domain feature search (gafds) method proposed electroencephalogram (eeg) analysis epilepsy. method, frequency-domain features first searched combined nonlinear features. subsequently, features selected optimized classify eeg signals. extracted features analyzed experimentally. features extracted gafds show remarkable independence, superior nonlinear features terms ratio inter-class distance intra-class distance. moreover, proposed feature search method additionally search features instantaneous frequency signal hilbert transformation. classification results achieved using features reasonable, thus, gafds exhibits good extensibility. multiple classic classifiers (i.e., $k$-nearest neighbor, linear discriminant analysis, decision tree, adaboost, multilayer perceptron, na\""ive bayes) achieve good results using features generated gafds method optimized selection. specifically, accuracies two-classification three-classification problems may reach 99% 97%, respectively. results several cross-validation experiments illustrate gafds effective feature extraction eeg classification. therefore, proposed feature selection optimization model improve classification accuracy.",3
"framework sequential planning multi-agent settings. paper extends framework partially observable markov decision processes (pomdps) multi-agent settings incorporating notion agent models state space. agents maintain beliefs physical states environment models agents, use bayesian updates maintain beliefs time. solutions map belief states actions. models agents may include belief states related agent types considered games incomplete information. express agents autonomy postulating models directly manipulable observable agents. show important properties pomdps, convergence value iteration, rate convergence, piece-wise linearity convexity value functions carry framework. approach complements traditional approach interactive settings uses nash equilibria solution paradigm. seek avoid drawbacks equilibria may non-unique capture off-equilibrium behaviors. cost represent, process continuously revise models agents. since agents beliefs may arbitrarily nested, optimal solutions decision making problems asymptotically computable. however, approximate belief updates approximately optimal plans computable. illustrate framework using simple application domain, show examples belief updates value functions.",3
"proceedings 5th workshop machine learning interpretation neuroimaging (mlini) nips 2015. volume collection contributions 5th workshop machine learning interpretation neuroimaging (mlini) neural information processing systems (nips 2015) conference. modern multivariate statistical methods developed rapidly growing field machine learning increasingly applied various problems neuroimaging, cognitive state detection clinical diagnosis prognosis. multivariate pattern analysis methods designed examine complex relationships high-dimensional signals, brain images, outcomes interest, category stimulus, type mental state subject, specific mental disorder. techniques contrast traditional mass-univariate approaches dominated neuroimaging past treated individual imaging measurement isolation. believe machine learning prominent role shaping questions neuroscience framed, machine-learning mind set entering modern psychology behavioral studies. also equally important practical applications fields motivate rapidly evolving line research machine learning community. parallel, intense interest learning brain function context rich naturalistic environments scenes. efforts go beyond highly specific paradigms pinpoint single function, towards schemes measuring interaction natural varied scene made. goal workshop pinpoint pressing issues common challenges across neuroscience, neuroimaging, psychology machine learning fields, sketch future directions open questions light novel methodology.",18
"dense distributions sparse samples: improved gibbs sampling parameter estimators lda. introduce novel approach estimating latent dirichlet allocation (lda) parameters collapsed gibbs samples (cgs), leveraging full conditional distributions latent variable assignments efficiently average multiple samples, little computational cost drawing single additional collapsed gibbs sample. approach understood adapting soft clustering methodology collapsed variational bayes (cvb0) cgs parameter estimation, order get best techniques. estimators straightforwardly applied output existing implementation cgs, including modern accelerated variants. perform extensive empirical comparisons estimators standard collapsed inference algorithms real-world data unsupervised lda prior-lda, supervised variant lda multi-label classification. results show consistent advantage approach traditional cgs experimental conditions, cvb0 inference majority conditions. broadly, results highlight importance averaging multiple samples lda parameter estimation, use efficient computational techniques so.",18
"new unbiased efficient class lsh-based samplers estimators partition function computation log-linear models. log-linear models arguably successful class graphical models large-scale applications simplicity tractability. learning inference models require calculating partition function, major bottleneck intractable large state spaces. importance sampling (is) mcmc-based approaches lucrative. however, condition ""good"" proposal distribution often satisfied practice. paper, add new dimension efficient estimation via sampling. propose new sampling scheme unbiased estimator estimates partition function accurately sub-linear time. samples generated near-constant time using locality sensitive hashing (lsh), correlated unnormalized. demonstrate effectiveness proposed approach comparing accuracy speed estimating partition function state-of-the-art estimation techniques including efficient variant gumbel-max sampling. efficient sampling scheme, accurately train real-world language models using 1-2% computations.",18
"learning deep features one-class classification. propose deep learning-based solution problem feature learning one-class classification. proposed method operates top convolutional neural network (cnn) choice produces descriptive features maintaining low intra-class variance feature space given class. purpose two loss functions, compactness loss descriptiveness loss proposed along parallel cnn architecture. template matching-based framework introduced facilitate testing process. extensive experiments publicly available anomaly detection, novelty detection mobile active authentication datasets show proposed deep one-class (doc) classification method achieves significant improvements state-of-the-art.",3
"odd order $c_4$-face-magic $m \times n$ projective grid graphs $c_4$-face-magic value $2mn+1$ $2mn+3$. graph $g = (v, e)$ embedded projective plane, let $\mathcal{f}(g)$ denote set faces $g$. then, $g$ called $c_n$-face-magic projective graph exists bijection $f: v(g) \to \{1, 2, \dots, |v(g)|\}$ $f \in \mathcal{f}(g)$ $f \cong c_n$, sum vertex labels around $c_n$ constant $s$. consider $m \times n$ grid graph, denoted $\mathcal{p}_{m,n}$, embedded projective plane natural way. let $m \geqslant 3$ $n \geqslant 3$ odd integers. known $c_4$-face-magic value $c_4$-face-magic labeling $\mathcal{p}_{m,n}$ either $2mn+1$, $2mn+2$, $2mn+3$. characterization $c_4$-face-magic labelings $\mathcal{p}_{m,n}$ $c_4$-face-magic value $2mn+2$ known. paper, determine category $c_4$-face-magic labelings $\mathcal{p}_{m,n}$ $c_4$-face-magic value either $2mn+1$ $2mn+3$. conjectured $c_4$-face-magic labeling $\mathcal{p}_{m,n}$ $c_4$-face-magic value $2mn+1$ $2mn+3$.",10
"causes query answers databases: datalog abduction, view-updates, integrity constraints. causality recently introduced databases, model, characterize, possibly compute causes query answers. connections qa-causality consistency-based diagnosis database repairs (wrt. integrity constraint violations) already established. work establish precise connections qa-causality abductive diagnosis view-update problem databases, allowing us obtain new algorithmic complexity results qa-causality. also obtain new results complexity view-conditioned causality, investigate notion qa-causality presence integrity constraints, obtaining complexity results connection view-conditioned causality. abduction connection integrity constraints allows us obtain algorithmic tools qa-causality.",3
"machine learning neural decoding. machine learning tools rapidly advancing, majority neural decoding approaches still use last century's methods. improving performance neural decoding algorithms allows us better understand information contained brain, help advance engineering applications brain machine interfaces. here, apply modern machine learning techniques, including neural networks gradient boosting, decode spiking activity 1) motor cortex, 2) somatosensory cortex, 3) hippocampus. compare predictive ability modern methods traditional decoding methods wiener kalman filters. modern methods, particular neural networks ensembles, significantly outperformed traditional approaches. instance, three brain areas, lstm decoder explained 40% unexplained variance wiener filter. results suggest modern machine learning techniques become standard methodology neural decoding. provide code facilitate wider implementation methods.",15
"searching image information content, discovery, extraction, representation. image information content known complicated controvercial problem. paper posits new image information content definition. following theory solomonoff-kolmogorov-chaitin's complexity, define image information content set descriptions imafe data structures. three levels description generally distinguished: 1)the global level, coarse structure entire scene initially outlined; 2) intermediate level, structures separate, non-overlapping image regions usually associated individual scene objects deliniated; 3) low-level description, local image structures observed limited restricted field view resolved. technique creating image information content descriptors developed. algorithm presented elucidated examples, demonstrate effectiveness proposed approach.",3
"gradient regularization improves accuracy discriminative models. regularizing gradient norm output neural network respect inputs powerful technique, first proposed drucker & lecun (1991) named double backpropagation. idea independently rediscovered several times since then, often goal making models robust adversarial sampling. paper presents evidence gradient regularization consistently significantly improve classification accuracy vision tasks, especially amount training data small. introduce regularizers members broader class jacobian-based regularizers, compare theoretically empirically. straightforward objection minimizing gradient norm training points locally optimal solution, model small gradients training points, may possibly contain large changes regions. demonstrate experiments real synthetic tasks stochastic gradient descent unable find locally optimal globally unproductive solutions. instead, forced find solutions generalize well.",3
"model independence $(\infty,2)$-categorical nerves. models $(\infty,2)$-categories embedding $\infty$-category 2-categories $(\infty,2)$-categories constructed form nerve construction flavor. prove nerve embeddings induce equivalent functors, modulo change model. also show nerve embeddings realize $\infty$-category 2-categories sub-$\infty$-category $(\infty,2)$-categories local respect certain class maps.",10
"all-order resurgence complexified path integral quantum mechanical system integrability. discuss all-order transseries one simplest quantum mechanical systems: u(1) symmetric single-degree-of-freedom system first-order time derivative term. following procedure lefschetz thimble method, explicitly evaluate path integral generating function noether charge derive exact transseries expression. using conservation law, find complex saddle points action, responsible non-perturbative effects resurgence structure model. all-order power-series contributions around saddle point generated one-loop determinant help differential equations obeyed generating function. transseries constructed summing contributions relevant saddle points, identify determining intersection numbers dual thimbles original path integration contour. confirm borel ambiguities perturbation series canceled non-perturbative ambiguities originating discontinuous jumps intersection numbers. transseries computed path-integral formalism agrees exact generating function, whose explicit form obtained operator formalism thanks integrable nature model. agreement indicates non-perturbative completeness transseries obtained semi-classical expansion path integral based lefschetz thimble method.",9
"local area damage detection composite structures using piezoelectric transducers. integrated automated smart structures approach structural health monitoring presented, utilizing array piezoelectric transducers attached embedded within structure actuation sensing. system actively interrogates structure via broadband excitation multiple actuators across desired frequency range. structure's vibration signature characterized computing transfer functions actuator/sensor pair, compared baseline signature. experimental results applying system local area damage detection md explorer rotorcraft composite flexbeam presented.",3
"solving linear inverse problems using gan priors: algorithm provable guarantees. recent works, sparsity-based methods well learning-based methods proven successful solving several challenging linear inverse problems. however, sparsity priors natural signals images suffer poor discriminative capability, learning-based methods seldom provide concrete theoretical guarantees. work, advocate idea replacing hand-crafted priors, sparsity, generative adversarial network (gan) solve linear inverse problems compressive sensing. particular, propose projected gradient descent (pgd) algorithm effective use gan priors linear inverse problems, also provide theoretical guarantees rate convergence algorithm. moreover, show empirically algorithm demonstrates superior performance existing method leveraging gans compressive sensing.",18
"movable cone calabi--yau threefolds ruled fano manifolds. describe explicitly chamber structure movable cone general complete intersection calabi--yau threefold non-split $(n + 4)$-dimensional $\mathbb{p}^{n}$-ruled fano manifold index $n + 1$. moreover, birational minimal models calabi--yau threefolds found whose number finite.",10
"task-driven adaptive statistical compressive sensing gaussian mixture models. framework adaptive non-adaptive statistical compressive sensing developed, statistical model replaces standard sparsity model classical compressive sensing. propose within framework optimal task-specific sensing protocols specifically jointly designed classification reconstruction. two-step adaptive sensing paradigm developed, online sensing applied detect signal class first step, followed reconstruction step adapted detected class observed samples. approach based information theory, tailored gaussian mixture models (gmms), information-theoretic objective relationship sensed signals representation specific task interest maximized. experimental results using synthetic signals, landsat satellite attributes, natural images different sizes different noise levels show improvements achieved using proposed framework compared standard sensing protocols. underlying formulation applied beyond gmms, price higher mathematical computational complexity.",3
"development information criterion change-point analysis. change-point analysis flexible computationally tractable tool analysis times series data systems transition discrete states whose observables corrupted noise. change-point algorithm used identify time indices (change points) system transitions discrete states. present unified information-based approach testing existence change points. new approach reconciles two previously disparate approaches change-point analysis (frequentist information-based) testing transitions states. resulting method statistically principled, parameter prior free widely applicable wide range change-point problems.",14
"fast optical transport ultracold molecules long distances. optically trapped laser-cooled polar molecules hold promise new science technology quantum information quantum simulation. large numerical aperture optical access long trap lifetimes needed many studies, requirements challenging achieve magneto-optical trap (mot) vacuum chamber connected cryogenic buffer gas beam source, case molecule laser cooling experiments far. long distance transport molecules greatly eases fulfilling requirements molecules placed region separate mot chamber. realize fast transport method ultracold molecules based electronically focus-tunable lens combined optical lattice. high transport speed achieved 1d red-detuned optical lattice, generated interference focus-tunable laser beam focus-fixed laser beam. efficiency 48(8)% realized transport ultracold calcium monofluoride (caf) molecules 46 cm distance 50 ms, moderate heating 32(2) {\mu}k 53(4) {\mu}k. positional stability molecular cloud allows stable loading optical tweezer array single molecules.",14
"local support vector machines:formulation analysis. provide formulation local support vector machines (lsvms) generalizes previous formulations, brings explicit connections local polynomial learning used nonparametric estimation literature. investigate simplest type lsvms called local linear support vector machines (llsvms). first time establish conditions llsvms make bayes consistent predictions test point $x_0$. also establish rates local risk llsvms converges minimum value expected local risk point $x_0$. using stability arguments establish generalization error bounds llsvms.",18
"correspondence factor analysis big data sets: case study 30 million words; contrasting analytics using apache solr correspondence analysis r. consider large number text data sets. cooking recipes. term distribution distributional properties data investigated. aim look various analytical approaches allow mining information high low detail scales. metric space embedding fundamental interest semantic properties data. consider projection data analyses aggregated versions data. contrast projection aggregated versions data analyses data. analogously term set, look analysis selected terms. also look inherent term associations singular plural. addition use correspondence analysis r, latent semantic space mapping, also use apache solr. setting solr server carrying querying described. novelty querying supported solr based principal factor plane mapping data. uses bounding box query, based factor projections.",3
"image restoration: general wavelet frame based model asymptotic analysis. image restoration one important areas imaging science. mathematical tools widely used image restoration, wavelet frame based approach one successful examples. paper, introduce generic wavelet frame based image restoration model, called ""general model"", includes existing wavelet frame based models special cases. moreover, general model also includes examples new literature. motivated earlier studies [1-3], provide asymptotic analysis general model image resolution goes infinity, establishes connection general model discrete setting new variatonal model continuum setting. variational model also includes existing variational models special cases, total generalized variational model proposed [4]. end, introduce algorithm solving general model present one numerical simulation example.",10
"model selection nonlinear embedding unsupervised domain adaptation. domain adaptation deals adapting classifiers trained data source distribution, work effectively data target distribution. paper, introduce nonlinear embedding transform (net) unsupervised domain adaptation. net reduces cross-domain disparity nonlinear domain alignment. also embeds domain-aligned data similar data points clustered together. results enhanced classification. determine parameters net model (and unsupervised domain adaptation models), introduce validation procedure sampling source data points similar distribution target data. test net validation procedure using popular image datasets compare classification results across competitive procedures unsupervised domain adaptation.",3
max-plus statistical leverage scores. statistical leverage scores complex matrix $a\in\mathbb{c}^{n\times d}$ record degree alignment col$(a)$ coordinate axes $\mathbb{c}^n$. score used random sampling algorithms solving certain numerical linear algebra problems. paper present max-plus algebraic analogue statistical leverage scores. show max-plus statistical leverage scores used calculate exact asymptotic behavior conventional statistical leverage scores generic matrices puiseux series also provide novel way approximate conventional statistical leverage scores fixed complex matrix. advantage approximating complex matrices scores max-plus scores max-plus scores computed quickly. approximation typically accurate within order magnitude useful practical problems true scores known vary widely.,18
"lossless compression angiogram foreground visual quality preservation background. increasing volume telemedicine information, need medical image compression become important. angiographic images, small ratio entire image usually belongs vasculature provides crucial information diagnosis. parts image diagnostically less important compressed higher compression ratio. however, quality parts affect visual perception image well. existing methods compress foreground background angiographic images using different techniques. paper first utilize convolutional neural network segment vessels represent hierarchical block processing algorithm capable eliminating background redundancies preserving overall visual quality angiograms.",3
"optimization first-order surrogate functions. paper, study optimization methods consisting iteratively minimizing surrogates objective function. proposing several algorithmic variants simple convergence analyses, make two main contributions. first, provide unified viewpoint several first-order optimization techniques accelerated proximal gradient, block coordinate descent, frank-wolfe algorithms. second, introduce new incremental scheme experimentally matches outperforms state-of-the-art solvers large-scale optimization problems typically arising machine learning.",18
"local variation statistical hypothesis test. goal image oversegmentation divide image several pieces, ideally part object. one simplest yet effective oversegmentation algorithms known local variation (lv) (felzenszwalb huttenlocher 2004). work, study algorithm show algorithms similar lv devised applying different statistical models decisions, thus providing theoretical justification well-founded explanation unexpected high performance lv approach. algorithms based statistics natural images hypothesis testing decision; denote algorithms probabilistic local variation (plv). best plv algorithm, relies censored estimation, presents state-of-the-art results keeping computational complexity lv algorithm.",3
"deep semantic classification 3d lidar data. robots expected operate autonomously dynamic environments. understanding underlying dynamic characteristics objects key enabler achieving goal. paper, propose method pointwise semantic classification 3d lidar data three classes: non-movable, movable dynamic. concentrate understanding specific semantics characterize important information required autonomous system. non-movable points scene belong unchanging segments environment, whereas remaining classes corresponds changing parts scene. difference movable dynamic class motion state. dynamic points perceived moving, whereas movable objects move, perceived static. learn distinction movable non-movable points environment, introduce approach based deep neural network detecting dynamic points, estimate pointwise motion. propose bayes filter framework combining learned semantic cues motion cues infer required semantic classification. extensive experiments, compare approach methods standard benchmark dataset report competitive results comparison existing state-of-the-art. furthermore, show improvement classification points combining semantic cues retrieved neural network motion cues.",3
"global attractor weakly damped forced kawahara equation torus. study long time behaviour solutions weakly damped forced kawahara equation torus. precisely, prove existence global attractor $l^2$, time passes solutions draw closer. fact, show global attractor turns lie smoother space $h^2$ bounded therein. further, give upper bound size attractor $h^2$ depends damping parameter norm forcing term.",10
"learning multi-domain convolutional neural networks visual tracking. propose novel visual tracking algorithm based representations discriminatively trained convolutional neural network (cnn). algorithm pretrains cnn using large set videos tracking ground-truths obtain generic target representation. network composed shared layers multiple branches domain-specific layers, domains correspond individual training sequences branch responsible binary classification identify target domain. train network respect domain iteratively obtain generic target representations shared layers. tracking target new sequence, construct new network combining shared layers pretrained cnn new binary classification layer, updated online. online tracking performed evaluating candidate windows randomly sampled around previous target state. proposed algorithm illustrates outstanding performance compared state-of-the-art methods existing tracking benchmarks.",3
"regularization wasserstein gans. since invention, generative adversarial networks (gans) become popular approach learning model distribution real (unlabeled) data. convergence problems training overcome wasserstein gans minimize distance model empirical distribution terms different metric, thereby introduce lipschitz constraint optimization problem. simple way enforce lipschitz constraint class functions, modeled neural network, weight clipping. proposed training improved instead augmenting loss regularization term penalizes deviation gradient critic (as function network's input) one. present theoretical arguments using weaker regularization term enforcing lipschitz constraint preferable. arguments supported experimental results toy data sets.",18
"patterns scalable bayesian inference. datasets growing size complexity, creating demand rich models quantification uncertainty. bayesian methods excellent fit demand, scaling bayesian inference challenge. response challenge, considerable recent work based varying assumptions model structure, underlying computational resources, importance asymptotic correctness. result, zoo ideas clear overarching principles. paper, seek identify unifying principles, patterns, intuitions scaling bayesian inference. review existing work utilizing modern computing resources mcmc variational approximation techniques. taxonomy ideas, characterize general principles proven successful designing scalable inference procedures comment path forward.",18
"genetic algorithms astronomy astrophysics. genetic algorithms (gas) emulate process biological evolution, computational setting, order generate good solutions difficult search optimisation problems. ga-based optimisers tend extremely robust versatile compared traditional techniques used solve optimisation problems. review paper provides brief introduction gas outlines utility astronomy astrophysics.",0
"stochastic gradient method accelerated stochastic dynamics. paper, propose novel technique implement stochastic gradient methods, beneficial learning large datasets, accelerated stochastic dynamics. stochastic gradient method based mini-batch learning reducing computational cost amount data large. stochasticity gradient mitigated injection gaussian noise, yields stochastic langevin gradient method; method used bayesian posterior sampling. however, performance stochastic langevin gradient method depends mixing rate stochastic dynamics. study, propose violating detailed balance condition enhance mixing rate. recent studies revealed violating detailed balance condition accelerates convergence stationary state reduces correlation time samplings. implement violation detailed balance condition stochastic gradient langevin method test method simple model demonstrate performance.",18
"functional principal component analysis randomized sparse clustering algorithm medical image analysis. due advances sensors, growing large complex medical image data ability visualize pathological change cellular even molecular level anatomical changes tissues organs. consequence, medical images potential enhance diagnosis disease, prediction clinical outcomes, characterization disease progression, management health care development treatments, also pose great methodological computational challenges representation selection features image cluster analysis. address challenges, first extend one dimensional functional principal component analysis two dimensional functional principle component analyses (2dfpca) fully capture space variation image signals. image signals contain large number redundant irrelevant features provide additional useful information cluster analysis. widely used methods removing redundant irrelevant features sparse clustering algorithms using lasso-type penalty select features. however, accuracy clustering using lasso-type penalty depends select penalty parameters threshold selecting features. practice, difficult determine. recently, randomized algorithms received great deal attention big data analysis. paper presents randomized algorithm accurate feature selection image cluster analysis. proposed method applied ovarian kidney cancer histology image data tcga database. results demonstrate randomized feature selection method coupled functional principal component analysis substantially outperforms current sparse clustering algorithms image cluster analysis.",18
"adaptation learning networks nonlinear system modeling. chapter, analyze nonlinear filtering problems distributed environments, e.g., sensor networks peer-to-peer protocols. scenarios, agents environment receive measurements streaming fashion, required estimate common (nonlinear) model alternating local computations communications neighbors. focus important distinction single-task problems, underlying model common agents, multitask problems, agent might converge different model due to, e.g., spatial dependencies factors. currently, literature distributed learning nonlinear case focused single-task case, may strong limitation real-world scenarios. introducing problem reviewing existing approaches, describe simple kernel-based algorithm tailored multitask case. evaluate proposal simulated benchmark task, conclude detailing currently open problems lines research.",18
"provably fast accurate recovery evolutionary trees harmonic greedy triplets. give greedy learning algorithm reconstructing evolutionary tree based certain harmonic average triplets terminal taxa. pairwise distances terminal taxa estimated sequence data, algorithm runs o(n^2) time using o(n) work space, n number terminal taxa. time space complexities optimal sense size input distance matrix n^2 size output tree n. moreover, jukes-cantor model evolution, algorithm recovers correct tree topology high probability using sample sequences length polynomial (1) n, (2) logarithm error probability, (3) inverses two small parameters.",3
"sentiment analysis based rhetorical structure theory: learning deep neural networks discourse trees. prominent applications sentiment analysis countless, covering areas marketing, customer service communication. conventional bag-of-words approach measuring sentiment merely counts term frequencies; however, neglects position terms within discourse. remedy, develop discourse-aware method builds upon discourse structure documents. purpose, utilize rhetorical structure theory label (sub-)clauses according hierarchical relationships assign polarity scores individual leaves. learn resulting rhetorical structure, propose tensor-based, tree-structured deep neural network (named discourse-lstm) order process complete discourse tree. underlying attention mechanism infers salient passages narrative materials. addition, suggest two algorithms data augmentation (node reordering artificial leaf insertion) increase training set reduce overfitting. benchmarks demonstrate superior performance approach. moreover, attention mechanism reveals salient text passages thereby provides explanatory insights.",3
"interpretability artificial neural network models artificial intelligence vs. neuroscience. computationally explicit hypotheses brain function derived machine learning (ml)-based models recently revolutionized neuroscience. despite unprecedented ability artificial neural networks (anns) capture responses biological neural networks (brains), full access internal model components (unlike brain), anns often referred black-boxes limited interpretability. interpretability, however, multi-faceted construct used differently across fields. particular, interpretability, explainability, efforts artificial intelligence (ai) focus understanding different model components contribute output (i.e., decision making). contrast, neuroscientific interpretability anns requires explicit alignment model components neuroscientific constructs (e.g., different brain areas phenomena, like recurrence top-down feedback). given widespread calls improve interpretability ai systems, highlight different notions interpretability argue neuroscientific interpretability anns pursued parallel with, independently from, ongoing efforts ai. certain ml techniques (e.g., deep dream) leveraged fields, ask stimulus optimally activates specific model features (feature visualization optimization), different features contribute model's output (feature attribution). however, without appropriate brain alignment, certain features remain uninterpretable neuroscientists.",15
"fast semantic segmentation video using motion vector-based feature interpolation. models optimized accuracy challenging, dense prediction tasks semantic segmentation entail significant inference costs, prohibitively slow run frame video. since nearby video frames spatially similar, however, substantial opportunity reuse computation. existing work explored basic feature reuse feature warping based optical flow, encountered limits speedup attainable techniques. paper, present new, two part approach accelerating inference video. firstly, propose fast feature propagation scheme utilizes block motion vector maps present compressed video cheaply propagate features frame frame. secondly, develop novel feature estimation scheme, termed feature interpolation, fuses features propagated enclosing keyframes render accurate feature estimates, even sparse keyframe frequencies. evaluate system cityscapes dataset, comparing frame-by-frame baseline related work. find able substantially accelerate segmentation video, achieving almost twice average inference speed prior work target accuracy level.",3
"early universe quantum gravity. present new, testable picture early universe finite nonlocal quantum gravity, conformally invariant classical quantum levels. high-energy regime theory consists two phases, conformally invariant trans-planckian phase sub-planckian higgs phase described action quadratic ricci tensor cosmos evolves according standard radiation-dominated model. first phase, issues hot big bang singularity, flatness, horizon problems find universal simple non-inflationary solution means weyl conformal invariance, regardless microscopic details theory. second phase, conformal symmetry spontaneously broken, logarithmic quantum corrections action make primordial tensor spectrum (from graviton fluctuations) scalar spectrum (from thermal fluctuations) quasi scale invariant. although nonlocal quantum gravity explicit realization scenario, finite conformal theory quadratic limit would yield results. particular, value scalar spectral index depends beta function scale ratio, consistent observations. theory also predicts positive tensor spectral index $n_{\rm t} =1-n_{\rm s}$, fixed tensor-to-scalar ratio $r_{0.05}\gtrsim 0.01$, blue-tilted stochastic gravitational-wave background, universal predictions testable immediate near future.",6
"$a^{4}nt$: author attribute anonymity adversarial training neural machine translation. text-based analysis methods allow reveal privacy relevant author attributes gender, age identify text's author. methods compromise privacy anonymous author even author tries remove privacy sensitive content. paper, propose automatic method, called adversarial author attribute anonymity neural translation ($a^4nt$), combat text-based adversaries. combine sequence-to-sequence language models used machine translation generative adversarial networks obfuscate author attributes. unlike machine translation techniques need paired data, method trained unpaired corpora text containing different authors. importantly, propose evaluate techniques impose constraints $a^4nt$ preserve semantics input text. $a^4nt$ learns make minimal changes input text successfully fool author attribute classifiers, aiming maintain meaning input. show experiments two different datasets three settings proposed method effective fooling author attribute classifiers thereby improving anonymity authors.",3
"short-term mortality prediction elderly patients using medicare claims data. risk prediction central clinical medicine public health. many machine learning models developed predict mortality, rarely applied clinical literature, classification tasks typically rely logistic regression. one reason existing machine learning models often seek optimize predictions incorporating features present databases readily available providers policy makers, limiting generalizability implementation. tested number machine learning classifiers prediction six-month mortality population elderly medicare beneficiaries, using administrative claims database kind available majority health care payers providers. show machine learning classifiers substantially outperform current widely-used methods risk prediction used improved feature set incorporating insights clinical medicine, developed study. work applications supporting patient provider decision making end life, well population health-oriented efforts identify patients high risk poor outcomes.",18
"predicting parameters deep learning. demonstrate significant redundancy parameterization several deep learning models. given weight values feature possible accurately predict remaining values. moreover, show parameter values predicted, many need learned all. train several different architectures learning small number weights predicting rest. best case able predict 95% weights network without drop accuracy.",3
"generalization machine learning via analytical learning theory. paper introduces novel measure-theoretic learning theory analyze generalization behaviors practical interest. proposed learning theory following abilities: 1) utilize qualities learned representation path raw inputs outputs representation learning, 2) guarantee good generalization errors possibly arbitrarily rich hypothesis spaces (e.g., arbitrarily large capacity rademacher complexity) non-stable/non-robust learning algorithms, 3) clearly distinguish individual problem instance other. generalization bounds relative representation data, hold true even representation learned. discuss several consequences results deep learning, one-shot learning curriculum learning. unlike statistical learning theory, proposed learning theory analyzes problem instance individually via measure theory, rather set problem instances via statistics. differences assumptions objectives, proposed learning theory meant complementary previous learning theory designed compete it.",18
"unusual gas structure otherwise normal spiral galaxy hosting grb 171205a / sn 2017iuk. study structure atomic hydrogen (hi) host galaxy grb 171205a / sn 2017iuk z=0.037 hi 21cm emission line observations karl g. jansky large array. observations reveal unusual morphology kinematics hi otherwise apparently normal galaxy. high column density, cold hi absent extended north-south region passing optical centre galaxy, instead extended towards south, sides galaxy. moreover, hi kinematics show continuous change along major axis galaxy expected classical rotating disk. explore several scenarios explain hi structure kinematics galaxy: feedback central starburst and/or active galactic nucleus, ram pressure stripping, accretion, tidal interaction companion galaxy. options ruled out. viable remaining explanation penetrating passage satellite disk myr ago, redistributing hi grb host without yet affecting stellar distribution. also lead rapid formation peculiar stars due violent induced shock. location grb 171205a vicinity distorted area suggests progenitor star(s) originated extreme conditions share origin peculiarities hi. could explain atypical location grb 171205a host galaxy.",0
"numerical sensitivity efficiency treatment epistemic aleatory uncertainty. treatment aleatory epistemic uncertainty recent methods often requires high computational effort. abstract, propose numerical sampling method allowing lighten computational burden treating information means so-called fuzzy random variables.",3
"alternating direction method multipliers sparse convolutional neural networks. storage computation requirements convolutional neural networks (cnns) prohibitive exploiting models low-power embedded devices. paper reduces computational complexity cnns minimizing objective function, including recognition loss augmented sparsity-promoting penalty term. sparsity structure network identified using alternating direction method multipliers (admm), widely used large optimization problems. method alternates promoting sparsity network optimizing recognition performance, allows us exploit two-part structure corresponding objective functions. particular, take advantage separability sparsity-inducing penalty functions decompose minimization problem sub-problems solved sequentially. applying method variety state-of-the-art cnn models, proposed method able simplify original model, generating models less computation fewer parameters, maintaining often improving generalization performance. accomplishments variety models strongly verify proposed admm-based method useful tool simplifying improving deep cnns.",3
"protocols learning classifiers distributed data. consider problem learning classifiers labeled data distributed across several nodes. goal find single classifier, small approximation error, across datasets minimizing communication nodes. setting models real-world communication bottlenecks processing massive distributed datasets. present several general sampling-based solutions well two-way protocols provable exponential speed-up one-way protocol. focus core problems noiseless data distributed across two nodes. techniques introduce reminiscent active learning, rather actively probing labels, nodes actively communicate other, node simultaneously learning important data another node.",18
"arbitrary-oriented scene text detection via rotation proposals. paper introduces novel rotation-based framework arbitrary-oriented text detection natural scene images. present rotation region proposal networks (rrpn), designed generate inclined proposals text orientation angle information. angle information adapted bounding box regression make proposals accurately fit text region terms orientation. rotation region-of-interest (rroi) pooling layer proposed project arbitrary-oriented proposals feature map text region classifier. whole framework built upon region-proposal-based architecture, ensures computational efficiency arbitrary-oriented text detection compared previous text detection systems. conduct experiments using rotation-based framework three real-world scene text detection datasets demonstrate superiority terms effectiveness efficiency previous approaches.",3
"existence boundary behaviour radial solutions weighted elliptic systems gradient terms. concerned existence boundary behaviour positive radial solutions system \begin{equation*} \left\{ \begin{aligned} \delta u&=|x|^{a}v^{p} &&\quad\mbox{ } \omega, \\ \delta v&=|x|^{b}v^{q}f(|\nabla u|) &&\quad\mbox{ } \omega, \end{aligned} \right. \end{equation*} $\omega \subset \br^n$ either ball centered origin whole space $\br^n$, $a$, $b$, $p$, $q> 0$, $f \in c^1[0, \infty)$ increasing function $f(t)> 0$ $t> 0$. firstly, study existence positive radial solutions case system posed ball corresponding behaviour boundary. next, take $f(t) = t^s$, $s> 1$, $\omega = \br^n$ use dynamical system techniques able describe behaviour infinity positive radial solutions.",10
"fast sample efficient inductive matrix completion via multi-phase procrustes flow. revisit inductive matrix completion problem aims recover rank-$r$ matrix ambient dimension $d$ given $n$ features side prior information. goal make use known $n$ features reduce sample computational complexities. present analyze new gradient-based non-convex optimization algorithm converges true underlying matrix linear rate sample complexity linearly depending $n$ logarithmically depending $d$. best knowledge, previous algorithms either quadratic dependency number features sample complexity sub-linear computational convergence rate. addition, provide experiments synthetic real world data demonstrate effectiveness proposed algorithm.",18
stochastic quasi-newton adaptive step lengths large-scale problems. provide numerically robust fast method capable exploiting local geometry solving large-scale stochastic optimisation problems. key innovation auxiliary variable construction coupled inverse hessian approximation computed using receding history iterates gradients. markov chain nature classic stochastic gradient algorithm enables development. construction offers mechanism stochastic line search adapting step length. numerically evaluate compare current state-of-the-art encouraging performance real-world benchmark problems number observations unknowns order millions.,18
"empirical study budget-aware online kernel algorithms streams graphs. kernel methods considered effective technique on-line learning. many approaches developed compactly representing dual solution kernel method problem imposes memory constraints. however, literature work specifically tailored streams graphs. motivated fact size feature space representation many state-of-the-art graph kernels relatively small thus explicitly computable, study whether executing kernel algorithms feature space effective classical dual approach. study three different algorithms various strategies managing budget. efficiency efficacy proposed approaches experimentally assessed relatively large graph streams exhibiting concept drift. turns that, strict memory budget constraints enforced, working feature space, given current state art graph kernels, viable alternative dual approaches, terms speed classification performance.",3
"influence mass black holes radio/x-ray correlation diagram. radio/x-ray correlation diagram, black hole x-ray binaries (bhxbs) hard state, depicts connection might exists radio jets x-ray emitting accretion discs. current version radio/x-ray correlation diagram shows two populations bhxbs. two populations evolve along two different correlation tracks ($f_{rad}= \kappa f_{x}^{b}$), namely standard track outliers. past years, key question explain existence two tracks. paper, investigate impact black hole mass radio/x-ray correlation sample 17 bhxbs. led conclude least one following full consideration large uncertainties mass estimates:(i) reported mass estimates black holes incorrect insufficiently accurate infer impact radio/x-ray correlation diagram. (ii) estimated radio luminosities x-ray luminosites still reliable enough two reasons. one reason due lack associated errors observational data. another sources might still transit one track other. (iii) mass bh significant influence track, source belongs to, radio/x-ray correlation diagram.",0
"wikireading: novel large-scale language understanding task wikipedia. present wikireading, large-scale natural language understanding task publicly-available dataset 18 million instances. task predict textual values structured knowledge base wikidata reading text corresponding wikipedia articles. task contains rich variety challenging classification extraction sub-tasks, making well-suited end-to-end models deep neural networks (dnns). compare various state-of-the-art dnn-based architectures document classification, information extraction, question answering. find models supporting rich answer space, word character sequences, perform best. best-performing model, word-level sequence sequence model mechanism copy out-of-vocabulary words, obtains accuracy 71.8%.",3
"existence boundedness minimizing measures general form non-local energies. paper consider general form non-local energy integral form, covers usual ones (for instance, sum positive negative power). instead admitting sets, $l^\infty$ functions, admissible objects, define energy radon measures. prove existence optimal measures wide generality, show several cases optimal measures actually $l^\infty$ functions, providing priori bound norm. also derive uniqueness result minimizers, discuss possible form optimal functions $1$-dimensional case (which far trivial, lot still remains open).",10
"survey resampling techniques improving classification performance unbalanced datasets. number classification problems need deal data imbalance classes. often desired high recall minority class maintaining high precision majority class. paper, review number resampling techniques proposed literature handle unbalanced datasets study effect classification performance.",18
"queue-aware distributive resource control delay-sensitive two-hop mimo cooperative systems. paper, consider queue-aware distributive resource control algorithm two-hop mimo cooperative systems. shall illustrate relay buffering effective way reduce intrinsic half-duplex penalty cooperative systems. complex interactions queues source node relays modeled average-cost infinite horizon markov decision process (mdp). traditional approach solving mdp problem involves centralized control huge complexity. obtain distributive low complexity solution, introduce linear structure approximates value function associated bellman equation sum per-node value functions. derive distributive two-stage two-winner auction-based control policy function local csi local qsi only. furthermore, estimate best fit approximation parameter, propose distributive online stochastic learning algorithm using stochastic approximation theory. finally, establish technical conditions almost-sure convergence show heavy traffic, proposed low complexity distributive control global optimal.",3
"propositional abduction implicit hitting sets. logic-based abduction finds important applications artificial intelligence related areas. one application example finding explanations observed phenomena. propositional abduction restriction abduction propositional domain, complexity-wise second level polynomial hierarchy. recent work shown exploiting implicit hitting sets propositional satisfiability (sat) solvers provides efficient approach propositional abduction. paper investigates earlier work proposes number algorithmic improvements. improvements shown yield exponential reductions number sat solver calls. importantly, experimental results show significant performance improvements compared best approaches propositional abduction.",3
fitness landscape analysis dynamic resource allocation multiuser ofdm based cognitive radio systems. paper withdrawn.,3
"isomorphisms properties tar reconfiguration graphs zero forcing $x$-set parameters. $x$-tar (token addition/removal) reconfiguration graph vertices sets satisfy property $x$, edge two sets one obtained adding removing one element. paper considers $x$-tar graph $x-$ sets vertices base graph $g$ $x$-sets $g$ must satisfy certain conditions. dominating sets, power dominating sets, zero forcing sets, positive semidefinite zero forcing sets examples $x$-sets. graphs $g$ $g'$ isolated vertices, shown $g$ $g'$ isomorphic $x$-tar reconfiguration graphs relabeling vertices $g'$ $g$ $g'$ exactly $x$-sets. concept $x$-irrelevant vertex introduced facilitate analysis $x$-tar graph isomorphisms. furthermore, results related connectedness zero forcing tar graph given. present families graphs exceed known lower bounds connectedness parameters.",10
"intentions vulnerable road users - detection forecasting means machine learning. avoiding collisions vulnerable road users (vrus) using sensor-based early recognition critical situations one manifold opportunities provided current development field intelligent vehicles. especially pedestrians cyclists agile variety movement options, modeling behavior traffic scenes challenging task. article propose movement models based machine learning methods, particular artificial neural networks, order classify current motion state predict future trajectory vrus. model types also combined enable application specifically trained motion predictors based continuously updated pseudo probabilistic state classification. furthermore, architecture used evaluate motion-specific physical models starting stopping video-based pedestrian motion classification. comprehensive dataset consisting 1068 pedestrian 494 cyclist scenes acquired urban intersection used optimization, training, evaluation different models. results show substantial higher classification rates ability earlier recognize motion state changes machine learning approaches compared interacting multiple model (imm) kalman filtering. trajectory prediction quality also improved kinds test scenes, especially starting stopping motions included. here, 37\% 41\% lower position errors achieved average, respectively.",3
"stnet: selective tuning convolutional networks object localization. visual attention modeling recently gained momentum developing visual hierarchies provided convolutional neural networks. despite recent successes feedforward processing abstraction concepts form raw images, inherent nature feedback processing remained computationally controversial. inspired computational models covert visual attention, propose selective tuning convolutional networks (stnet). composed streams bottom-up top-down information processing selectively tune visual representation convolutional networks. experimentally evaluate performance stnet weakly-supervised localization task imagenet benchmark dataset. demonstrate stnet successfully surpasses state-of-the-art results also generates attention-driven class hypothesis maps.",3
"dialogue session segmentation embedding-enhanced texttiling. human-computer conversation systems, context user-issued utterance particularly important provides useful background information conversation. however, unwise track previous utterances current session equally important. paper, address problem session segmentation. propose embedding-enhanced texttiling approach, inspired observation conversation utterances highly noisy, word embeddings provide robust way capturing semantics. experimental results show approach achieves better performance texttiling, mmd approaches.",3
"high-dimensional vector semantics. paper explore ""vector semantics"" problem perspective ""almost orthogonal"" property high-dimensional random vectors. show intriguing property used ""memorize"" random vectors simply adding them, provide efficient probabilistic solution set membership problem. also, discuss several applications word context vector embeddings, document sentences similarity, spam filtering.",3
"high-accuracy total variation compressed video sensing. numerous total variation (tv) regularizers, engaged image restoration problem, encode gradients means simple $[-1,1]$ fir filter. despite low computational processing, filter severely deviates signal's high frequency components pertinent edge/discontinuous information cause several deficiency issues known texture geometric loss. paper addresses problem proposing alternative model tv regularization problem via high order accuracy differential fir filters preserve rapid transitions signal recovery. numerical encoding scheme designed extend tv model multidimensional representation (tensorial decomposition). adopt design regulate spatial temporal redundancy compressed video sensing problem jointly recover frames under-sampled measurements. seek solution via alternating direction methods multipliers find unique solution quadratic minimization step capability handling different boundary conditions. resulting algorithm uses much lower sampling rate highly outperforms alternative state-of-the-art methods. evaluated terms restoration accuracy visual quality recovered frames.",10
"probabilistic structure discovery time series data. existing methods structure discovery time series data construct interpretable, compositional kernels gaussian process regression models. learned gaussian process model provides posterior mean variance estimates, typically structure learned via greedy optimization procedure. restricts space possible solutions leads over-confident uncertainty estimates. introduce fully bayesian approach, inferring full posterior structures, reliably captures uncertainty model.",18
"measuring star formation black hole accretion rates tandem using mid-infrared spectra local infrared-luminous galaxies. present results stacking analysis performed spitzer/infrared spectrograph high-resolution mid-infrared spectra luminous infrared galaxies (lirgs) great observatories all-sky lirg survey (goals). binning mid-infrared active galactic nucleus (agn) fraction stacking spectra, detect bright emission lines [ne ii] [ne iii], trace star formation, fainter emission lines [ne v] [o iv], trace agn activity, throughout sample. find [ne ii] luminosity fairly constant across agn fraction bins, [o iv] [ne v] luminosities increase order magnitude. measured average line ratios, [ne v]/[ne ii] [o iv]/[ne ii], low agn fraction similar h ii galaxies line ratios high agn fraction similar liners seyferts. decompose [o iv] luminosity star-formation agn components fitting [o iv] luminosity function [ne ii] luminosity mid-infrared agn fraction. [o iv] luminosity lirgs dominated star formation mid-infrared agn fractions $\lesssim0.3$. corrected [o iv] luminosity, calculate black hole accretion rates ranging $10^{-5}$ m$_{\odot}$/yr low agn fractions 0.2 m$_{\odot}$/yr highest agn fractions. find using [o iv] luminosity, without correcting star formation, lead overestimate bhar factor 30 starburst dominated lirgs. finally, show bhar/sfr ratio increases three orders magnitude function mid-infrared agn fraction lirgs.",0
"deep learning classification lung nodules. deep learning, promising new area machine learning, attracted rapidly increasing attention field medical imaging. compared conventional machine learning methods, deep learning requires hand-tuned feature extractor, shown superior performance many visual object recognition applications. study, develop deep convolutional neural network (cnn) apply thoracic ct images classification lung nodules. present cnn architecture classification accuracy original images lung nodules. order understand features lung nodules, construct new datasets, based combination artificial geometric nodules transformations original images, well stochastic nodule shape model. found simplistic geometric nodules cannot capture important features lung nodules.",15
"efficient exploration bayesian deep q-networks. propose bayesian deep q-network (bdqn), practical thompson sampling based reinforcement learning (rl) algorithm. thompson sampling allows targeted exploration high dimensions posterior sampling usually computationally expensive. address limitation introducing uncertainty output layer network bayesian linear regression (blr) model. layer trained fast closed-form updates samples drawn efficiently gaussian distribution. apply method wide range atari games arcade learning environments. since bdqn carries efficient exploration, able reach higher rewards substantially faster key baseline, double deep q network (ddqn).",3
"tagging multimedia stimuli ontologies. successful management emotional stimuli pivotal issue concerning affective computing (ac) related research. subfield artificial intelligence, ac concerned design computer systems accompanying hardware recognize, interpret, process human emotions, also development systems trigger human emotional response ordered controlled manner. requires maximum attainable precision efficiency extraction data emotionally annotated databases databases use keywords tags description semantic content, provide either necessary flexibility leverage needed efficiently extract pertinent emotional content. therefore, extent propose introduction ontologies new paradigm description emotionally annotated data. ability select sequence data based semantic attributes vital study involving metadata, semantics ontological sorting like semantic web social semantic desktop, approach described paper facilitates reuse areas well.",3
"survey description logic ontologies benefit formal concept analysis. although notion concept collection objects sharing certain properties, notion conceptual hierarchy fundamental formal concept analysis description logics, ways concepts described obtained differ significantly two research areas. despite differences, several attempts bridge gap two formalisms, attempts apply methods one field other. present work aims give overview research done combining description logics formal concept analysis.",3
"conservative binary dynamics spinning black hole $\mathcal{o}(g^3)$ scattering amplitudes. compute conservative two-body hamiltonian compact binary system spinning black hole $\mathcal{o}(g^3)$ orders velocity, including linear quadratic spin terms. obtain results calculate classical limit two-loop amplitude scattering massive scalar particle massive spin-1 particle minimally coupled gravity. employ modern scattering amplitude loop integration techniques, particular numerical unitarity, integration-by-parts identities, method regions. conservative potential terms rest-frame spin vectors extracted matching non-relativistic effective field theory. also apply kosower-maybee-o'connell (kmoc) formalism calculate impulse covariant spin formalism directly amplitude. work systematically conventional dimensional regularization explicitly evaluate divergent integrals appear full- effective-theory amplitudes, well phase-space integrals arise kmoc formalism.",9
"text segmentation based semantic word embeddings. explore use semantic word embeddings text segmentation algorithms, including c99 segmentation algorithm new algorithms inspired distributed word vector representation. developing general framework discussing class segmentation objectives, study effectiveness greedy versus exact optimization approaches suggest new iterative refinement technique improving performance greedy strategies. compare results known benchmarks, using known metrics. demonstrate state-of-the-art performance untrained method content vector segmentation (cvs) choi test set. finally, apply segmentation procedure in-the-wild dataset consisting text extracted scholarly articles arxiv.org database.",3
"deep learning medical image segmentation. report provides overview current state art deep learning architectures optimisation techniques, uses adni hippocampus mri dataset example compare effectiveness efficiency different convolutional architectures task patch-based 3-dimensional hippocampal segmentation, important diagnosis alzheimer's disease. found slightly unconventional ""stacked 2d"" approach provides much better classification performance simple 2d patches without requiring significantly computational power. also examined popular ""tri-planar"" approach used recently published studies, found provides much better results 2d approaches, also moderate increase computational power requirement. finally, evaluated full 3d convolutional architecture, found provides marginally better results tri-planar approach, cost significant increase computational power requirement.",3
"scaling laws fluctuations statistics word frequencies. paper combine statistical analysis large text databases simple stochastic models explain appearance scaling laws statistics word frequencies. besides sublinear scaling vocabulary size database size (heaps' law), report new scaling fluctuations around average (fluctuation scaling analysis). explain scaling laws modeling usage words simple stochastic processes overall distribution word-frequencies fat tailed (zipf's law) frequency single word subject fluctuations across documents (as topic models). framework, mean variance vocabulary size expressed quenched averages, implying that: i) inhomogeneous dissemination words cause reduction average vocabulary size comparison homogeneous case, ii) correlations co-occurrence words lead increase variance vocabulary size becomes non-self-averaging quantity. address implications observations measurement lexical richness. test results three large text databases (google-ngram, enlgish wikipedia, collection scientific articles).",14
"basics positively multiplicative graphs algebras. graph said positively multiplicative adjacency matrix $a$ embeds matrix algebra admitting basis $\mathfrak{b}$ nonnegative structure constants matrix multiplication $a$ coincides $a$. goal paper present basics notion expose, various simple examples, relates highly non trivial problems like combinatorial description fusion rules, determination spectrum cayley graphs, description minimal boundary graded graphs study random walks alcove tilings.",10
"extended littlestone's dimension learning mistakes abstentions. paper studies classification abstention option online setting. setting, examples arrive sequentially, learner given hypothesis class $\mathcal h$, goal learner either predict label example abstain, ensuring make pre-specified number mistakes predict label. previous work problem left open two main challenges. first, much known optimality algorithms, particular, optimal algorithmic strategy individual hypothesis class. second, realizable case studied, realistic non-realizable scenario well-understood. paper, address challenges. first, provide novel measure, called extended littlestone's dimension, captures number abstentions needed ensure certain number mistakes. second, explore non-realizable case, provide upper lower bounds number abstentions required algorithm guarantee specified number mistakes.",3
"few-example object detection model communication. paper, study object detection using large pool unlabeled images labeled images per category, named ""few-example object detection"". key challenge consists generating trustworthy training samples many possible pool. using training examples seeds, method iterates model training high-confidence sample selection. training, easy samples generated first and, poorly initialized model undergoes improvement. model becomes discriminative, challenging reliable samples selected. that, another round model improvement takes place. improve precision recall generated training samples, embed multiple detection models framework, proven outperform single model baseline model ensemble method. experiments pascal voc'07, ms coco'14, ilsvrc'13 indicate using three four samples selected category, method produces competitive results compared state-of-the-art weakly-supervised approaches using large number image-level labels.",3
"online learning switching costs adaptive adversaries. study power different types adaptive (nonoblivious) adversaries setting prediction expert advice, full-information bandit feedback. measure player's performance using new notion regret, also known policy regret, better captures adversary's adaptiveness player's behavior. setting losses allowed drift, characterize ---in nearly complete manner--- power adaptive adversaries bounded memories switching costs. particular, show switching costs, attainable rate bandit feedback $\widetilde{\theta}(t^{2/3})$. interestingly, rate significantly worse $\theta(\sqrt{t})$ rate attainable switching costs full-information case. via novel reduction experts bandits, also show bounded memory adversary force $\widetilde{\theta}(t^{2/3})$ regret even full information case, proving switching costs easier control bounded memory adversaries. lower bounds rely new stochastic adversary strategy generates loss processes strong dependencies.",3
"efficient tracking growing number experts. consider variation problem prediction expert advice, new forecasters unknown may appear round. often prediction expert advice, designing algorithm achieves near-optimal regret guarantees straightforward, using aggregation experts. however, comparison class sufficiently rich, instance best expert set experts changes time, strategies naively require maintain prohibitive number weights (typically exponential time horizon). contrast, designing strategies achieve near-optimal regret maintain reasonable number weights highly non-trivial. consider three increasingly challenging objectives (simple regret, shifting regret sparse shifting regret) extend existing notions defined fixed expert ensemble; case, design strategies achieve tight regret bounds, adaptive parameters comparison class, computationally inexpensive. moreover, algorithms anytime, agnostic number incoming experts completely parameter-free. remarkable results made possible thanks two simple highly effective recipes: first ""abstention trick"" comes specialist framework enables handle least challenging notions regret, limited addressing sophisticated objectives. second, ""muting trick"" introduce give flexibility. show combine two tricks order handle challenging class comparison strategies.",18
"use deep learning interventional radiotherapy (brachytherapy): review focus open source open data. deep learning advanced one important technologies almost medical fields. especially areas, related medical imaging plays big role. however, interventional radiotherapy (brachytherapy) deep learning still early phase. review, first, investigated scrutinised role deep learning processes interventional radiotherapy directly related fields. additionally summarised recent developments. reproduce results deep learning algorithms source code training data must available. therefore, second focus work analysis availability open source, open data open models. analysis, able show deep learning plays already major role areas interventional radiotherapy, still hardly presented others. nevertheless, impact increasing years, partly self-propelled also influenced closely related fields. open source, data models growing number still scarce unevenly distributed among different research groups. reluctance publishing code, data models limits reproducibility restricts evaluation mono-institutional datasets. summarised, deep learning change positively workflow interventional radiotherapy room improvement comes reproducible results standardised evaluation methods.",14
"study clear sky models singapore. estimation total solar irradiance falling earth's surface important field solar energy generation forecasting. several clear-sky solar radiation models developed last decades. models based empirical distribution various geographical parameters; models consider various atmospheric effects solar energy estimation. paper, perform comparative analysis several popular clear-sky models, tropical region singapore. important countries like singapore, primarily focused reliable efficient solar energy generation. analyze compare three popular clear-sky models widely used literature. validate solar estimation results using actual solar irradiance measurements obtained collocated weather stations. finally conclude reliable clear sky model singapore, based clear sky days year.",14
"analytic solution model language competition bilingualism interlinguistic similarity. in-depth analytic study model language dynamics presented: model tackles problem coexistence two languages within closed community speakers taking account bilingualism incorporating parameter measure distance languages. previous numerical simulations, model yielded coexistence might lead survival languages within monolingual speakers along bilingual community extinction weakest tongue depending different parameters. paper, study closed thorough analytical calculations settle results robust way previous results refined modifications. present analysis possible almost completely assay number nature equilibrium points model, depend parameters, well build phase space based them. also, obtain conclusions way languages evolve time. rigorous considerations also suggest ways improve model facilitate comparison consequences approaches real data.",14
"analysis fast alternating minimization structured dictionary learning. methods exploiting sparsity popular imaging signal processing applications including compression, denoising, imaging inverse problems. data-driven approaches dictionary learning transform learning enable one discover complex image features datasets provide promising performance analytical models. alternating minimization algorithms particularly popular dictionary transform learning. work, study properties alternating minimization structured (unitary) sparsifying operator learning. algorithm converges stationary points non-convex problem general, prove rapid local linear convergence underlying generative model mild assumptions. experiments show unitary operator learning algorithm robust initialization.",3
"improving strong neural parser conjunction-specific features. dependency parsers reach high overall accuracy, dependency relations much harder others. particular, dependency parsers perform poorly coordination construction (i.e., correctly attaching ""conj"" relation). extend state-of-the-art dependency parser conjunction-specific features, focusing similarity conjuncts head words. training extended parser yields improvement ""conj"" attachment well overall dependency parsing accuracy stanford dependency conversion penn treebank.",3
"full wafer redistribution wafer embedding key technologies multi-scale neuromorphic hardware cluster. together kirchhoff-institute physics(kip) fraunhofer izm developed full wafer redistribution embedding technology base large-scale neuromorphic hardware system. paper give overview neuromorphic computing platform kip associated hardware requirements drove described technological developments. first phase project standard redistribution technologies wafer level packaging adapted enable high density reticle-to-reticle routing 200mm cmos wafers. neighboring reticles interconnected across scribe lines 8{\mu}m pitch routing based semi-additive copper metallization. passivation photo sensitive benzocyclobutene used enable second intra-reticle routing layer. final io pads flash gold generated top reticle. concept neuromorphic systems based full wafers could assembled tested. fabricated high density inter-reticle routing revealed high yield larger 99.9%. order allow upscaling system size large number wafers feasible effort full wafer embedding concept printed circuit boards developed proven second phase project. wafers thinned 250{\mu}m laminated additional prepreg layers copper foils core material. lamination pcb panel reticle ios embedded wafer accessed micro via drilling, copper electroplating, lithography subtractive etching pcb wiring structure. created wiring 50um line width enabled access reticle ios embedded wafer well board level routing. panels embedded wafers subsequently stressed 1000 thermal cycles 0c 100c shown severe failure formation cycle time.",3
"video analysis ""youtube funnies"" aid study human gait falls - preliminary results proof concept. falls funny, youtube video sharing sites contain large repository real-life falls. propose extracting gait balance information videos help us better understand factors contribute falls. proof-of-concept explored single video containing multiple (n=14) falls/non-falls presence unexpected obstacle. analysis explores: computing spatiotemporal parameters gait video captured arbitrary viewpoint; relationship parameters gait last steps obstacle falling vs. falling; predictive capacity multivariate model predicting fall presence unexpected obstacle. homography transformations correct perspective projection distortion allow consistent tracking gait parameters individual walks arbitrary direction scene. synthetic top view allows computing average stride length synthetic side view allows measuring motions head. leave-one-out cross-validation, able correctly predict whether person would fall 11 14 cases (78.6%), looking average stride length range vertical head motion 1-4 recent steps prior reaching obstacle.",3
"beam search strategies neural machine translation. basic concept neural machine translation (nmt) train large neural network maximizes translation performance given parallel corpus. nmt using simple left-to-right beam-search decoder generate new translations approximately maximize trained conditional probability. current beam search strategy generates target sentence word word left-to- right keeping fixed amount active candidates time step. first, simple search less adaptive also expands candidates whose scores much worse current best. secondly, expand hypotheses within best scoring candidates, even scores close best one. latter one avoided increasing beam size performance improvement observed. reach better performance, draw- back slower decoding speed. paper, concentrate speeding decoder applying flexible beam search strategy whose candidate size may vary time step depending candidate scores. speed original decoder 43% two language pairs german-english chinese-english without losing translation quality.",3
"distributed evolutionary computation: new technique solving large number equations. evolutionary computation techniques mostly used solve various optimization learning problems successfully. evolutionary algorithm effective gain optimal solution(s) solve complex problems traditional methods. case problems large set parameters, evolutionary computation technique incurs huge computational burden single processing unit. taking limitation account, paper presents new distributed evolutionary computation technique, decomposes decision vectors smaller components achieves optimal solution short time. technique, jacobi-based time variant adaptive (jbtva) hybrid evolutionary algorithm distributed incorporating cluster computation. moreover, two new selection methods named best selection (bas) twin selection (ts) introduced selecting best fit solution vector. experimental results show optimal solution achieved different kinds problems huge parameters considerable speedup obtained proposed distributed system.",3
"relax stay control: value algorithms online markov decision processes. online learning algorithms designed perform non-stationary environments, generally notion dynamic state model constraints current future actions function past actions. state-based models common stochastic control settings, commonly used frameworks markov decision processes (mdps) assume known stationary environment. recent years, growing interest combining two frameworks considering mdp setting cost function allowed change arbitrarily time step. however, work area algorithmic: given problem, one would develop algorithm almost scratch. moreover, presence state assumption arbitrarily varying environment complicate theoretical analysis development computationally efficient methods. paper describes broad extension ideas proposed rakhlin et al. give general framework deriving algorithms mdp setting arbitrarily changing costs. framework leads unifying view existing methods provides general procedure constructing new ones. several new methods presented, one shown important advantages similar method developed scratch via online version approximate dynamic programming.",3
"stochastic backward euler: implicit gradient descent algorithm $k$-means clustering. paper, propose implicit gradient descent algorithm classic $k$-means problem. implicit gradient step backward euler solved via stochastic fixed-point iteration, randomly sample mini-batch gradient every iteration. average fixed-point trajectory carried next gradient step. draw connections proposed stochastic backward euler recent entropy stochastic gradient descent (entropy-sgd) improving training deep neural networks. numerical experiments various synthetic real datasets show proposed algorithm finds global minimum (or neighborhood) high probability, given correct number clusters. method provides better clustering results compared $k$-means algorithms sense decreased objective function (the cluster) much robust initialization.",10
"improved calibration wavelength dependence metallicity cepheid leavitt law. cepheid period-luminosity (pl) relation (or leavitt law) served first rung widely used extragalactic distance ladders central determination local value hubble constant ($h_0$). investigate influence metallicity cepheid brightness, term significantly improves overall fit distance ladder, better define wavelength dependence. aim, compare pl relations obtained three cepheid samples distinct chemical composition (in milky way magellanic clouds) focusing use improved recent data covering metallicity range 1 dex. estimate metallicity effect (hereafter $\gamma$) 15 filters mid-infrared optical wavelengths, including 5 wesenheit indices, derive significant metallicity term filters, accord recent empirical studies models, sense metal-rich cepheids brighter metal-poor ones. describe contribution various systematic effects determination $\gamma$ term. find evidence $\gamma$ changing wavelength range 0.5 - 4.5 $\mu$m indicating main influence metallicity cepheids luminosity rather color. identify factors sharpen empirical constraints metallicity term past studies including corrections depth magellanic clouds, better calibrated cepheid photometry, improved milky way extinction estimates revised expanded metallicity measurements lmc.",0
"computational narrative intelligence: human-centered goal artificial intelligence. narrative intelligence ability craft, tell, understand, respond affectively stories. argue instilling artificial intelligences computational narrative intelligence affords number applications beneficial humans. lay machine learning challenges necessary solve achieve computational narrative intelligence. finally, argue computational narrative practical step towards machine enculturation, teaching sociocultural values machines.",3
"siamese deep forest. siamese deep forest (sdf) proposed paper. based deep forest gcforest proposed zhou feng viewed gcforest modification. also regarded alternative well-known siamese neural networks. sdf uses modified training set consisting concatenated pairs vectors. moreover, defines class distributions deep forest weighted sum tree class probabilities weights determined order reduce distances similar pairs increase dissimilar points. show weights obtained solving quadratic optimization problem. sdf aims prevent overfitting takes place neural networks limited training data available. numerical experiments illustrate proposed distance metric method.",18
"efficient quantum algorithm generative machine learning. central task field quantum computing find applications quantum computer could provide exponential speedup classical computer. machine learning represents important field broad applications quantum computer may offer significant speedup. several quantum algorithms discriminative machine learning found based efficient solving linear algebraic problems, potential exponential speedup runtime assumption effective input quantum random access memory. machine learning, generative models represent another large class widely used supervised unsupervised learning. here, propose efficient quantum algorithm machine learning based quantum generative model. prove proposed model exponentially powerful represent probability distributions compared classical generative models exponential speedup training inference least instances reasonable assumption computational complexity theory. result opens new direction quantum machine learning offers remarkable example quantum algorithm shows exponential improvement classical algorithm important application field.",17
"reinforcement learning via aixi approximation. paper introduces principled approach design scalable general reinforcement learning agent. approach based direct approximation aixi, bayesian optimality notion general reinforcement learning agents. previously, unclear whether theory aixi could motivate design practical algorithms. answer hitherto open question affirmative, providing first computationally feasible approximation aixi agent. develop approximation, introduce monte carlo tree search algorithm along agent-specific extension context tree weighting algorithm. empirically, present set encouraging results number stochastic, unknown, partially observable domains.",3
"integrating boundary center correlation filters visual tracking aspect ratio variation. aspect ratio variation frequently appears visual tracking severe influence performance. although many correlation filter (cf)-based trackers also suggested scale adaptive tracking, studies given handle aspect ratio variation cf trackers. paper, make first attempt address issue introducing family 1d boundary cfs localize left, right, top, bottom boundaries videos. allows us cope aspect ratio variation flexibly tracking. specifically, present novel tracking model integrate 1d boundary 2d center cfs (ibccf) boundary center filters enforced near-orthogonality regularization term. optimize ibccf model, develop alternating direction method multipliers. experiments several datasets show ibccf effectively handle aspect ratio variation, achieves state-of-the-art performance terms accuracy robustness.",3
"gadaboost: accelerating adaboost feature selection genetic algorithms. boosted cascade simple features, viola jones, one famous object detection frameworks. however, suffers lengthy training process. due vast features space exhaustive search nature adaboost. paper propose gadaboost: genetic algorithm accelerate training procedure natural feature selection. specifically, propose limit adaboost search within subset huge feature space, evolving subset following genetic algorithm. experiments demonstrate proposed gadaboost 3.7 times faster adaboost. also demonstrate price speedup mere decrease (3%, 4%) detection accuracy tested fddb benchmark face detection set, caltech web faces respectively.",3
"low-shot learning large-scale diffusion. paper considers problem inferring image labels labelled examples available training time. setup often referred low-shot learning literature, standard approach re-train last layers convolutional neural network learned separate classes. consider semi-supervised setting exploit large collection images support label propagation. made possible leveraging recent advances large-scale similarity graph construction. show despite conceptual simplicity, scaling label propagation hundred millions images leads state art accuracy low-shot learning regime.",3
"cluster-based kriging approximation algorithms complexity reduction. kriging gaussian process regression applied many fields non-linear regression model well surrogate model field evolutionary computation. however, computational space complexity kriging, cubic quadratic number data points respectively, becomes major bottleneck data available nowadays. paper, propose general methodology complexity reduction, called cluster kriging, whole data set partitioned smaller clusters multiple kriging models built top them. addition, four kriging approximation algorithms proposed candidate algorithms within new framework. algorithms applied much larger data sets maintaining advantages power kriging. proposed algorithms explained detail compared empirically broad set existing state-of-the-art kriging approximation methods well-defined testing framework. according empirical study, proposed algorithms consistently outperform existing algorithms. moreover, practical suggestions provided using proposed algorithms.",3
"regulargcc matrix constraint. study propagation regulargcc global constraint. ensures row matrix decision variables satisfies regular constraint, column satisfies gcc constraint. negative side, prove propagation np-hard even strong restrictions (e.g. 3 values, 4 states automaton, 5 columns matrix). positive side, identify two cases propagation fixed parameter tractable. addition, show improve propagation simple decomposition separate regular gcc constraints identifying necessary insufficient conditions solution. enforce conditions additional weighted row automata. experimental results demonstrate potential methods standard benchmark problems.",3
"distributional convergence sliced wasserstein process. motivated statistical computational challenges computing wasserstein distances high-dimensional contexts, machine learning researchers defined modified wasserstein distances based computing distances one-dimensional projections measures. different choices aggregate projected distances (averaging, random sampling, maximizing) give rise different distances, requiring different statistical analyses. define \emph{sliced wasserstein process}, stochastic process defined empirical wasserstein distance projections empirical probability measures one-dimensional subspaces, prove uniform distributional limit theorem process. result, obtain unified framework prove distributional limit results wasserstein distances based one-dimensional projections. illustrate results number examples distributional limits previously known.",10
"collective behavior stock prices time crisis response external stimulus. analyze interaction stock prices big companies usa germany using granger causality. claim increase pair-wise granger causality interaction prices times crisis consequence simultaneous response markets outside events external stimulus considered common driver stocks, result real causal predictability prices themselves. alternative approach recurrence analysis single stock price series supports claim. observed patterns price stocks modelled adding multiplicative exogenous term representative external factors geometric brownian motion model stock prices. altogether, detect model effects great recession consequence mortgage crisis 2007/2008 well impacts covid out-break early 2020",16
"assessing performance bounds local feature detectors: taking inspiration electronics design practices. since local feature detection one active research areas computer vision, large number detectors proposed. rendered task characterizing performance various feature detection methods important issue vision research. inspired good practices electronic system design, generic framework based improved repeatability measure presented paper allows assessment upper lower bounds detector performance effort design reliable effective vision systems. framework employed establish operating guarantee regions several state-of-the art detectors jpeg compression uniform light changes. results obtained using newly acquired, large image database (15092 images) 539 different scenes. results provide new insights behavior detectors also useful vision systems design perspective.",3
"accelerating reinforcement learning implicit imitation. imitation viewed means enhancing learning multiagent environments. augments agent's ability learn useful behaviors making intelligent use knowledge implicit behaviors demonstrated cooperative teachers experienced agents. propose study formal model implicit imitation accelerate reinforcement learning dramatically certain cases. roughly, observing mentor, reinforcement-learning agent extract information capabilities in, relative value of, unvisited parts state space. study two specific instantiations model, one learning agent mentor identical abilities, one designed deal agents mentors different action sets. illustrate benefits implicit imitation integrating prioritized sweeping, demonstrating improved performance convergence observation single multiple mentors. though make stringent assumptions regarding observability possible interactions, briefly comment extensions model relax restricitions.",3
"minimizing misclassification error rate using surrogate convex loss. carefully study well minimizing convex surrogate loss functions, corresponds minimizing misclassification error rate problem binary classification linear predictors. particular, show amongst convex surrogate losses, hinge loss gives essentially best possible bound, convex loss functions, misclassification error rate resulting linear predictor terms best possible margin error rate. also provide lower bounds specific convex surrogates show different commonly used losses qualitatively differ other.",3
"determinants mobile money adoption pakistan. work, analyze problem adoption mobile money pakistan using call detail records major telecom company input. results highlight fact different sections society different patterns adoption digital financial services user mobility related features important one comes adopting using mobile money services.",18
solving multi-class using orthogonal coding matrices. probability estimates desirable statistical classification gauging accuracy classification result calibration. describe method solving conditional probabilities multi-class classification using orthogonal error correcting codes. method tested six different datasets using support vector machines compares favorably existing technique based one-versus-one multi-class method. probabilities validated based cumulative sum boolean evaluation correctness class label divided estimated probability. probability estimation using orthogonal coding simple efficient potential faster classification results one-versus-one method.,18
"mining discovering biographical information difangzhi language-model-based approach. present results expanding contents china biographical database text mining historical local gazetteers, difangzhi. goal database see people connected together, kinship, social connections, places offices served. gazetteers single important collection names offices covering song qing periods. although begin local officials shall eventually include lists local examination candidates, people locality served government, notable local figures biographies. data collect connections emerge. value systematic text mining work identify relevant connections either directly informative become useful without deep historical research. academia sinica developing name database officials central governments ming qing dynasties.",3
"pointwise space-time estimates 3d bipolar compressible navier-stokes-poisson system unequal viscosities. space-time behaviors 3d bipolar navier-stokes-poisson system (bnsp) unequal viscosities unequal pressure functions given. know, space-time estimate electric field \nabla\phi important one deducing generalized huygens' principle bnsp since estimate obtained relation \nabla\phi=\frac{\nabla}{\delta}(\rho-n) poisson equation. thus, requires prove space-time estimate \rho-n contains diffusion wave. appearance unequal coefficients results one cannot follow ideas special case \cite{wu4}, original system rewritten navier-stokes system (ns) unipolar navier-stokes-poisson system (nsp) suitable linear combination unknowns. additionally linear combination brings special structure nonlinear terms, special structure also used get desired space-time estimate \rho-n. importantly, green's function subsystem nsp contain huygens' wave equally important proof \cite{wu4}. however, general case here, benefits linear combination exist longer. first, directly consider 8\times8 green's matrix. second, entries green's function low frequency actually contain wave operators. generally produces huygens' wave entries green's function, result, one cannot achieve space-time estimate $z\rho-n$ contains diffusion wave usual. overcome difficulty making detailed spectral analysis developing new estimates arising cancellations green's function. third, due loss special structure nonlinear terms linear combination, shall develop new nonlinear convolution estimates.",10
"mining smart card data travelers' mini activities. context public transport modeling simulation, address problem mismatch simulated transit trips observed ones. point weakness current travel demand modeling process; trips generates over-optimistic reflect real passenger choices. introduce notion mini activities travelers trips; explain deviation simulated trips observed trips. propose mine smart card data extract mini activities. develop technique integrate generated trips learn integration two available sources, trip history trip planner recommendations. input travel demand, build markov chain trip collection apply monte carlo markov chain algorithm integrate mini activities way selected characteristics converge desired distributions. test method different settings passenger trip collection nancy, france. report experimental results demonstrating important mismatch reduction.",3
"on-line learning unlearnable true teacher mobile ensemble teachers. on-line learning hierarchical learning model studied method statistical mechanics. model student simple perceptron learns true teacher directly, ensemble teachers learn true teacher perceptron learning rule. since true teacher ensemble teachers expressed non-monotonic perceptron simple ones, respectively, ensemble teachers go around unlearnable true teacher distance fixed asymptotic steady state. generalization performance student shown exceed ensemble teachers transient state, shown similar ensemble-teachers models. further, found moving ensemble teachers even steady state, contrast fixed ensemble teachers, efficient performance student.",2
"convolutional kernel networks. important goal visual recognition devise image representations invariant particular transformations. paper, address goal new type convolutional neural network (cnn) whose invariance encoded reproducing kernel. unlike traditional approaches neural networks learned either represent data solving classification task, network learns approximate kernel feature map training data. approach enjoys several benefits classical ones. first, teaching cnns invariant, obtain simple network architectures achieve similar accuracy complex ones, easy train robust overfitting. second, bridge gap neural network literature kernels, natural tools model invariance. evaluate methodology visual recognition tasks cnns proven perform well, e.g., digit recognition mnist dataset, challenging cifar-10 stl-10 datasets, accuracy competitive state art.",3
multi-optional many-sorted past present future structures description. cognitive theory true conditions (cttc) proposal describe model-theoretic semantics symbolic cognitive architectures design implementation cognitive abilities. cttc formulated mathematically using multi-optional many-sorted past present future(mmppf) structures. article defines mathematically mmppf structures formal languages proposed describe cttc.,3
"support consistency direct sparse-change learning markov networks. study problem learning sparse structure changes two markov networks $p$ $q$. rather fitting two markov networks separately two sets data figuring differences, recent work proposed learn changes \emph{directly} via estimating ratio two markov network models. paper, give sufficient conditions \emph{successful change detection} respect sample size $n_p, n_q$, dimension data $m$, number changed edges $d$. using unbounded density ratio model prove true sparse changes consistently identified $n_p = \omega(d^2 \log \frac{m^2+m}{2})$ $n_q = \omega({n_p^2})$, exponentially decaying upper-bound learning error. sample complexity improved $\min(n_p, n_q) = \omega(d^2 \log \frac{m^2+m}{2})$ boundedness density ratio model assumed. theoretical guarantee applied wide range discrete/continuous markov networks.",18
"really use post-hoc tests based mean-ranks?. statistical comparison multiple algorithms multiple data sets fundamental machine learning. typically carried friedman test. friedman test rejects null hypothesis, multiple comparisons carried establish significant differences among algorithms. multiple comparisons usually performed using mean-ranks test. aim technical note discuss inconsistencies mean-ranks post-hoc test goal discouraging use machine learning well medicine, psychology, etc.. show outcome mean-ranks test depends pool algorithms originally included experiment. words, outcome comparison algorithms b depends also performance algorithms included original experiment. lead paradoxical situations. instance difference b could declared significant pool comprises algorithms c, d, e significant pool comprises algorithms f, g, h. overcome issues, suggest instead perform multiple comparison using test whose outcome depends two algorithms compared, sign-test wilcoxon signed-rank test.",3
"atp presentation service mizar formalizations. paper describes automated reasoning mizar (mizar) service, integrates several automated reasoning, artificial intelligence, presentation tools mizar authoring environment. service provides atp assistance mizar authors finding explaining proofs, offers generation mizar problems challenges atp systems. service based sound translation mizar language first-order atp systems, relies recent progress application atp systems large theories containing tens thousands available facts. present main features mizar services, followed account initial experiments finding proofs atp assistance. initial experience indicates tool offers substantial help exploring mizar library preparing new mizar articles.",3
"pushbroom stereo high-speed navigation cluttered environments. present novel stereo vision algorithm capable obstacle detection mobile-cpu processor 120 frames per second. system performs subset standard block-matching stereo processing, searching obstacles single depth. using onboard imu state-estimator, recover position obstacles depths, building updating full depth-map framerate. here, describe algorithm implementation high-speed, small uav, flying 20 mph (9 m/s) close obstacles. system requires external sensing computation is, best knowledge, first high-framerate stereo detection system running onboard small uav.",3
"counterfactual language model adaptation suggesting phrases. mobile devices use language models suggest words phrases use text entry. traditional language models based contextual word frequency static corpus text. however, certain types phrases, offered writers suggestions, may systematically chosen often frequency would predict. paper, propose task generating suggestions writers accept, related distinct task making accurate predictions. although task fundamentally interactive, propose counterfactual setting permits offline training evaluation. find even simple language model capture text characteristics improve acceptability.",3
"generalized inverse classification. inverse classification process perturbing instance meaningful way likely conform specific class. historical methods address problem often framed leverage single classifier, specific set classifiers. works often accompanied naive assumptions. work propose generalized inverse classification (gic), avoids restricting classification model used. incorporate formulation refined framework gic takes place. framework, gic operates features immediately actionable. change incurs individual cost, either linear non-linear. changes subjected occur within specified level cumulative change (budget). furthermore, framework incorporates estimation features change consequence direct actions taken (indirectly changeable features). solve problem, propose three real-valued heuristic-based methods two sensitivity analysis-based comparison methods, evaluated two freely available real-world datasets. results demonstrate validity benefits formulation, framework, methods.",3
"optimization beyond prediction: prescriptive price optimization. paper addresses novel data science problem, prescriptive price optimization, derives optimal price strategy maximize future profit/revenue basis massive predictive formulas produced machine learning. prescriptive price optimization first builds sales forecast formulas multiple products, basis historical data, reveal complex relationships sales prices, price elasticity demand cannibalization. then, constructs mathematical optimization problem basis predictive formulas. present optimization problem formulated instance binary quadratic programming (bqp). although bqp problems np-hard general computationally intractable, propose fast approximation algorithm using semi-definite programming (sdp) relaxation, closely related goemans-williamson's max-cut approximation. experiments simulation real retail datasets show prescriptive price optimization simultaneously derives optimal prices tens/hundreds products practical computational time, potentially improve 8.2% gross profit products.",10
"qualitative models decision uncertainty without commensurability assumption. paper investigates purely qualitative version savage's theory decision making uncertainty. now, representation theorems preference acts rely numerical representation utility uncertainty utility uncertainty commensurate. disrupting tradition, relax assumption introduce purely ordinal axiom requiring decision maker (dm) preference two acts depends relative position consequences state. within qualitative framework, determine possible form decision rule investigate instances compatible transitivity strict preference. finally propose mild relaxation ordinality axiom, leaving room new family qualitative decision rules compatible transitivity.",3
"large scale novel object discovery 3d. present method discovering never-seen-before objects 3d point clouds obtained sensors like microsoft kinect. generate supervoxels directly point cloud data use siamese network, built recently proposed 3d convolutional neural network architecture. use known objects train non-linear embedding supervoxels, optimizing criteria supervoxels fall object closer fall different objects, embedding space. test unknown objects, seen training, perform clustering learned embedding space supervoxels effectively perform novel object discovery. validate method extensive experiments, quantitatively showing discover numerous unseen objects trained dense 3d models. also show good qualitative results object discovery point cloud data test objects, either specific instances even categories, never seen training.",3
"learning select pre-trained deep representations bayesian evidence framework. propose bayesian evidence framework facilitate transfer learning pre-trained deep convolutional neural networks (cnns). framework formulated top least squares svm (ls-svm) classifier, simple fast training testing, achieves competitive performance practice. regularization parameters ls-svm estimated automatically without grid search cross-validation maximizing evidence, useful measure select best performing cnn multiple candidates transfer learning; evidence optimized efficiently employing aitken's delta-squared process, accelerates convergence fixed point update. proposed bayesian evidence framework also provides good solution identify best ensemble heterogeneous cnns greedy algorithm. bayesian evidence framework transfer learning tested 12 visual recognition datasets illustrates state-of-the-art performance consistently terms prediction accuracy modeling efficiency.",3
"swift heuristic method work order scheduling skilled-workforce constraint. considered problem optimally allocate set jobs technicians different skills number technicians skill exceed number persons skill designation. key motivation quick sensitivity analysis terms workforce size quite necessary many industries presence unexpected work orders. time-indexed mathematical model proposed minimize total weighted completion time jobs. proposed model decomposed number single-skill sub-problems one combination series nested binary knapsack problems. heuristic procedure proposed solve problem. experimental results, based real-world case study, reveal proposed method quickly produces schedule statistically close optimal one classical optimal procedure time-consuming.",3
affine-invariant diffusion geometry analysis deformable 3d shapes. introduce (equi-)affine invariant diffusion geometry surfaces go squeeze shear transformations still properly analyzed. definition affine invariant metric enables us construct invariant laplacian local global geometric structures extracted. applications proposed framework demonstrate power generalizing enriching existing set tools shape analysis.,3
"spectroscopic binaries high mass functions near gaia dr3 main sequence?. 3rd data release gaia mission includes orbital solutions $> 10^5$ single-lined spectroscopic binaries, representing order magnitude increase sample size previous studies. dataset treasure trove searches quiescent black hole + normal star binaries. investigate one population black hole candidate binaries highlighted data release: sources near main sequence color-magnitude diagram (cmd) dynamically-inferred companion masses $m_2$ larger cmd-inferred mass luminous star. model light curves, spectral energy distributions, archival spectra 14 objects dr3 high-significance orbital solutions inferred $m_2 > 3\,m_{\odot}$. find 100\% sources mass-transfer binaries containing highly stripped lower giant donor ($0.2 \lesssim m/m_{\odot} \lesssim 0.4$) much massive ($2 \lesssim m/m_{\odot} \lesssim 2.5$) main-sequence accretor. gaia orbital solutions donors, contribute half light gaia rvs bandpass $\lesssim 20\%$ $g-$band. accretors' broad spectral features likely prevented sources classified double-lined. donors close roche lobe-filling ($r/r_{\rm roche\,lobe}>0.8$), modeling suggests majority detached ($r/r_{\rm roche\,lobe}<1$). binary evolution models predict systems soon become detached helium white dwarf + main sequence ""el cvn"" binaries. investigation highlights power gaia data selecting interesting sub-populations binaries ways binary evolution bamboozle standard cmd-based stellar mass estimates.",0
"temporal-needle: view appearance invariant video descriptor. ability detect similar actions across videos useful real-world applications many fields. however, task still challenging existing systems, since videos present action, taken significantly different viewing directions, performed different actors backgrounds various video qualities. video descriptors play significant role systems. work propose ""temporal-needle"" descriptor captures dynamic behavior, invariant viewpoint appearance. descriptor computed using multi temporal scales video computing self-similarity every patch time every temporal scale. descriptor computed every pixel video. however, find similar actions across videos, consider small subset descriptors - statistical significant descriptors. allow us find good correspondences across videos efficiently. using descriptor, able detect behavior across videos variety scenarios. demonstrate use descriptor tasks temporal spatial alignment, action detection even show potential unsupervised video clustering categories. work handled videos taken stationary cameras, descriptor extended handle moving camera well.",3
"event-camera dataset simulator: event-based data pose estimation, visual odometry, slam. new vision sensors, dynamic active-pixel vision sensor (davis), incorporate conventional global-shutter camera event-based sensor pixel array. sensors great potential high-speed robotics computer vision allow us combine benefits conventional cameras event-based sensors: low latency, high temporal resolution, high dynamic range. however, new algorithms required exploit sensor characteristics cope unconventional output, consists stream asynchronous brightness changes (called ""events"") synchronous grayscale frames. purpose, present release collection datasets captured davis variety synthetic real environments, hope motivate research new algorithms high-speed high-dynamic-range robotics computer-vision applications. addition global-shutter intensity images asynchronous events, provide inertial measurements ground-truth camera poses motion-capture system. latter allows comparing pose accuracy ego-motion estimation algorithms quantitatively. data released standard text files binary files (i.e., rosbag). paper provides overview available data describes simulator release open-source create synthetic event-camera data.",3
"zero shot recognition unreliable attributes. principle, zero-shot learning makes possible train recognition model simply specifying category's attributes. example, classifiers generic attributes like \emph{striped} \emph{four-legged}, one construct classifier zebra category enumerating properties possesses---even without providing zebra training images. practice, however, standard zero-shot paradigm suffers attribute predictions novel images hard get right. propose novel random forest approach train zero-shot models explicitly accounts unreliability attribute predictions. leveraging statistics attribute's error tendencies, method obtains robust discriminative models unseen classes. devise extensions handle few-shot scenario unreliable attribute descriptions. three datasets, demonstrate benefit visual category learning zero training examples, critical domain rare categories categories defined fly.",3
"deep driven fmri decoding visual categories. deep neural networks developed drawing inspiration brain visual pathway, implementing end-to-end approach: image data video object classes. however building fmri decoder typical structure convolutional neural network (cnn), i.e. learning multiple level representations, seems impractical due lack brain data. possible solution, work presents first hybrid fmri deep features decoding approach: collected fmri deep learnt representations video object classes linked together means kernel canonical correlation analysis. decoding, allows exploiting discriminatory power cnn relating fmri representation last layer cnn (fc7). show effectiveness embedding fmri data onto subspace related deep features distinguishing semantic visual categories based solely brain imaging data.",18
"new da white dwarf models asteroseismology zz ceti stars. asteroseismology powerful tool infer evolutionary status chemical stratification white dwarf (wd) stars, explore physical processes lead formation. particularly true variable h-rich atmosphere (da) wds, known dav zz ceti stars. present new grid da wd models take account last advances modeling input physics progenitor wd stars, thus avoiding improving several shortcomings present set da wd models employed asteroseismological analyses zz ceti stars carried previous works. new models derived self-consistent way changes internal chemical distribution result mixing core chemical components induced mean molecular-weight inversions, $^{22}$ne diffusion, coulomb sedimentation, residual nuclearburning. addition, expected nuclear-burning history mixing events along progenitor evolution accounted for, particular occurrence third dredge-up, determines properties core envelope post-agb white dwarf stars, well wd initial-final mass relation. range h envelopes new zz star models extends log(m_h/m_*)= -4 -5, log(m_h/m_*)= -13.5. allows, first time, consider seismological solutions zz ceti stars extremely thin h envelopes. new h-burning post-agb models predict chemical profiles c near stellar centre zz ceti stars substantially different used previous works. find pulsation periods $g$ modes mode-trapping properties new models differ significantly characterizing zz ceti models previous works, particularly long periods.",0
"optimizing cvar via sampling. conditional value risk (cvar) prominent risk measure used extensively various domains. develop new formula gradient cvar form conditional expectation. based formula, propose novel sampling-based estimator cvar gradient, spirit likelihood-ratio method. analyze bias estimator, prove convergence corresponding stochastic gradient descent algorithm local cvar optimum. method allows consider cvar optimization new domains. example, consider reinforcement learning application, learn risk-sensitive controller game tetris.",18
"empirical measure large deviations reinforced chains finite spaces. let $a$ transition probability kernel finite state space $\delta^o =\{1, \ldots , d\}$ $a(x,y)>0$ $x,y \in \delta^o$. consider reinforced chain given sequence $\{x_n, \; n \in \mathbb{n}_0\}$ $\delta^o$-valued random variables, defined recursively according to, $$l^n = \frac{1}{n}\sum_{i=0}^{n-1} \delta_{x_i}, \;\; p(x_{n+1} \in \cdot \mid x_0, \ldots, x_n) = l^n a(\cdot).$$ establish large deviation principle $\{l^n\}$. rate function takes strikingly different form donsker-varadhan rate function associated empirical measure markov chain transition kernel $a$ described terms novel deterministic infinite horizon discounted cost control problem associated linear controlled dynamics nonlinear running cost involving relative entropy function. proofs based analysis time-reversal controlled dynamics representations log-transforms exponential moments, weak convergence methods.",10
"efficient effective single-document summarizations word-embedding measurement quality. task generate effective summary given document specific realtime requirements. use softplus function enhance keyword rankings favor important sentences, based present number summarization algorithms using various keyword extraction topic clustering methods. show algorithms meet realtime requirements yield best rouge recall scores duc-02 previously-known algorithms. show algorithms meet realtime requirements yield best rouge recall scores duc-02 previously-known algorithms. evaluate quality summaries without human-generated benchmarks, define measure called wesm based word-embedding using word mover's distance. show orderings rouge wesm scores algorithms highly comparable, suggesting wesm may serve viable alternative measuring quality summary.",3
"saliency fusion eigenvector space multi-channel pulse coupled neural network. saliency computation become popular research field many applications due useful information provided saliency maps. saliency map, local relations around salient regions multi-channel perspective taken consideration aiming uniformity region interest internal approach. and, irrelevant salient regions avoided much possible. works achieve criteria external processing modules; however, accomplished conspicuity map fusion process. therefore, paper, new model proposed saliency/conspicuity map fusion two concepts: a) input image transformation relying principal component analysis (pca), b) saliency conspicuity map fusion multi-channel pulsed coupled neural network (m-pcnn). experimental results, evaluated precision, recall, f-measure, area curve (auc), support reliability proposed method enhancing saliency computation.",3
"provable inductive matrix completion. consider movie recommendation system apart ratings information, side information user's age movie's genre also available. unlike standard matrix completion, setting one able predict inductively new users/movies. paper, study problem inductive matrix completion exact recovery setting. is, assume ratings matrix generated applying feature vectors low-rank matrix goal recover back underlying matrix. furthermore, generalize problem low-rank matrix estimation using rank-1 measurements. study generic problem provide conditions set measurements satisfy alternating minimization method (which otherwise non-convex method convergence guarantees) able recover back {\em exact} underlying low-rank matrix. addition inductive matrix completion, show two low-rank estimation problems studied framework: a) general low-rank matrix sensing using rank-1 measurements, b) multi-label regression missing labels. problems, provide novel interesting bounds number measurements required alternating minimization provably converges {\em exact} low-rank matrix. particular, analysis general low rank matrix sensing problem significantly improves required storage computational cost required rip-based matrix sensing methods \cite{rechtfp2007}. finally, provide empirical validation approach demonstrate alternating minimization able recover true matrix mentioned problems using small number measurements.",3
"train test tightness lp relaxations structured prediction. structured prediction used areas computer vision natural language processing predict structured outputs segmentations parse trees. settings, prediction performed map inference or, equivalently, solving integer linear program. complex scoring functions required obtain accurate predictions, learning inference typically require use approximate solvers. propose theoretical explanation striking observation approximations based linear programming (lp) relaxations often tight real-world instances. particular, show learning lp relaxed inference encourages integrality training instances, tightness generalizes train test data.",18
"efficient tabling mechanisms transaction logic programs. paper present efficient evaluation algorithms horn transaction logic (a generalization regular horn logic programs state updates). present two complementary methods optimizing implementation transaction logic. first method based tabling modified proof theory table calls answers states (practically, equivalent dynamic programming). call-answer table indexed call signature state call made. answer columns contain answer unification signature state call executed. states signed efficiently using technique based tries counting. second method based incremental evaluation applies data oracle contains derived relations. deletions insertions (executed transaction oracle) change state database. using heuristic inertia (only part state changes response elementary updates), time cheaper compute changes state recompute entire state scratch. two methods complementary fact first method optimizes evaluation call repeated state, second method optimizes evaluation new state call-state pair found tabling mechanism (i.e. first method). proof theory transaction logic application tabling incremental evaluation sound complete respect model theory.",3
"fast memory-efficient significant pattern mining via permutation testing. present novel algorithm, westfall-young light, detecting patterns, itemsets subgraphs, statistically significantly enriched one two classes. method corrects rigorously multiple hypothesis testing correlations patterns westfall-young permutation procedure, empirically estimates null distribution pattern frequencies class via permutations. experiments, westfall-young light dramatically outperforms current state-of-the-art approach terms runtime memory efficiency popular real-world benchmark datasets pattern mining. key efficiency unlike existing methods, algorithm neither needs solve underlying frequent itemset mining problem anew permutation needs store occurrence list frequent patterns. westfall-young light opens door significant pattern mining large datasets previously led prohibitive runtime memory costs.",18
"linear programming driven genetic algorithm meta-scheduling utility grids. user-level brokers grids consider individual application qos requirements minimize cost without considering demands users. results contention resources sub-optimal schedules. meta-scheduling grids aims address scheduling problem, np hard due combinatorial nature. thus, many heuristic-based solutions using genetic algorithm (ga) proposed, apart traditional algorithms greedy fcfs. propose linear programming/integer programming model (lp/ip) scheduling applications multiple resources. also propose novel algorithm lpga (linear programming driven genetic algorithm) combines capabilities lp ga. aim algorithm obtain best metaschedule utility grids minimize combined cost users coordinated manner. simulation results show proposed integrated algorithm offers best schedule minimum processing cost negligible time overhead.",3
"self-reflective risk-aware artificial cognitive modeling robot response human behaviors. order cooperative robots (""co-robots"") respond human behaviors accurately efficiently human-robot collaboration, interpretation human actions, awareness new situations, appropriate decision making crucial abilities co-robots. purpose, human behaviors interpreted co-robots manner human peers. address issue, novel interpretability indicator introduced robot actions appropriate current human behaviors. addition, complete consideration potential situations robot's environment nearly impossible real-world applications, making difficult co-robot act appropriately safely new scenarios. true even pretrained model highly accurate known situation. effective safe teaming humans, introduce new generalizability indicator allows co-robot self-reflect reason observation falls outside co-robot's learned model. based topic modeling two novel indicators, propose new self-reflective risk-aware artificial cognitive (srac) model. co-robots able consider action risks identify new situations better decisions made. experiments using real-world datasets physical robots suggest srac model significantly outperforms traditional methodology enables better decision making response human activities.",3
"lepton flavour violation identification tau decay ($τ^{-} \rightarrow μ^{-}μ^{-}μ^{+}$) using artificial intelligence. discovery neutrino oscillation, proving neutrinos masses, reveals misfits particles current standard model (sm) theory. theory, neutrinos masses could result lepton flavour symmetry called lepton flavour violation (lfv). sm theory extensions allowed lfv processes, branching fractions small, making unobservable even strongest equipment up-to-date. that, scientists recent years generated lfv-like processes combined lhcb monte-carlo-simulated data attempt identify lfv using artificial intelligence (ai), specifically machine learning (ml) deep learning (dl). paper, performance several algorithms ai presented, xgboost, lightgbm, custom 1-d dense block neural networks (dbnns), custom 1-d convolutional neural networks (cnns) identifying lfv signals, specifically $\tau^{-} \rightarrow \mu^{-}\mu^{-}\mu^{+}$ decay combined lhcb monte-carlo-simulated data imitates signatures said decay. kolmogorov-smirnov (ks) cramer-von mises (cvm) tests also conducted verify validity predictions trained algorithms. result shows decent performances among algorithms, except lightgbm, failing cvm test, 20-layered cnn recorded considerably low auc. meanwhile, xgboost 10-layered dbnn recorded highest auc 0.88. main contribution paper extensive experiment involving custom dbnn cnn algorithms different layers, rarely used past years identifying lfv-like signatures, unlike gbms tree-based algorithms, popular said task.",8
"self-training method semi-supervised gans. since creation generative adversarial networks (gans), much work done improve training stability, generated image quality, range application nearly none explored self-training potential. self-training used advent deep learning order allow training limited labelled training data shown impressive results semi-supervised learning. work, combine two ideas make gans self-trainable semi-supervised learning tasks exploiting infinite data generation potential. results show using even simplest form self-training yields improvement. also show results complex self-training scheme performs least well basic self-training scheme significantly less data augmentation.",3
"localizing branes bifurcating bulks. study problem evolution bulk 5-fluids embedded braneworld flat, de sitter, anti-de sitter geometry. introduce new variables express einstein equations dynamical system depends equation state parameter $\gamma$ exponent $\lambda$. linear fluids (i.e., $\lambda=1$), formulation leads partial decoupling equations thus exact solution. find fluid develops transcritical bifurcation around value $\gamma=-1/2$, study behaviour affects stability solutions. nonlinear fluids, situation diverse. find overall attractor $\lambda=1/2$ draw enough phase portraits exhibit detail overall dynamics. show value $\lambda =3/2$ structurally unstable typical forms $\lambda$. consequently, observe noticeable dependence qualitative behaviour solutions different `polytropic' forms fluid bulk. addition, prove existence dulac function nonlinear fluids, signifying impossibility closed orbits certain subsets phase space. also provide ample numerical evidence gravity localizing solutions brane satisfy energy conditions.",9
"client profiling anti-money laundering system. present data mining approach profiling bank clients order support process detection anti-money laundering operations. first present overall system architecture, focus relevant component paper. detail experiments performed real world data financial institution, allowed us group clients clusters generate set classification rules. discuss relevance founded client profiles generated classification rules. according defined overall agent-based architecture, rules incorporated knowledge base intelligent agents responsible signaling suspicious transactions.",3
linear isoperimetric functions surfaces hyperbolic groups. show word-hyperbolic groups satisfy linear isoperimetric functions homotopy types surface diagrams. generalises linear isoperimetric functions disc annular diagrams.,10
"human group behaviour simulation evaluation framework utilising composition video analysis. work present modular crowd simulation evaluation composition framework (csec) provides quantitative comparison different pedestrian crowd simulation approaches. evaluation made based comparison source footage synthetic video created novel composition techniques. proposed framework seeks reduce complexity simulation evaluation provide platform comparison differing simulation algorithms well parametric tuning conducted improve simulation accuracy providing measures similarity crowd simulation algorithms source data. use features designed mimic human visual system (hvs), specific simulation properties evaluated relative sample footage. validation performed number popular crowd datasets comparisons multiple pedestrian crowd simulation algorithms.",3
"sparse linear regression missing data. paper proposes fast accurate method sparse regression presence missing data. underlying statistical model encapsulates low-dimensional structure incomplete data matrix sparsity regression coefficients, proposed algorithm jointly learns low-dimensional structure data linear regressor sparse coefficients. proposed stochastic optimization method, sparse linear regression missing data (slrm), performs alternating minimization procedure scales well problem size. large deviation inequalities shed light impact various problem-dependent parameters expected squared loss learned regressor. extensive simulations synthetic real datasets show slrm performs better competing algorithms variety contexts.",18
"spatial temporal graph convolutional networks skeleton-based action recognition. dynamics human body skeletons convey significant information human action recognition. conventional approaches modeling skeletons usually rely hand-crafted parts traversal rules, thus resulting limited expressive power difficulties generalization. work, propose novel model dynamic skeletons called spatial-temporal graph convolutional networks (st-gcn), moves beyond limitations previous methods automatically learning spatial temporal patterns data. formulation leads greater expressive power also stronger generalization capability. two large datasets, kinetics ntu-rgbd, achieves substantial improvements mainstream methods.",3
"formalized lambek calculus higher order logic (hol4). project, rather complete proof-theoretical formalization lambek calculus (non-associative arbitrary extensions) ported coq proof assistent hol4 theorem prover, improvements new theorems. three deduction systems (syntactic calculus, natural deduction sequent calculus) lambek calculus defined many related theorems proved. equivalance systems formally proved. finally, formalization sequent calculus proofs (where coq built-in supports) designed implemented hol4. basic results including sub-formula properties so-called ""cut-free"" proofs formally proved. work considered preliminary work towards language parser based category grammars multimodal still ability support context-sensitive languages customized extensions.",3
"learning distributions via support measure machines. paper presents kernel-based discriminative learning framework probability measures. rather relying large collections vectorial training examples, framework learns using collection probability distributions constructed meaningfully represent training data. representing probability distributions mean embeddings reproducing kernel hilbert space (rkhs), able apply many standard kernel-based learning techniques straightforward fashion. accomplish this, construct generalization support vector machine (svm) called support measure machine (smm). analyses smms provides several insights relationship traditional svms. based insights, propose flexible svm (flex-svm) places different kernel functions training example. experimental results synthetic real-world data demonstrate effectiveness proposed framework.",18
"framework culture-aware robots based fuzzy logic. cultural adaptation, i.e., matching robot's behaviours cultural norms preferences user, well known key requirement success assistive application. however, culture-dependent robot behaviours often implicitly set designers, thus allowing easy automatic adaptation different cultures. paper presents method design culture-aware robots, automatically adapt behaviour conform given culture. propose mapping cultural factors related parameters robot behaviours relies linguistic variables encode heterogeneous cultural factors uniform formalism, fuzzy rules encode qualitative relations among multiple variables. illustrate approach two practical case studies.",3
"minimally faithful inversion graphical models. inference amortization methods allow sharing statistical strength across related observations learning perform posterior inference. generally requires inversion dependency structure generative model, modeller must design learn distribution approximate posterior. previous methods invert dependency structure heuristic way fail capture dependencies model, therefore limiting performance eventual inference algorithm. introduce algorithm faithfully minimally inverting graphical model structure generative model. inversion two crucial properties: a) encode independence assertions absent model, b) given inversion, encodes many true independence assertions possible. algorithm works simulating variable elimination generative model reparametrize distribution. show experiments minimal inversions assist performing better inference.",18
"incremental model-based learners formal learning-time guarantees. model-based learning algorithms shown use experience efficiently learning solve markov decision processes (mdps) finite state action spaces. however, high computational cost due repeatedly solving internal model inhibits use large-scale problems. propose method based real-time dynamic programming (rtdp) speed two model-based algorithms, rmax mbie (model-based interval estimation), resulting computationally much faster algorithms little loss compared existing bounds. specifically, two new learning algorithms, rtdp-rmax rtdp-ie, considerably smaller computational demands rmax mbie. develop general theoretical framework allows us prove efficient learners pac (probably approximately correct) sense. also present experimental evaluation new algorithms helps quantify tradeoff computational experience demands.",3
"dirichlet fragmentation processes. tree structures ubiquitous data across many domains, many datasets naturally modelled unobserved tree structures. paper, first review theory random fragmentation processes [bertoin, 2006], number existing methods modelling trees, including popular nested chinese restaurant process (ncrp). define general class probability distributions trees: dirichlet fragmentation process (dfp) novel combination theory dirichlet processes random fragmentation processes. dfp presents stick-breaking construction, relates ncrp way dirichlet process relates chinese restaurant process. furthermore, develop novel hierarchical mixture model dfp, empirically compare new model similar models machine learning. experiments show dfp mixture model convincingly better existing state-of-the-art approaches hierarchical clustering density modelling.",18
"$o(n\log(n))$ algorithm projecting onto ordered weighted $\ell_1$ norm ball. ordered weighted $\ell_1$ (owl) norm newly developed generalization octogonal shrinkage clustering algorithm regression (oscar) norm. norm desirable statistical properties used perform simultaneous clustering regression. paper, show compute projection $n$-dimensional vector onto owl norm ball $o(n\log(n))$ operations. addition, illustrate performance algorithm synthetic regression test.",10
"experimentally comparing uncertain inference systems probability. paper examines biases performance several uncertain inference systems: mycin, variant mycin. simplified version probability using conditional independence assumptions. present axiomatic arguments using minimum cross entropy inference best way uncertain inference. mycin variant found special situations performance good, also situations performance worse random guessing, data interpreted opposite true import found three systems usually gave accurate results, conditional independence assumptions gave robust results. illustrate importance biases may quantitatively assessed ranked. considerations robustness might critical factor selecting uls's given application.",3
"automatic spatially-adaptive balancing energy terms image segmentation. image segmentation techniques predominately based parameter-laden optimization. objective function typically involves weights balancing competing image fidelity segmentation regularization cost terms. setting weights suitably painstaking, empirical process. even ideal weights found novel image, current approaches fix weight across whole image domain, ignoring spatially-varying properties object shape image appearance. propose novel technique autonomously balances terms spatially-adaptive manner incorporation image reliability graph-based segmentation framework. validate synthetic data achieving reduction mean error 47% (p-value << 0.05) compared best fixed parameter segmentation. also present results medical images (including segmentations corpus callosum brain tissue mri data) natural images.",3
"triaging content severity online mental health forums. mental health forums online communities people express issues seek help moderators users. forums, often posts severe content indicating user acute distress risk attempted self-harm. moderators need respond severe posts timely manner prevent potential self-harm. however, large volume daily posted content makes difficult moderators locate respond critical posts. present framework triaging user content four severity categories defined based indications self-harm ideation. models based feature-rich classification framework includes lexical, psycholinguistic, contextual topic modeling features. approaches improve state art triaging content severity mental health forums large margins (up 17% improvement f-1 scores). using proposed model, analyze mental state users show overall, long-term users forum demonstrate decreased severity risk time. analysis interaction moderators users indicates without automatic way identify critical content, indeed challenging moderators provide timely response users need.",3
"contrary inferences classical histories within consistent histories formulation quantum theory. histories formulation quantum theory, sets coarse-grained histories, called consistent, obey classical probability rules. argued sets describe semi-classical behaviour closed quantum systems. physical scenarios admit multiple different consistent sets one view consistent set separate context. using propositions different consistent sets make inferences leads paradoxes contrary inferences first noted kent [physical review letters, 78(15):2874, 1997]. proponents consistent histories formulation argue one mix propositions coming different consistent sets making logical arguments, paradoxes aforementioned contrary inferences nothing else usual microscopic paradoxes quantum contextuality first demonstrated kochen specker theorem. contribution use consistent histories describe macroscopic (semi-classical) system show paradoxes involving contextuality (mixing different consistent sets) persist even semi-classical limit. distinctively different contextuality standard quantum theory, contextuality paradoxes persist semi-classical limit. specifically, consider different consistent sets arrival time semi-classical wave packet infinite square well. surprisingly, get consistent sets disagree whether motion semi-classical system, started within subregion, ever left subregion not. results point need constraints, additional consistency condition, recover correct semi-classical limit formalism lead motto `all consistent sets equal', `some consistent sets equal others'.",17
"location sensitive deep convolutional neural networks segmentation white matter hyperintensities. anatomical location imaging features crucial importance accurate diagnosis many medical tasks. convolutional neural networks (cnn) huge successes computer vision, lack natural ability incorporate anatomical location decision making process, hindering success medical image analysis tasks. paper, integrate anatomical location information network, propose several deep cnn architectures consider multi-scale patches take explicit location features training. apply compare proposed architectures segmentation white matter hyperintensities brain mr images large dataset. result, observe cnns incorporate location information substantially outperform conventional segmentation method hand-crafted features well cnns integrate location information. test set 46 scans, best configuration networks obtained dice score 0.791, compared 0.797 independent human observer. performance levels machine independent human observer statistically significantly different (p-value=0.17).",3
"practical algorithms learning near-isometric linear embeddings. propose two practical non-convex approaches learning near-isometric, linear embeddings finite sets data points. given set training points $\mathcal{x}$, consider secant set $s(\mathcal{x})$ consists pairwise difference vectors $\mathcal{x}$, normalized lie unit sphere. problem formulated finding symmetric positive semi-definite matrix $\boldsymbol{\psi}$ preserves norms vectors $s(\mathcal{x})$ distortion parameter $\delta$. motivated non-negative matrix factorization, reformulate problem frobenius norm minimization problem, solved alternating direction method multipliers (admm) develop algorithm, fromax. another method solves projection matrix $\boldsymbol{\psi}$ minimizing restricted isometry property (rip) directly set symmetric, postive semi-definite matrices. applying admm moreau decomposition proximal mapping, develop another algorithm, nile-pro, dimensionality reduction. fromax shown converge faster smaller $\delta$ nile-pro converges faster larger $\delta$. non-convex approaches empirically demonstrated computationally efficient prior convex approaches number applications machine learning signal processing.",18
"3d reconstruction incomplete archaeological objects using generative adversarial network. introduce data-driven approach aid repairing conservation archaeological objects: organ, object reconstruction generative adversarial network (gan). using encoder-decoder 3d deep neural network gan architecture, combining two loss objectives: completion loss improved wasserstein gan loss, train network effectively predict missing geometry damaged objects. archaeological objects greatly differ them, network conditioned variable, culture, region metadata object. results, show method recover information damaged objects, even cases half voxels missing, without producing many errors.",3
"simultaneous detection quantification retinal fluid deep learning. propose new deep learning approach automatic detection segmentation fluid within retinal oct images. proposed framework utilizes resnet encoder-decoder neural network architectures. training network, apply novel data augmentation method called myopic warping together standard rotation-based augmentation increase training set size 45 times original amount. finally, network output post-processed energy minimization algorithm (graph cut) along knowledge guided morphological operations finalize segmentation process. based oct imaging data ground truth retouch challenge, proposed system achieves dice indices 0.522, 0.682, 0.612, average absolute volume differences 0.285, 0.115, 0.156 mm$^3$ intaretinal fluid, subretinal fluid, pigment epithelial detachment respectively.",3
"robust combination local controllers. planning problems hard, motion planning, example, ispspace-hard. problems even difficult presence uncertainty. although, markov decision processes (mdps) provide formal framework problems, finding solutions high dimensional continuous mdps usually difficult, especially actions time measurements continuous. fortunately, problem-specific knowledge allows us design controllers good locally, though global guarantees. propose method nonparametrically combining local controllers obtain globally good solutions. apply formulation two types problems : motion planning (stochastic shortest path) discounted mdps. motion planning, argue usual mdp optimality criterion (expected cost) may practically relevant. wepropose alternative: finding minimum cost path,subject constraint robot must reach goal withhigh probability. problem, prove polynomial number samples sufficient obtain high probability path. discounted mdps, propose formulation explicitly deals model uncertainty, i.e., problem introduced transition probabilities known exactly. formulate problem robust linear program directly incorporates type uncertainty.",3
"insights origin halo mass profiles machine learning. mass distribution dark matter haloes result hierarchical growth initial density perturbations mass accretion mergers. use interpretable machine-learning framework provide physical insights origin spherically-averaged mass profile dark matter haloes. train gradient-boosted-trees algorithm predict final mass profiles cluster-sized haloes, measure importance different inputs provided algorithm. find two primary scales initial conditions (ics) impact final mass profile: density approximately scale haloes' lagrangian patch $r_l$ ($r\sim 0.7\, r_l$) large-scale environment ($r\sim 1.7~r_l$). model also identifies three primary time-scales halo assembly history affect final profile: (i) formation time virialized, collapsed material inside halo, (ii) dynamical time, captures dynamically unrelaxed, infalling component halo first orbit, (iii) third, recent time-scale, captures impact outer profile recent massive merger events. inner profile retains memory ics, information alone insufficient yield accurate predictions outer profile. add information haloes' mass accretion history, find significant improvement predicted profiles radii. machine-learning framework provides novel insights role ics mass assembly history determining final mass profile cluster-sized haloes.",0
"proof orthogonal double machine learning $z$-estimators. consider two stage estimation non-parametric first stage generalized method moments second stage, simpler setting (chernozhukov et al. 2016). give alternative proof theorem given (chernozhukov et al. 2016) orthogonal second stage moments, sample splitting $n^{1/4}$-consistency first stage, imply $\sqrt{n}$-consistency asymptotic normality second stage estimates. proof variant estimator, based empirical version moment condition (z-estimator), rather minimization norm empirical vector moments (m-estimator). note meant primarily expository purposes, rather new technical contribution.",18
"knowing look: adaptive attention via visual sentinel image captioning. attention-based neural encoder-decoder frameworks widely adopted image captioning. methods force visual attention active every generated word. however, decoder likely requires little visual information image predict non-visual words ""the"" ""of"". words may seem visual often predicted reliably language model e.g., ""sign"" ""behind red stop"" ""phone"" following ""talking cell"". paper, propose novel adaptive attention model visual sentinel. time step, model decides whether attend image (and so, regions) visual sentinel. model decides whether attend image where, order extract meaningful information sequential word generation. test method coco image captioning 2015 challenge dataset flickr30k. approach sets new state-of-the-art significant margin.",3
"causal discovery continuous additive noise models. consider problem learning causal directed acyclic graphs observational joint distribution. one use graphs predict outcome interventional experiments, data often available. show observational distribution follows structural equation model additive noise structure, directed acyclic graph becomes identifiable distribution mild conditions. constitutes interesting alternative traditional methods assume faithfulness identify markov equivalence class graph, thus leaving edges undirected. provide practical algorithms finitely many samples, resit (regression subsequent independence test) two methods based independence score. prove resit correct population setting provide empirical evaluation.",18
"track facial points unconstrained videos. tracking facial points unconstrained videos challenging due non-rigid deformation changes time. paper, propose exploit incremental learning person-specific alignment wild conditions. approach takes advantage part-based representation cascade regression robust efficient alignment frame. unlike existing methods usually rely models trained offline, incrementally update representation subspace cascade regressors unified framework achieve personalized modeling fly. alleviate drifting issue, fitting results evaluated using deep neural network, well-aligned faces picked incrementally update representation fitting models. image video datasets employed valid proposed method. results demonstrate superior performance approach compared existing approaches terms fitting accuracy efficiency.",3
"generation two-layer monotonic functions. problem implementing class functions particular conditions using monotonic multilayer functions considered. genetic algorithm used create monotonic functions certain class, implemented two-layer monotonic functions. existence solution given problem suggests two monotone functions, monotonic function dimensions created. new algorithm based genetic algorithm proposed, easily implemented two-layer monotonic functions specific class six variables.",3
"applications algorithmic probability philosophy mind. paper presents formulae solve various seemingly hopeless philosophical conundrums. discuss simulation argument, teleportation, mind-uploading, rationality utilitarianism, ethics exploiting artificial general intelligence. approach arises combining essential ideas formalisms algorithmic probability, universal intelligence measure, space-time-embedded intelligence, hutter's observer localization. argue universal models yield ultimate solutions, novel research direction would required order find computationally efficient approximations thereof.",3
"prolongation smap spatio-temporally seamless coverage continental us using deep learning neural network. soil moisture active passive (smap) mission delivered valuable sensing surface soil moisture since 2015. however, short time span irregular revisit schedule. utilizing state-of-the-art time-series deep learning neural network, long short-term memory (lstm), created system predicts smap level-3 soil moisture data atmospheric forcing, model-simulated moisture, static physiographic attributes inputs. system removes bias model simulations improves predicted moisture climatology, achieving small test root-mean-squared error (<0.035) high correlation coefficient >0.87 75\% continental united states, including forested southeast. first application lstm hydrology, show proposed network avoids overfitting robust temporal spatial extrapolation tests. lstm generalizes well across regions distinct climates physiography. high fidelity smap, lstm shows great potential hindcasting, data assimilation, weather forecasting.",18
"scalable text mining sparse generative models. information age brought deluge data. much text form, insurmountable scope humans incomprehensible structure computers. text mining expanding field research seeks utilize information contained vast document collections. general data mining methods based machine learning face challenges scale text data, posing need scalable text mining methods. thesis proposes solution scalable text mining: generative models combined sparse computation. unifying formalization generative text models defined, bringing together research traditions used formally equivalent models, ignored parallel developments. framework allows use methods developed different processing tasks retrieval classification, yielding effective solutions across different text mining tasks. sparse computation using inverted indices proposed inference probabilistic models. reduces computational complexity common text mining operations according sparsity, yielding probabilistic models scalability modern search engines. proposed combination provides sparse generative models: solution text mining general, effective, scalable. extensive experimentation text classification ranked retrieval datasets conducted, showing proposed solution matches outperforms leading task-specific methods effectiveness, order magnitude decrease classification times wikipedia article categorization million classes. developed methods applied two 2014 kaggle data mining prize competitions hundred competing teams, earning first second places.",3
"augur: modeling language data-parallel probabilistic inference. time-consuming error-prone implement inference procedures new probabilistic model. probabilistic programming addresses problem allowing user specify model compiler automatically generate inference procedure it. approach practical, important generate inference code reasonable performance. paper, present probabilistic programming language compiler bayesian networks designed make effective use data-parallel architectures gpus. language fully integrated within scala programming language benefits tools ide support, type-checking, code completion. show compiler generate data-parallel inference code scalable thousands gpu cores making use conditional independence relationships bayesian network.",18
"fused elastic net logistic regression model multi-task binary classification. multi-task learning shown significantly enhance performance multiple related learning tasks variety situations. present fused logistic regression, sparse multi-task learning approach binary classification. specifically, introduce sparsity inducing penalties parameter differences related logistic regression models encode similarity across related tasks. resulting joint learning task cast form lends efficiently optimized recursive variant alternating direction method multipliers. show results synthetic data describe regime settings multi-task approach achieves significant improvements single task learning approach discuss implications applying fused logistic regression different real world settings.",18
"encrypted accelerated least squares regression. information stored encrypted format is, definition, usually amenable statistical analysis machine learning methods. paper present detailed analysis coordinate accelerated gradient descent algorithms capable fitting least squares penalised ridge regression models, using data encrypted fully homomorphic encryption scheme. gradient descent shown dominate terms encrypted computational speed, theoretical results proven give parameter bounds ensure correctness decryption. characteristics encrypted computation empirically shown favour non-standard acceleration technique. demonstrates possibility approximating conventional statistical regression methods using encrypted data without compromising privacy.",18
"filtering algorithms multiset ordering constraint. constraint programming (cp) used great success tackle wide variety constraint satisfaction problems computationally intractable general. global constraints one important factors behind success cp. paper, study new global constraint, multiset ordering constraint, shown useful symmetry breaking searching leximin optimal solutions cp. propose efficient effective filtering algorithms propagating global constraint. show algorithms sound complete discuss possible extensions. also consider alternative propagation methods based existing constraints cp toolkits. experimental results number benchmark problems demonstrate propagating multiset ordering constraint via dedicated algorithm beneficial.",3
"leril : collaborative effort creating lexical resources. paper reports efforts taken create lexical resources pertaining indian languages, using collaborative model. lexical resources developed are: (1) transfer lexicon grammar english several indian languages. (2) dependencey tree bank annotated corpora several indian languages. dependency trees based paninian model. (3) bilingual dictionary 'core meanings'.",3
"computational limits matrix completion. matrix completion problem recovering unknown real-valued low-rank matrix subsample entries. important recent results show problem solved efficiently assumption unknown matrix incoherent subsample drawn uniformly random. assumptions necessary? well known matrix completion full generality np-hard. however, little known make additional assumptions incoherence permit algorithm output matrix slightly higher rank. paper prove matrix completion remains computationally intractable even unknown matrix rank $4$ allowed output constant rank matrix, even additionally assume unknown matrix incoherent shown $90%$ entries. result relies conjectured hardness $4$-coloring problem. also consider positive semidefinite matrix completion problem. show similar hardness result standard assumption $\mathrm{p}\ne \mathrm{np}.$ results greatly narrow gap existing feasibility results computational lower bounds. particular, believe results give first complexity-theoretic justification distributional assumptions needed beyond incoherence assumption order obtain positive results. technical side, contribute several new ideas encode hard combinatorial problems low-rank optimization problems. hope techniques helpful understanding computational limits matrix completion related problems.",3
"learning efficient sparse low rank models. parsimony, including sparsity low rank, shown successfully model data numerous machine learning signal processing tasks. traditionally, modeling approaches rely iterative algorithm minimizes objective function parsimony-promoting terms. inherently sequential structure data-dependent complexity latency iterative optimization constitute major limitation many applications requiring real-time performance involving large-scale data. another limitation encountered modeling techniques difficulty inclusion discriminative learning scenarios. work, propose move emphasis model pursuit algorithm, develop process-centric view parsimonious modeling, learned deterministic fixed-complexity pursuit process used lieu iterative optimization. show principled way construct learnable pursuit process architectures structured sparse robust low rank models, derived iteration proximal descent algorithms. architectures learn approximate exact parsimonious representation fraction complexity standard optimization methods. also show appropriate training regimes allow naturally extend parsimonious models discriminative settings. state-of-the-art results demonstrated several challenging problems image audio processing several orders magnitude speedup compared exact optimization algorithms.",3
"tensor completion algorithms big data analytics. tensor completion problem filling missing unobserved entries partially observed tensors. due multidimensional character tensors describing complex datasets, tensor completion algorithms applications received wide attention achievement data mining, computer vision, signal processing, neuroscience, etc. survey, provide modern overview recent advances tensor completion algorithms perspective big data analytics characterized diverse variety, large volume, high velocity. towards better comprehension comparison vast existing advances, summarize categorize four groups including general tensor completion algorithms, tensor completion auxiliary information (variety), scalable tensor completion algorithms (volume) dynamic tensor completion algorithms (velocity). besides, introduce applications real-world data-driven problems present open-source package covering several widely used tensor decomposition completion algorithms. goal summarize popular methods introduce researchers promoting research process field give available repository practitioners. end, also discuss challenges promising research directions community future explorations.",18
"detecting english writing styles non native speakers. paper presents first attempt, knowledge, classify english writing styles scale challenge classifying day day language written writers different backgrounds covering various areas topics.the paper proposes simple machine learning algorithms simple generate features solve hard problems. relying scale data available large sources knowledge like wikipedia. believe sources data crucial generate robust solutions web high accuracy easy deploy practice. paper achieves 74\% accuracy classifying native versus non native speakers writing styles. moreover, paper shows interesting observations similarity different languages measured similarity users english writing styles. technique could used show well known facts languages grouping families, experiments support.",3
"note equivalence hierarchical variational models auxiliary deep generative models. note compares two recently published machine learning methods constructing flexible, tractable families variational hidden-variable posteriors. first method, called ""hierarchical variational models"" enriches inference model extra variable, other, called ""auxiliary deep generative models"", enriches generative model instead. conclude two methods mathematically equivalent.",18
"learning probability measures respect optimal transport metrics. study problem estimating, sense optimal transport metrics, measure assumed supported manifold embedded hilbert space. establishing precise connection optimal transport metrics, optimal quantization, learning theory, derive new probabilistic bounds performance classic algorithm unsupervised learning (k-means), used produce probability measure derived data. course analysis, arrive new lower bounds, well probabilistic upper bounds convergence rate empirical law large numbers, which, unlike existing bounds, applicable wide class measures.",3
"likelihood plausibility. several authors explained likelihood ratio measures strength evidence represented observations statistical problems. idea works fine goal evaluate strength available evidence simple hypothesis versus another simple hypothesis. however, applicability idea limited simple hypotheses likelihood function primarily defined points (simple hypotheses) parameter space. paper define general weight evidence applicable simple composite hypotheses. based dempster-shafer concept plausibility shown generalization likelihood ratio. functional models fundamental importance general weight evidence proposed paper. relevant concepts ideas explained means familiar urn problem general analysis real-world medical problem presented.",3
"importance normative data speech-based assessment. data sets identifying alzheimer's disease (ad) often relatively sparse, limits ability train generalizable models. here, augment data set, dementiabank, two normative data sets, wisconsin longitudinal study talk2me, employs speech-based picture-description assessment. minority class oversampling adasyn, outperform state-of-the-art results binary classification people without ad dementiabank. work highlights effectiveness combining sparse difficult-to-acquire patient data relatively large easily accessible normative datasets.",3
"malware detection eating whole exe. work introduce malware detection raw byte sequences fruitful research area larger machine learning community. building neural network problem presents number interesting challenges occurred tasks image processing nlp. particular, note detection raw bytes presents sequence problem two million time steps problem batch normalization appear hinder learning process. present initial work building solution tackle problem, linear complexity dependence sequence length, allows interpretable sub-regions binary identified. discuss many challenges building neural network process data scale, methods used work around them.",18
"multi-frame image super-resolution fast upscaling technique. multi-frame image super-resolution (misr) aims fuse information low-resolution (lr) image sequence compose high-resolution (hr) one, applied extensively many areas recently. different single image super-resolution (sisr), sub-pixel transitions multiple frames introduce additional information, attaching significance fusion operator alleviate ill-posedness misr. reconstruction-based approaches, inevitable projection reconstruction errors lr space hr space commonly tackled interpolation operator, however crude interpolation may fit natural image generate annoying blurring artifacts, especially fusion operator. paper, propose end-to-end fast upscaling technique replace interpolation operator, design upscaling filters lr space periodic sub-locations respectively shuffle filter results derive final reconstruction errors hr space. proposed fast upscaling technique reduce computational complexity upscaling operation utilizing shuffling operation avoid complex operation hr space, also realize superior performance fewer blurring artifacts. extensive experimental results demonstrate effectiveness efficiency proposed technique, whilst, combining proposed technique bilateral total variation (btv) regu-larization, misr approach outperforms state-of-the-art methods.",3
"transition-based parsing context enhancement future reward reranking. paper presents novel reranking model, future reward reranking, re-score actions transition-based parser using global scorer. different conventional reranking parsing, model searches best dependency tree feasible trees constraining sequence actions get future reward sequence. scorer based first-order graph-based parser bidirectional lstm, catches different parsing view compared transition-based parser. besides, since context enhancement shown substantial improvement arc-stand transition-based parsing parsing accuracy, implement context enhancement arc-eager transition-base parser stack lstms, dynamic oracle dropout supporting achieve improvement. global scorer context enhancement, results show uas parser increases much 1.20% english 1.66% chinese, las increases much 1.32% english 1.63% chinese. moreover, get state-of-the-art lass, achieving 87.58% chinese 93.37% english.",3
"harder, better, faster, stronger convergence rates least-squares regression. consider optimization quadratic objective function whose gradients accessible stochastic oracle returns gradient given point plus zero-mean finite variance random error. present first algorithm achieves jointly optimal prediction error rates least-squares regression, terms forgetting initial conditions o(1/n 2), terms dependence noise dimension problem, o(d/n). new algorithm based averaged accelerated regularized gradient descent, may also analyzed finer assumptions initial conditions hessian matrix, leading dimension-free quantities may still small "" optimal "" terms large. order characterize tightness new bounds, consider application non-parametric regression use known lower bounds statistical performance (without computational limits), happen match bounds obtained single pass data thus show optimality algorithm wide variety particular trade-offs bias variance.",10
"natural data structure extracted neighborhood-similarity graphs. 'big' high-dimensional data commonly analyzed low-dimensions, performing dimensionality-reduction step inherently distorts data structure. purpose, clustering methods also often used. methods also introduce bias, either starting assumption particular geometric form clusters, using iterative schemes enhance cluster contours, uncontrollable consequences. goal data analysis should, however, encode detect structural data features scales densities simultaneously, without assuming parametric form data point distances, modifying them. propose novel approach directly encodes data point neighborhood similarities sparse graph. non-iterative framework permits transparent interpretation data, without altering original data dimension metric. several natural synthetic data applications demonstrate efficacy novel approach.",18
"characterization gradient dominance regularity conditions neural networks. past decade witnessed successful application deep learning solving many challenging problems machine learning artificial intelligence. however, loss functions deep neural networks (especially nonlinear networks) still far well understood theoretical aspect. paper, enrich current understanding landscape square loss functions three types neural networks. specifically, parameter matrices square, provide explicit characterization global minimizers linear networks, linear residual networks, nonlinear networks one hidden layer. then, establish two quadratic types landscape properties square loss neural networks, i.e., gradient dominance condition within neighborhood full rank global minimizers, regularity condition along certain directions within neighborhood global minimizers. two landscape properties desirable optimization around global minimizers loss function neural networks.",18
"filtrations torsion pairs abramovich polishchuk's heart. study abelian subcategories torsion pairs abramovich polishchuk's heart. construct stability conditions full triangulated subcategory $\mathcal{d}^{\leq 1}_s$ $d(x\times s)$, arbitrary smooth projective variety s. also define notion $l$-th level stability, generalization slope stability gieseker stability. show object e abramovich polishchuk's heart, unique filtration whose factors $l$-th level semistable, phase vectors decreasing lexicographic order.",10
"comparison time-frequency representations environmental sound classification using convolutional neural networks. recent successful applications convolutional neural networks (cnns) audio classification speech recognition motivated search better input representations efficient training. visual displays audio signal, various time-frequency representations spectrograms offer rich representation temporal spectral structure original signal. letter, compare various popular signal processing methods obtain representation, short-time fourier transform (stft) linear mel scales, constant-q transform (cqt) continuous wavelet transform (cwt), assess impact classification performance two environmental sound datasets using cnns. study supports hypothesis time-frequency representations valuable learning useful features sound classification. moreover, actual transformation used shown impact classification accuracy, mel-scaled stft outperforming discussed methods slightly baseline mfcc features large degree. additionally, observe optimal window size transformation dependent characteristics audio signal architecturally, 2d convolution yielded better results cases compared 1d.",3
"conjugacy classes maximal cyclic subgroups metacyclic $p$-groups. paper, set $\eta (g)$ number conjugacy classes maximal cyclic subgroups finite group $g$. compute $\eta (g)$ metacyclic $p$-groups. show $g$ metacyclic $p$-group order $p^n$ dihedral, generalized quaternion, semi-dihedral, $\eta (g) \ge n-2$, determine equality holds.",10
"action-driven object detection top-down visual attentions. dominant paradigm deep learning based object detection relies ""bottom-up"" approach using ""passive"" scoring class agnostic proposals. approaches efficient lack holistic analysis scene-level context. paper, present ""action-driven"" detection mechanism using ""top-down"" visual attention model. localize object taking sequential actions attention model provides. attention model conditioned image region provides required actions get closer toward target object. action time step weak ensemble sequential actions makes bounding-box accurately converge target object boundary. attention model call attentionnet composed convolutional neural network. whole detection procedure, utilize actions single attentionnet without modules object proposals post bounding-box regression. evaluate top-down detection mechanism pascal voc series ilsvrc cls-loc dataset, achieve state-of-the-art performances compared major bottom-up detection methods. particular, detection mechanism shows strong advantage elaborate localization outperforming faster r-cnn margin +7.1% pascal voc 2007 increase iou threshold positive detection 0.7.",3
"non-constant bounded holomorphic functions hyperbolic numbers - candidates hyperbolic activation functions. liouville theorem states bounded holomorphic complex functions necessarily constant. holomorphic functions fulfill socalled cauchy-riemann (cr) conditions. cr conditions mean complex $z$-derivative independent direction. holomorphic functions ideal activation functions complex neural networks, liouville theorem makes useless. yet recently use hyperbolic numbers, lead construction hyperbolic number neural networks. describe cauchy-riemann conditions hyperbolic numbers show exists new interesting type bounded holomorphic functions hyperbolic numbers, constant. give examples functions. therefore substantially expand available candidates holomorphic activation functions hyperbolic number neural networks. keywords: hyperbolic numbers, liouville theorem, cauchy-riemann conditions, bounded holomorphic functions",3
"random multi-constraint projection: stochastic gradient methods convex optimization many constraints. consider convex optimization problems subject large number constraints. focus stochastic problems objective takes form expected values feasible set intersection large number convex sets. propose class algorithms perform stochastic gradient descent random feasibility updates simultaneously. every iteration, algorithms sample number projection points onto randomly selected small subsets constraints. three feasibility update schemes considered: averaging random projected points, projecting onto distant sample, projecting onto special polyhedral set constructed based sample points. prove almost sure convergence algorithms, analyze iterates' feasibility error optimality error, respectively. provide new convergence rate benchmarks stochastic first-order optimization many constraints. rate analysis numerical experiments reveal algorithm using polyhedral-set projection scheme efficient one within known algorithms.",18
"$c^\infty$-manifolds skeletal diffeology. formulate study notion $d$-skeletal diffeology, generalizes wire diffeology, introducing dual notion $d$-coskeletal diffeology. first show paracompact finite-dimensional $c^\infty$-manifolds $m_d$ $d$-skeletal diffeology inherit good topologies smooth paracompactness $m$. study pathology $m_d$. all, prove following: $d<{\rm dim}\ m$, every immersion $f:m\longrightarrow n$ isolated diffeological space $\dcal(m_d, n_d)$ smooth maps $d$-dimensional smooth homotopy group $m_d$ uncountable.",10
"algorithmic identification probabilities. tthe problem identify probability associated set natural numbers, given infinite data sequence elements set. given sequence drawn i.i.d. probability mass function involved (the target) belongs computably enumerable (c.e.) co-computably enumerable (co-c.e.) set computable probability mass functions, algorithm almost surely identify target limit. technical tool strong law large numbers. set finite elements sequence dependent sequence typical sense martin-l\""of least one measure belonging c.e. co-c.e. set computable measures, algorithm identify limit computable measure sequence typical (there may one measure). technical tool theory kolmogorov complexity. give algorithms consider associated predictions.",3
"post-inference prior swapping. bayesian methods praised ability incorporate useful prior knowledge, practice, convenient priors allow computationally cheap tractable inference commonly used. paper, investigate following question: given model, possible compute inference result convenient false prior, afterwards, given target prior interest, quickly transform result target posterior? potential solution use importance sampling (is). however, demonstrate fail many choices target prior, depending parametric form similarity false prior. instead, propose prior swapping, method leverages pre-inferred false posterior efficiently generate accurate posterior samples arbitrary target priors. prior swapping lets us apply less-costly inference algorithms certain models, incorporate new updated prior information ""post-inference"". give theoretical guarantees method, demonstrate empirically number models priors.",18
"note classes subgraphs locally finite graphs. investigate question `small' graph be, contains members given class locally finite graphs subgraphs induced subgraphs. precisely, give necessary sufficient conditions existence connected, locally finite graph $h$ containing elements graph class $\mathcal g$. conditions imply graph $h$ exists class $\mathcal g_d$ consisting graphs maximum degree $<d$ raises question whether case $h$ chosen bounded maximum degree. show case, thereby answering question recently posed huynh et al.",10
"creep failure amorphous solids tensile stress. applying constant tensile stress piece amorphous solid results slow extension, followed eventual rapid mechanical collapse. ""creep"" process paramount engineering concern, subject study variety materials, century. predictive theories $\tau_w$, expected time collapse, lacking, mainly due dependence bewildering variety parameters, including temperature, system size, tensile force, also detailed microscopic interactions constituents. complex dependence collapse time parameters discussed below, using simulations strip amorphous material. different scenarios observed ductile brittle materials, resulting serious difficulties creating all-encompassing theory could offer safety measures given conditions. central aim paper employ scaling concepts, achieve data collapse probability distribution function (pdf) $\ln{\tau_w}$. scaling ideas result universal function provides prediction pdf $\ln{\tau_w}$ out-of-sample systems, measurements values parameters. predictive power scaling theory demonstrated ductile brittle systems. finally, present derivation universal scaling function brittle materials. ductile case appears due plastic necking instability left future research.",2
"first step toward model-free, anonymous object tracking recurrent neural networks. paper, propose study novel visual object tracking approach based convolutional networks recurrent networks. proposed approach distinct existing approaches visual object tracking, filtering-based ones tracking-by-detection ones, sense tracking system explicitly trained off-line track anonymous objects noisy environment. proposed visual tracking model end-to-end trainable, minimizing adversarial effect mismatches object representation true underlying dynamics learning dynamics. empirically show proposed tracking approach works well various scenarios generating artificial video sequences varying conditions; number objects, amount noise match training shapes test shapes.",3
"excitation surface plasmon polaritons diffraction-free vector beams. surface plasmon polaritons (spps) traditionally excited plane waves within rayleigh range focused transverse magnetic (tm) gaussian beam. here, investigate confirm coupling spps two-dimensional gaussian bessel-gauss wave packets, well one-dimensional light sheets space-time wave packets. encode incoming wavefronts spatially varying states polarization couple respective tm components radial azimuthal vector beam profiles confirm spp polarization-correlation spatial-mode selectivity. results require material optimization multi-dimensional confinement via periodically corrugated metal surfaces achieve coupling greater extents. hereby, outlining pivotal, yet commonly overlooked, path towards development long-range biosensors all-optical integrated plasmonic circuits.",14
"submodular variational inference network reconstruction. real-world online social networks, individuals receive transmit information real time. cascading information transmissions (e.g. phone calls, text messages, social media posts) may understood realization diffusion process operating network, branching path represented directed tree. process traverses thus reveals limited portion edges. network reconstruction/inference problem infer unrevealed connections. existing approaches derive likelihood attempt find network topology maximizing likelihood, problem highly intractable. paper, focus network reconstruction problem broad class real-world diffusion processes, exemplified network diffusion scheme called respondent-driven sampling (rds). prove realistic general models network diffusion, posterior distribution observed rds realization bayesian log-submodular model.we propose vine (variational inference network reconstruction), novel, accurate, computationally efficient variational inference algorithm, network reconstruction problem model. crucially, assume particular probabilistic model underlying network. vine recovers connected graph high accuracy shown experimental results real-life networks.",3
"detecting pulmonary embolism computed tomography using convolutional neural network. clinical symptoms pulmonary embolism (pe) diverse non-specific, makes difficult diagnose. addition, pulmonary embolism multiple triggers one major causes vascular death. therefore, detected treated quickly, significantly reduce risk death hospitalized patients. detection process, cost computed tomography pulmonary angiography (ctpa) high, angiography requires injection contrast agents, increase risk damage patient. therefore, study use deep learning approach detect pulmonary embolism patients take ct image chest using convolutional neural network. proposed pulmonary embolism detection system, detect possibility pulmonary embolism time patient's first ct image, schedule ctpa test immediately, saving week ct image screening time providing timely diagnosis treatment patient.",5
zero resource speech challenge 2017. describe new challenge aimed discovering subword word units raw speech. challenge followup zero resource speech challenge 2015. aims constructing systems generalize across languages adapt new speakers. design features evaluation metrics challenge presented results seventeen models discussed.,3
"supernova neutrino decoupling altered flavor conversion. large neutrino density deep interior core-collapse supernovae makes neutrino flavor evolution non-linear coherent forward scattering neutrinos among themselves. assumption spherical symmetry, model neutrino decoupling matter present first non-linear simulation flavor evolution presence charged current neutral current collisions neutrino advection. flavor transformation occurs neutrinos fully decoupled matter, dynamically affecting flavor distributions neutrino species shifting location neutrino decoupling surfaces. findings may implications explosion mechanism supernovae, nucleosynthesis heavy elements, well observable neutrino signal, yet assessed.",8
"sets sum difference structure. nonempty sets $a,b$ nonnegative integers integer $n$, let $r_{a,b}(n)$ number representations $n$ $a+b$ $d_{a,b}(n)$ number representations $n$ $a-b$, $a\in a, b\in b$. paper, determine sets $a,b$ $r_{a,b}(n)=1$ every nonnegative integer $n$. also consider \emph{difference} structure prove that: exist sets $a$ $b$ nonnegative integers $r_{a,b}(n)\ge 1$ large $n$, $a(x)b(x)=(1+o(1))x$ given nonnegative integer $c$, $d_{a,b}(n)=c$ infinitely many positive integers $n$. related results also contained.",10
"german flux: detecting metaphoric change via word entropy. paper explores information-theoretic measure entropy detect metaphoric change, transferring ideas hypernym detection research language change. also build first diachronic test set german standard metaphoric change annotation. model shows high performance, unsupervised, language-independent generalizable processes semantic change.",3
"epistemological problems knowledge level cognitive architectures. article addresses open problem area cognitive systems architectures: namely problem handling (in terms processing reasoning capabilities) complex knowledge structures least plausibly comparable, terms size typology encoded information, knowledge humans process daily executing everyday activities. handling huge amount knowledge, selectively retrieve ac- cording needs emerging different situational scenarios, important aspect human intelligence. task, fact, humans adopt wide range heuristics (gigerenzer todd) due bounded rationality (simon, 1957). perspective, one re- quirements considered design, realization evaluation intelligent cognitively inspired systems represented ability heuristically identify retrieve, general knowledge stored artificial long term memory (ltm), one synthetically contextually relevant. require- ment, however, often neglected. currently, artificial cognitive systems architectures able, de facto, deal complex knowledge structures even slightly comparable knowledge heuris- tically managed humans. paper argue technological problem also epistemological one briefly sketch proposal possible solution.",3
"bi-text alignment movie subtitles spoken english-arabic statistical machine translation. describe efforts towards getting better resources english-arabic machine translation spoken text. particular, look movie subtitles unique, rich resource, subtitles one language often get translated languages. movie subtitles new resource explored previous research; however, create much larger bi-text (the biggest date), generate better quality alignment it. given subtitles movie different languages, key problem align fragment level. typically, done using length-based alignment, movie subtitles, also time information. exploit information develop original algorithm outperforms current best subtitle alignment tool, subalign. evaluation results show adding bi-text iwslt training bi-text yields improvement two bleu points absolute.",3
"novel framework expedite systematic reviews automatically building information extraction training corpora. systematic review identifies collates various clinical studies compares data elements results order provide evidence based answer particular clinical question. process manual involves lot time. tool automate process lacking. aim work develop framework using natural language processing machine learning build information extraction algorithms identify data elements new primary publication, without go expensive task manual annotation build gold standards data element type. system developed two stages. initially, uses information contained existing systematic reviews identify sentences pdf files included references contain specific data elements interest using modified jaccard similarity measure. sentences treated labeled data.a support vector machine (svm) classifier trained labeled data extract data elements interests new article. conducted experiments cochrane database systematic reviews related congestive heart failure using inclusion criteria example data element. empirical results show proposed system automatically identifies sentences containing data element interest high recall (93.75%) reasonable precision (27.05% - means reviewers read 3.7 sentences average). empirical results suggest tool retrieving valuable information reference articles, even time-consuming identify manually. thus hope tool useful automatic data extraction biomedical research publications. future scope work generalize information framework types systematic reviews.",3
"visualizing textual models in-text word-as-pixel highlighting. explore two techniques use color make sense statistical text models. one method uses in-text annotations illustrate model's view particular tokens particular documents. another uses high-level, ""words-as-pixels"" graphic display entire corpus. together, methods offer zoomed-in zoomed-out perspectives model's understanding text. show interconnected methods help diagnose classifier's poor performance twitter slang, make sense topic model historical political texts.",18
"inferring executing programs visual reasoning. existing methods visual reasoning attempt directly map inputs outputs using black-box architectures without explicitly modeling underlying reasoning processes. result, black-box models often learn exploit biases data rather learning perform visual reasoning. inspired module networks, paper proposes model visual reasoning consists program generator constructs explicit representation reasoning process performed, execution engine executes resulting program produce answer. program generator execution engine implemented neural networks, trained using combination backpropagation reinforce. using clevr benchmark visual reasoning, show model significantly outperforms strong baselines generalizes better variety settings.",3
"active model aggregation via stochastic mirror descent. consider problem learning convex aggregation models, good best convex aggregation, binary classification problem. working stream based active learning setting, active learner make decision on-the-fly, wants query label point currently seen stream, propose stochastic-mirror descent algorithm, called smd-ama, entropy regularization. establish excess risk bounds loss convex aggregate returned smd-ama order $o\left(\sqrt{\frac{\log(m)}{{t^{1-\mu}}}}\right)$, $\mu\in [0,1)$ algorithm dependent parameter, trades-off number labels queried, excess risk.",18
"detecting unusual part digital image. purpose paper introduce algorithm detect unusual part digital image. unusual part given shape defined part image maximal distance non intersecting shapes form. method used scan image databases clear model interesting part large image databases, example medical databases.",3
"multispectral image intrinsic decomposition via low rank constraint. multispectral images contain many clues surface characteristics objects, thus widely used many computer vision tasks, e.g., recolorization segmentation. however, due complex illumination geometry structure natural scenes, spectra curves surface look different. paper, low rank multispectral image intrinsic decomposition model (lriid) presented decompose shading reflectance single multispectral image. extend retinex model, proposed rgb image intrinsic decomposition, multispectral domain. based this, low rank constraint proposed reduce ill-posedness problem make algorithm solvable. dataset 12 images given ground truth shadings reflectance, objective evaluations conducted. experiments demonstrate effectiveness proposed method.",3
"affine-gradient based local binary pattern descriptor texture classiffication. present novel affine-gradient based local binary pattern (aglbp) descriptor texture classification. hard describe complicated texture using single type information, local binary pattern (lbp), utilizes sign information difference pixel local neighbors. descriptor three characteristics: 1) order make full use information contained texture, affine-gradient, different euclidean-gradient invariant affine transformation incorporated aglbp. 2) improved method proposed rotation invariance, depends reference direction calculating respect local neighbors. 3) feature selection method, considering statistical frequency intraclass variance training dataset, also applied reduce dimensionality descriptors. experiments three standard texture datasets, outex12, outex10 kth-tips2, conducted evaluate performance aglbp. results show proposed descriptor gets better performance comparing state-of-the-art rotation texture descriptors texture classification.",3
"mean actor critic. propose new algorithm, mean actor-critic (mac), discrete-action continuous-state reinforcement learning. mac policy gradient algorithm uses agent's explicit representation action values estimate gradient policy, rather using actions actually executed. significantly reduces variance gradient updates removes need variance reduction baseline. show empirical results two control domains mac performs well better policy gradient approaches, five atari games, mac competitive state-of-the-art policy search algorithms.",18
"personalized machine learning robot perception affect engagement autism therapy. robots great potential facilitate future therapies children autism spectrum. however, existing robots lack ability automatically perceive respond human affect, necessary establishing maintaining engaging interactions. moreover, inference challenge made harder fact many individuals autism atypical unusually diverse styles expressing affective-cognitive states. tackle heterogeneity behavioral cues children autism, use latest advances deep learning formulate personalized machine learning (ml) framework automatic perception childrens affective states engagement robot-assisted autism therapy. key approach novel shift traditional ml paradigm - instead using 'one-size-fits-all' ml models, personalized ml framework optimized child leveraging relevant contextual information (demographics behavioral assessment scores) individual characteristics child. designed evaluated framework using dataset multi-modal audio, video autonomic physiology data 35 children autism (age 3-13) 2 cultures (asia europe), participating 25-minute child-robot interaction (~500k datapoints). experiments confirm feasibility robot perception affect engagement, showing clear improvements due model personalization. proposed approach potential improve existing therapies autism offering efficient monitoring summarization therapy progress.",3
"english conversational telephone speech recognition humans machines. one difficult speech recognition tasks accurate recognition human human communication. advances deep learning last years produced major speech recognition improvements representative switchboard conversational corpus. word error rates years ago 14% dropped 8.0%, 6.6% recently 5.8%, believed within striking range human performance. raises two issues - human performance, far still drive speech recognition error rates? recent paper microsoft suggests already achieved human performance. trying verify statement, performed independent set human performance measurements two conversational tasks found human performance may considerably better earlier reported, giving community significantly harder goal achieve. also report efforts area, presenting set acoustic language modeling techniques lowered word error rate english conversational telephone lvcsr system level 5.5%/10.3% switchboard/callhome subsets hub5 2000 evaluation, - least writing paper - new performance milestone (albeit measure human performance!). acoustic side, use score fusion three models: one lstm multiple feature inputs, second lstm trained speaker-adversarial multi-task learning third residual net (resnet) 25 convolutional layers time-dilated convolutions. language modeling side, use word character lstms convolutional wavenet-style language models.",3
"actively learning attract followers twitter. twitter, popular social network, presents great opportunities on-line machine learning research. however, previous research focused almost entirely learning passively collected data. study problem learning acquire followers normative user behavior, opposed mass following policies applied many bots. formalize problem contextual bandit problem, consider retweeting content action chosen tweet (content) accompanied context. design reward signals based change followers. result month long experiment 60 agents suggests (1) aggregating experience across agents adversely impact prediction accuracy (2) twitter community's response different actions non-stationary. findings suggest actively learning on-line provide deeper insights attract followers machine learning passively collected data alone.",18
"nonparanormal: semiparametric estimation high dimensional undirected graphs. recent methods estimating sparse undirected graphs real-valued data high dimensional problems rely heavily assumption normality. show use semiparametric gaussian copula--or ""nonparanormal""--for high dimensional inference. additive models extend linear models replacing linear functions set one-dimensional smooth functions, nonparanormal extends normal transforming variables smooth functions. derive method estimating nonparanormal, study method's theoretical properties, show works well many examples.",18
"dictionary image recovery incomplete random measurements. paper tackles algorithmic theoretical aspects dictionary learning incomplete random block-wise image measurements performance adaptive dictionary sparse image recovery. problem related blind compressed sensing sparsifying dictionary basis viewed unknown variable subject estimation sparse recovery. however, unlike existing guarantees successful blind compressed sensing, results rely additional structural constraints learned dictionary measured signal. particular, rely spatial diversity compressive measurements guarantee solution unique high probability. moreover, distinguishing goal measure reduce estimation error respect ideal dictionary based complete image. using recent results random matrix theory, show applying slightly modified dictionary learning algorithm compressive measurements results accurate estimation ideal dictionary large-scale images. empirically, experiment space-invariant space-varying sensing matrices demonstrate critical role spatial diversity measurements. simulation results confirm presented algorithm outperforms typical non-adaptive sparse recovery based offline-learned universal dictionaries.",3
"truncated tensor schatten p-norm based approach spatiotemporal traffic data imputation complicated missing patterns. rapid advances sensor, wireless communication, cloud computing data science brought unprecedented amount data assist transportation engineers researchers making better decisions. however, traffic data reality often corrupted incomplete values due detector communication malfunctions. data imputation thus required ensure effectiveness downstream data-driven applications. end, numerous tensor-based methods treating imputation problem low-rank tensor completion (lrtc) attempted previous works. tackle rank minimization, core lrtc, aforementioned methods utilize tensor nuclear norm (nn) convex surrogate minimization. however, over-relaxation issue nn refrains desirable performance practice. paper, define innovative nonconvex truncated schatten p-norm tensors (tspn) approximate tensor rank impute missing spatiotemporal traffic data lrtc framework. model traffic data third-order tensor structure (time intervals,locations (sensors),days) introduce four complicated missing patterns, including random missing three fiber-like missing cases according tensor mode-n fibers. despite nonconvexity objective function model, derive global optimal solutions integrating alternating direction method multipliers (admm) generalized soft-thresholding (gst). addition, design truncation rate decay strategy deal varying missing rate scenarios. comprehensive experiments finally conducted using real-world spatiotemporal datasets, demonstrate proposed lrtc-tspn method performs well various missing cases, meanwhile outperforming sota tensor-based imputation models almost scenarios.",18
"fuzzy gene filter: classifier performance assesment. fuzzy gene filter (fgf) optimised fuzzy inference system designed rank genes order differential expression, based expression data generated microarray experiment. paper examines effectiveness fgf feature selection using various classification architectures. fgf compared three common gene ranking algorithms: t-test, wilcoxon test roc curve analysis. four classification schemes used compare performance fgf vis-a-vis standard approaches: k nearest neighbour (knn), support vector machine (svm), naive bayesian classifier (nbc) artificial neural network (ann). nested stratified leave-one-out cross validation scheme used identify optimal number top ranking genes, well optimal classifier parameters. two microarray data sets used comparison: prostate cancer data set lymphoma data set.",3
"backpropagation training fisher vectors within neural networks. fisher-vectors (fv) encode higher-order statistics set multiple local descriptors like sift features. already show good performance combination shallow learning architectures visual recognitions tasks. current methods using fv feature descriptor deep architectures assume original input features static. propose framework jointly learn representation original features, fv parameters parameters classifier style traditional neural networks. proof concept implementation improves performance fv pascal voc 2007 challenge multi-gpu setting comparison default svm setting. demonstrate fv embedded neural networks arbitrary positions, allowing end-to-end training back-propagation.",3
"restricted domain compressive sensing antenna metrology. paper, prove compressive sensing guarantee restricted measurement domains spherical near-field far-field transformations antenna metrology. first defining slepian functions measurement sub-domain $r$ rotation group $\mathrm{so}(3)$, full domain linear inverse problem associated spherical near-field far-field transformations. then, transform inverse problem measurement basis, bounded orthonormal system band-limited wigner $d$-functions $\mathrm{so}(3)$, slepian functions way preserves sparsity. contrasting methods using wigner $d$-functions require measurements $\mathrm{so}(3)$, show orthogonality structure slepian functions requires measurements sub-domain $r$, select-able. due particulars approach inherent presence slepian functions low concentrations $r$, approach gives highest accuracy signal study well concentrated $r$. provide numerical examples method comparison classical compressive sensing approaches. terms reconstruction quality, find method outperforms compressive sensing approaches test least good classical approaches significant reduction number measurements.",5
"biomedical term normalization ehrs umls. paper presents novel prototype biomedical term normalization electronic health record excerpts unified medical language system (umls) metathesaurus. despite multilingual cross-lingual design, first focus processing clinical text spanish existing tool language specific purpose. tool based apache lucene index metathesaurus generate mapping candidates input text. uses ixa pipeline basic language processing resolves ambiguities ukb toolkit. evaluated measuring agreement metamap two english-spanish parallel corpora. addition, present web-based interface tool.",3
"connection probabilities multiple fk-ising interfaces. find scaling limits general class boundary-to-boundary connection probabilities multiple interfaces critical planar fk-ising model, thus verifying predictions physics literature. also discuss conjectural formulas using coulomb-gas integrals corresponding quantities general critical planar random-cluster models cluster-weight $q \in [1,4)$. thus far, proofs convergence, including ours, rely discrete complex analysis techniques beyond reach values $q$ fk-ising model ($q=2$). given convergence interfaces, conjectural formulas values $q$ could verified similarly relatively minor technical work. limit interfaces variants sle$_\kappa$ curves (with $\kappa = 16/3$ $q=2$). partition functions, give connection probabilities, also satisfy properties predicted correlation functions conformal field theory (cft), expected describe scaling limits critical random-cluster models. verify properties $q \in [1,4)$, thus providing evidence expected cft description models.",10
"iterative gaussianization: ica random rotations. signal processing problems involve challenging task multidimensional probability density function (pdf) estimation. work, propose solution problem using family rotation-based iterative gaussianization (rbig) transforms. general framework consists sequential application univariate marginal gaussianization transform followed orthonormal transform. proposed procedure looks differentiable transforms known pdf unknown pdf estimated point original domain. particular, aim zero mean unit covariance gaussian convenience. rbig formally similar classical iterative projection pursuit (pp) algorithms. however, show that, unlike pp methods, particular class rotations used special qualitative relevance context, since looking interestingness critical issue pdf estimation. key difference approach focuses univariate part (marginal gaussianization) problem rather multivariate part (rotation). difference implies one may select convenient rotation suited practical application. differentiability, invertibility convergence rbig theoretically experimentally analyzed. relation methods, radial gaussianization (rg), one-class support vector domain description (svdd), deep neural networks (dnn) also pointed out. practical performance rbig successfully illustrated number multidimensional problems image synthesis, classification, denoising, multi-information estimation.",18
"zipf's law emerges asymptotically phase transitions communicative systems. zipf's law predicts power-law relationship word rank frequency language communication systems, widely reported texts yet remains enigmatic origins. computer simulations shown language communication systems emerge abrupt phase transition fidelity mappings symbols objects. since phase transition approximates heaviside step function, show zipfian scaling emerges asymptotically high rank based laplace transform. thereby demonstrate zipf's law gradually emerges moment phase transition communicative systems. show power-law scaling behavior explains emergence natural languages phase transitions. find emergence zipf's law language communication suggests use rare words lexicon critical construction effective communicative system phase transition.",14
"deep cheap learning work well?. show success deep learning could depend mathematics also physics: although well-known mathematical theorems guarantee neural networks approximate arbitrary functions well, class functions practical interest frequently approximated ""cheap learning"" exponentially fewer parameters generic ones. explore properties frequently encountered physics symmetry, locality, compositionality, polynomial log-probability translate exceptionally simple neural networks. argue statistical process generating data certain hierarchical form prevalent physics machine-learning, deep neural network efficient shallow one. formalize claims using information theory discuss relation renormalization group. prove various ""no-flattening theorems"" showing efficient linear deep networks cannot accurately approximated shallow ones without efficiency loss, example, show $n$ variables cannot multiplied using fewer 2^n neurons single hidden layer.",2
"consistency likelihood maximization vertex nomination scheme: bridging gap maximum likelihood estimation graph matching. given graph vertices deemed interesting priori, vertex nomination task order remaining vertices nomination list concentration interesting vertices top list. previous work yielded several approaches problem, theoretical results setting graph drawn stochastic block model (sbm), including vertex nomination analogue bayes optimal classifier. paper, prove maximum likelihood (ml)-based vertex nomination consistent, sense performance ml-based scheme asymptotically matches bayes optimal scheme. prove theorems form model parameters known unknown. additionally, introduce prove consistency related, scalable restricted-focus ml vertex nomination scheme. finally, incorporate vertex edge features ml-based vertex nomination briefly explore empirical effectiveness approach.",18
"image-level pixel-level labeling convolutional networks. interested inferring object segmentation leveraging object class information, considering minimal priors object segmentation task. problem could viewed kind weakly supervised segmentation task, naturally fits multiple instance learning (mil) framework: every training image known (or not) least one pixel corresponding image class label, segmentation task rewritten inferring pixels belonging class object (given one image, object class). propose convolutional neural network-based model, constrained training put weight pixels important classifying image. show test time, model learned discriminate right pixels well enough, performs well existing segmentation benchmark, adding smoothing priors. system trained using subset imagenet dataset segmentation experiments performed challenging pascal voc dataset (with fine-tuning model pascal voc). model beats state art results weakly supervised object segmentation task large margin. also compare performance model state art fully-supervised segmentation approaches.",3
"abzooba smart health informatics platform (ship) tm - patient experiences big data insights. paper describes technology connect patients information experiences patients using power structured big data. approach, implemented abzooba smart health informatics platform (ship),is distill concepts facts expressions conversations discussions health social media forums, use distilled concepts connecting patients experiences insights highly relevant particular. envision work, progress, provide new effective tools exploit richness content social media health outcomes research.",3
"nonparametric bayesian models hierarchical structure complex networks. analyzing understanding structure complex relational data important many applications including analysis connectivity human brain. networks prominent patterns different scales, calling hierarchically structured model. propose two non-parametric bayesian hierarchical network models based gibbs fragmentation tree priors, demonstrate ability capture nested patterns simulated networks. real networks demonstrate detection hierarchical structure show predictive performance par state art. envision methods employed exploratory analysis large scale complex networks example model human brain connectivity.",18
"fast svm-based feature elimination utilizing data radius, hard-margin, soft-margin. margin maximization hard-margin sense, proposed feature elimination criterion mfe-lo method, combined data radius utilization aim lower generalization error, several published bounds bound-related formulations pertaining lowering misclassification risk (or error) pertain radius e.g. product squared radius weight vector squared norm. additionally, propose additional novel feature elimination criteria that, instead soft-margin sense, utilize data radius, utilizing previously published bound-related formulations approaching radius soft-margin sense, whereby e.g. focus principle stated therein ""finding bound whose minima region small leave-one-out values may important tightness"". additional criteria propose combine radius utilization novel computationally low-cost soft-margin light classifier retraining approach devise named qp1; qp1 soft-margin alternative hard-margin lo. correct error mfe-lo description, find mfe-lo achieves highest generalization accuracy among previously published margin-based feature elimination (mfe) methods, discuss limitations mfe-lo, find novel methods herein outperform mfe-lo, attain lower test set classification error rate. several datasets large number features fall `large features samples' dataset category, datasets lower (low-to-intermediate) number features, novel methods give promising results. especially, among methods tunable ones, employ (the non-tunable) lo approach, tuned aggressively future herein, aim demonstrate even higher performance herein.",18
"bayesian nonparametric models ranked data. develop bayesian nonparametric extension popular plackett-luce choice model handle infinite number choice items. framework based theory random atomic measures, prior specified gamma process. derive posterior characterization simple effective gibbs sampler posterior simulation. develop time-varying extension model, apply new york times lists weekly bestselling books.",18
"deconvolutional latent-variable model text sequence matching. latent-variable model introduced text matching, inferring sentence representations jointly optimizing generative discriminative objectives. alleviate typical optimization challenges latent-variable models text, employ deconvolutional networks sequence decoder (generator), providing learned latent codes semantic information better generalization. model, trained unsupervised manner, yields stronger empirical predictive performance decoder based long short-term memory (lstm), less parameters considerably faster training. further, apply text sequence-matching problems. proposed model significantly outperforms several strong sentence-encoding baselines, especially semi-supervised setting.",3
"monte carlo tree search sampled information relaxation dual bounds. monte carlo tree search (mcts), famously used game-play artificial intelligence (e.g., game go), well-known strategy constructing approximate solutions sequential decision problems. primary innovation use heuristic, known default policy, obtain monte carlo estimates downstream values states decision tree. information used iteratively expand tree towards regions states actions optimal policy might visit. however, guarantee convergence optimal action, mcts requires entire tree expanded asymptotically. paper, propose new technique called primal-dual mcts utilizes sampled information relaxation upper bounds potential actions, creating possibility ""ignoring"" parts tree stem highly suboptimal choices. allows us prove despite converging partial decision tree limit, recommended action primal-dual mcts optimal. new approach shows significant promise used optimize behavior single driver navigating graph operating ride-sharing platform. numerical experiments real dataset 7,000 trips new jersey suggest primal-dual mcts improves upon standard mcts producing deeper decision trees exhibits reduced sensitivity size action space.",10
"pose-invariant face alignment single cnn. face alignment witnessed substantial progress last decade. one recent focuses aligning dense 3d face shape face images large head poses. dominant technology used based cascade regressors, e.g., cnn, shown promising results. nonetheless, cascade cnns suffers several drawbacks, e.g., lack end-to-end training, hand-crafted features slow training speed. address issues, propose new layer, named visualization layer, integrated cnn architecture enables joint optimization different loss functions. extensive evaluation proposed method multiple datasets demonstrates state-of-the-art accuracy, reducing training time half compared typical cascade cnns. addition, compare multiple cnn architectures visualization layer demonstrate advantage utilization.",3
"control gene regulatory networks noisy measurements uncertain inputs. paper concerned problem stochastic control gene regulatory networks (grns) observed indirectly noisy measurements uncertainty intervention inputs. partial observability gene states uncertainty intervention process accounted modeling grns using partially-observed boolean dynamical system (pobds) signal model noisy gene expression measurements. obtaining optimal infinite-horizon control strategy problem attainable general, apply reinforcement learning gaussian process techniques find near-optimal solution. pobds first transformed directly-observed markov decision process continuous belief space, gaussian process used modeling cost function belief intervention spaces. reinforcement learning used learn cost function available gene expression data. addition, employ sparsification, enables control large partially-observed grns. performance resulting algorithm studied comprehensive set numerical experiments using synthetic gene expression data generated melanoma gene regulatory network.",15
"clustering via content-augmented stochastic blockmodels. much data created web contains interactions users items. stochastic blockmodels, methods community detection clustering bipartite graphs, infer latent user communities latent item clusters interaction data. methods, however, typically ignore items' contents information provide item clusters, despite tendency items latent cluster share commonalities content. introduce content-augmented stochastic blockmodels (casb), use item content together user-item interaction data enhance user communities item clusters learned. comparisons several state-of-the-art benchmark methods, datasets arising scientists interacting scientific articles, show content-augmented stochastic blockmodels provide highly accurate clusters respect metrics representative underlying community structure.",18
"copula variational inference. develop general variational inference method preserves dependency among latent variables. method uses copulas augment families distributions used mean-field structured approximations. copulas model dependency captured original variational distribution, thus augmented variational family guarantees better approximations posterior. stochastic optimization, inference augmented distribution scalable. furthermore, strategy generic: applied inference procedure currently uses mean-field structured approach. copula variational inference many advantages: reduces bias; less sensitive local optima; less sensitive hyperparameters; helps characterize interpret dependency among latent variables.",18
"regularization effects anisotropic noise stochastic gradient descent. understanding generalization deep learning raised lots concerns recently, learning algorithms play important role generalization performance, stochastic gradient descent (sgd). along line, particularly study anisotropic noise introduced sgd, investigate importance generalization deep neural networks. thorough empirical analysis, shown anisotropic diffusion sgd tends follow curvature information loss landscape, thus beneficial escaping sharp poor minima effectively, towards stable flat minima. verify understanding comparing anisotropic diffusion full gradient descent plus isotropic diffusion (i.e. langevin dynamics) types position-dependent noise.",18
"gvnn: neural network library geometric computer vision. introduce gvnn, neural network library torch aimed towards bridging gap classic geometric computer vision deep learning. inspired recent success spatial transformer networks, propose several new layers often used parametric transformations data geometric computer vision. layers inserted within neural network much spirit original spatial transformers allow backpropagation enable end-to-end learning network involving domain knowledge geometric computer vision. opens applications learning invariance 3d geometric transformation place recognition, end-to-end visual odometry, depth estimation unsupervised learning warping parametric transformation image reconstruction error.",3
decoding finger movements ecog signals using switching linear models. one major challenges ecog-based brain-machine interfaces movement prediction human subject. several methods exist predict arm 2-d trajectory. fourth bci competition gives dataset aim predict individual finger movements (5-d trajectory). difficulty lies fact simple relation ecog signals finger movement. propose paper decode finger flexions using switching models. method permits simplify system described ensemble linear models depending internal state. show interesting accuracy prediction obtained model.,3
"generative adversarial network-based synthesis visible faces polarimetric thermal faces. large domain discrepancy faces captured polarimetric (or conventional) thermal visible domain makes cross-domain face recognition quite challenging problem human-examiners computer vision algorithms. previous approaches utilize two-step procedure (visible feature estimation visible image reconstruction) synthesize visible image given corresponding polarimetric thermal image. however, regarded two disjoint steps hence may hinder performance visible face reconstruction. argue joint optimization would better way reconstruct photo-realistic images computer vision algorithms human-examiners examine. end, paper proposes generative adversarial network-based visible face synthesis (gan-vfs) method synthesize photo-realistic visible face images corresponding polarimetric images. ensure encoded visible-features contain semantically meaningful information reconstructing visible face image, guidance sub-network involved training procedure. achieve photo realistic property preserving discriminative characteristics reconstructed outputs, identity loss combined perceptual loss optimized framework. multiple experiments evaluated different experimental protocols demonstrate proposed method achieves state-of-the-art performance.",3
"improved aberth-ehrlich root-finding algorithm application binary microlensing. gravitational microlensing formalism modeling binary light curves, key step solving binary lens equation. currently, combination newton's laguerre's methods first introduced skowron \& gould (sg) used modeling binary light curves. paper, first introduce fast root-finding algorithm univariate polynomials based aberth-ehrlich (ae) method first developed 1967 improvement newton's method. ae algorithm proven much faster newton's, laguerre's durand-kerner methods unlike root-finding algorithms, able produce roots simultaneously. improving basic ae algorithm discussing properties, optimize solving binary lens equations, fifth degree polynomials complex coefficients. method $1.8$ $2.0$ times faster sg algorithm. since, calculating magnification factors point-like finite source stars, necessary solve binary lens equation find positions produced images image plane first, new method improve speed accuracy binary microlensing modeling.",0
"sparcnn: spatially related convolutional neural networks. ability accurately detect classify objects varying pixel sizes cluttered scenes crucial many navy applications. however, detection performance existing state-of the-art approaches convolutional neural networks (cnns) degrade suffer applied cluttered multi-object detection tasks. conjecture spatial relationships objects image could exploited significantly improve detection accuracy, approach yet considered existing techniques (to best knowledge) time research conducted. introduce detection classification technique called spatially related detection convolutional neural networks (sparcnn) learns exploits probabilistic representation inter-object spatial configurations within images training sets effective region proposals use state-of-the-art cnns. empirical evaluation sparcnn voc 2007 dataset shows increases classification accuracy 8% compared region proposal technique exploit spatial relations. importantly, obtained higher performance boost 18.8% task difficulty test set increased including highly obscured objects increased image clutter.",3
"painting analysis using wavelets probabilistic topic models. paper, computer-based techniques stylistic analysis paintings applied five panels 14th century peruzzi altarpiece giotto di bondone. features extracted combining dual-tree complex wavelet transform hidden markov tree (hmt) model. hierarchical clustering used identify stylistic keywords image patches, keyword frequencies calculated sub-images contains many patches. generative hierarchical bayesian model learns stylistic patterns keywords; patterns used characterize styles sub-images; turn, permits discriminate paintings. results suggest unsupervised probabilistic topic models useful distill characteristic elements style.",3
"leads clothing fashion: style, color, texture? computational study. well known clothing fashion distinctive often habitual trend style person dresses. clothing fashions usually expressed visual stimuli style, color, texture. however, clear visual stimulus places higher/lower influence updating clothing fashion. study, computer vision machine learning techniques employed analyze influence different visual stimuli clothing-fashion updates. specifically, classification-based model proposed quantify influence different visual stimuli, visual stimulus's influence quantified corresponding accuracy fashion classification. experimental results demonstrate that, clothing-fashion updates, style holds higher influence color, color holds higher influence texture.",3
"monostream: minimal-hardware high accuracy device-free wlan localization system. device-free (df) localization emerging technology allows detection tracking entities carry devices participate actively localization process. typically, df systems require large number transmitters receivers achieve acceptable accuracy, available many scenarios homes small businesses. paper, introduce monostream accurate single-stream df localization system leverages rich channel state information (csi) well mimo information physical layer provide accurate df localization one stream. boost accuracy attain low computational requirements, monostream models df localization problem object recognition problem uses novel set csi-context features techniques proven accuracy efficiency. experimental evaluation two typical testbeds, side-by-side comparison state-of-the-art, shows monostream achieve accuracy 0.95m least 26% enhancement median distance error using single stream only. enhancement accuracy comes efficient execution less 23ms per location update typical laptop. highlights potential monostream usage real-time df tracking applications.",3
"neural network identify individuals health risk. risk diseases heart attack high blood pressure could reduced adequate physical activity. however, even though majority general population claims perform physical exercise, minority exercises enough keep healthy living style. thus, physical inactivity become one major concerns public health past decade. research shows highest decrease physical activity noticed high school college. thus, great importance quickly identify college students health risk due physical inactivity. research also shows level physical activity individual highly correlated demographic features race gender, well self motivation support family friends. information could collected student via 20 minute questionnaire, time needed distribute analyze questionnaire infeasible collegiate campus. thus, propose automatic identifier students risk, students could easier targeted collegiate campuses physical activity promotion departments. present paper preliminary results supervised backpropagation multilayer neural network classifying students at-risk at-risk group.",3
"accelerating potential evaluation unstructured meshes two dimensions. accurate efficient evaluation potentials great importance numerical solution partial differential equations. integration domain potential irregular discretized unstructured mesh, function spaces near field self-interactions non-compact, and, thus, computations cannot easily accelerated. paper, propose three novel complementary techniques accelerating evaluation potentials unstructured meshes. firstly, rigorously characterize geometry near field, show analysis used eliminate unnecessary near field interaction computations. secondly, near field made arbitrarily small increasing order far field quadrature rule, expensive near field interaction computation efficiently offloaded onto fmm-based far field interaction computation, leverages computational efficiency highly optimized parallel fmm libraries. finally, show separate interpolation mesh staggered quadrature mesh dramatically reduces cost constructing interpolants. besides contributions, present robust extensible framework evaluation interpolation 2-d volume potentials complicated geometries. demonstrate effectiveness techniques several numerical experiments.",10
"evaluating las vegas algorithms - pitfalls remedies. stochastic search algorithms among sucessful approaches solving hard combinatorial problems. large class stochastic search approaches cast framework las vegas algorithms (lvas). run-time behavior lvas characterized random variables, detailed knowledge run-time distributions provides important information analysis algorithms. paper propose novel methodology evaluating performance lvas, based identification empirical run-time distributions. exemplify approach applying stochastic local search (sls) algorithms satisfiability problem (sat) propositional logic. point pitfalls arising use improper empirical methods discuss benefits proposed methodology evaluating comparing lvas.",3
"hierarchical cellular automata visual saliency. saliency detection, finding important parts image, become increasingly popular computer vision. paper, introduce hierarchical cellular automata (hca) -- temporally evolving model intelligently detect salient objects. hca consists two main components: single-layer cellular automata (sca) cuboid cellular automata (cca). unsupervised propagation mechanism, single-layer cellular automata exploit intrinsic relevance similar regions interactions neighbors. low-level image features well high-level semantic information extracted deep neural networks incorporated sca measure correlation different image patches. hierarchical deep features, impact factor matrix coherence matrix constructed balance influences cell's next state. saliency values cells iteratively updated according well-defined update rule. furthermore, propose cca integrate multiple saliency maps generated sca different scales bayesian framework. therefore, single-layer propagation multi-layer integration jointly modeled unified hca. surprisingly, find sca improve existing methods applied to, resulting similar precision level regardless original results. cca act efficient pixel-wise aggregation algorithm integrate state-of-the-art methods, resulting even better results. extensive experiments four challenging datasets demonstrate proposed algorithm outperforms state-of-the-art conventional methods competitive deep learning based approaches.",3
"bandits experts metric spaces. multi-armed bandit problem, online algorithm chooses set strategies sequence trials maximize total payoff chosen strategies. performance bandit algorithms small finite strategy set quite well understood, bandit problems large strategy sets still topic active investigation, motivated practical applications online auctions web advertisement. goal research identify broad natural classes strategy sets payoff functions enable design efficient solutions. work study general setting multi-armed bandit problem strategies form metric space, payoff function satisfies lipschitz condition respect metric. refer problem ""lipschitz mab problem"". present solution multi-armed bandit problem setting. is, every metric space define isometry invariant bounds performance lipschitz mab algorithms metric space, present algorithm comes arbitrarily close meeting bound. furthermore, technique gives even better results benign payoff functions. also address full-feedback (""best expert"") version problem, every round payoffs arms revealed.",3
"motion compensated dynamic mri reconstruction local affine optical flow estimation. paper proposes novel framework reconstruct dynamic magnetic resonance images (dmri) motion compensation (mc). due inherent motion effects dmri acquisition, reconstruction dmri using motion estimation/compensation (me/mc) studied compressed sensing (cs) scheme. paper, embedding intensity-based optical flow (of) constraint traditional cs scheme, able couple dmri reconstruction motion field estimation. formulated optimization problem solved primal-dual algorithm linesearch due efficiency dealing non-differentiable problems. estimated motion field, dmri reconstruction refined mc. employing multi-scale coarse-to-fine strategy, able update variables(temporal image sequences motion vectors) refine image reconstruction alternately. moreover, proposed framework capable handling wide class prior information (regularizations) dmri reconstruction, sparsity, low rank total variation. experiments various dmri data, ranging vivo lung cardiac dataset, validate reconstruction quality improvement using proposed scheme comparison several state-of-the-art algorithms.",3
"knowledge projection deep neural networks. deeper wider neural networks actively pushing performance limits various computer vision machine learning tasks, often require large sets labeled data effective training suffer extremely high computational complexity. paper, develop new framework training deep neural networks datasets limited labeled samples using cross-network knowledge projection able improve network performance reducing overall computational complexity significantly. specifically, large pre-trained teacher network used observe samples training data. projection matrix learned project teacher-level knowledge visual representations intermediate layer teacher network intermediate layer thinner faster student network guide regulate training process. intermediate layers teacher network injection layers student network adaptively selected training evaluating joint loss function iterative manner. knowledge projection framework allows us use crucial knowledge learned large networks guide training thinner student networks, avoiding over-fitting, achieving better network performance, significantly reducing complexity. extensive experimental results benchmark datasets demonstrated proposed knowledge projection approach outperforms existing methods, improving accuracy 4% reducing network complexity 4 10 times, attractive practical applications deep neural networks.",3
"translucent players: explaining cooperative behavior social dilemmas. last decades, numerous experiments shown humans always behave maximize material payoff. cooperative behavior non-cooperation dominant strategy (with respect material payoffs) particularly puzzling. propose novel approach explain cooperation, assuming halpern pass call translucent players. typically, players assumed opaque, sense deviation one player normal-form game affect strategies used players. player may believe switches one strategy another, fact chooses switch may visible players. example, chooses defect prisoner's dilemma, player may sense guilt. show assuming translucent players, recover many regularities observed human behavior well-studied games prisoner's dilemma, traveler's dilemma, bertrand competition, public goods game.",3
"objective-reinforced generative adversarial networks (organ) sequence generation models. unsupervised data generation tasks, besides generation sample based previous observations, one would often like give hints model order bias generation towards desirable metrics. propose method combines generative adversarial networks (gans) reinforcement learning (rl) order accomplish exactly that. rl biases data generation process towards arbitrary metrics, gan component reward function ensures model still remembers information learned data. build upon previous results incorporated gans rl order generate sequence data test model several settings generation molecules encoded text sequences (smiles) context music generation, showing case effectively bias generation process towards desired metrics.",18
"machine learning deuteron: new architectures uncertainty quantification. solve ground state deuteron using variational neural network ansatz wave function momentum space. ansatz provides flexible representation $s$ $d$ states, relative errors energy within fractions percent full diagonalisation benchmark. extend previous work area two directions. first, study new architectures adding layers network exploring different connections states. second, provide better estimate numerical uncertainty taking account final oscillations end minimisation process. overall, find best performing architecture simple one-layer, state-independent network. two-layer networks show indications overfitting, regions probed fixed momentum basis calculations performed. cases, error associated model oscillations around real minimum larger stochastic initialisation uncertainties. conclusions draw generalised quantum mechanics settings.",13
"learning long term dependencies via fourier recurrent units. known fact training recurrent neural networks tasks long term dependencies challenging. one main reasons vanishing exploding gradient problem, prevents gradient information propagating early layers. paper propose simple recurrent architecture, fourier recurrent unit (fru), stabilizes gradients arise training giving us stronger expressive power. specifically, fru summarizes hidden states $h^{(t)}$ along temporal dimension fourier basis functions. allows gradients easily reach layer due fru's residual learning structure global support trigonometric functions. show fru gradient lower upper bounds independent temporal dimension. also show strong expressivity sparse fourier basis, fru obtains strong expressive power. experimental study also demonstrates fewer parameters proposed architecture outperforms recurrent architectures many tasks.",3
"global sensitivity analysis dependence measures. global sensitivity analysis variance-based measures suffers several theoretical practical limitations, since focus variance output handle multivariate variables limited way. paper, introduce new class sensitivity indices based dependence measures overcomes insufficiencies. approach originates idea compare output distribution conditional counterpart one input variables fixed. establish comparison yields previously proposed indices performed csiszar f-divergences, well sensitivity indices well-known dependence measures random variables. leads us investigate completely new sensitivity indices based recent state-of-the-art dependence measures, distance correlation hilbert-schmidt independence criterion. also emphasize potential feature selection techniques relying dependence measures alternatives screening high dimension.",10
"loss functions top-k error: analysis insights. order push performance realistic computer vision tasks, number classes modern benchmark datasets significantly increased recent years. increase number classes comes along increased ambiguity class labels, raising question top-1 error right performance measure. paper, provide extensive comparison evaluation established multiclass methods comparing top-k performance practical well theoretical perspective. moreover, introduce novel top-k loss functions modifications softmax multiclass svm losses provide efficient optimization schemes them. experiments, compare various datasets proposed established methods top-k error optimization. interesting insight paper softmax loss yields competitive top-k performance k simultaneously. specific top-k error, new top-k losses lead typically improvements faster train softmax.",18
"efficient convolutional neural network audio event detection. wireless distributed systems used sensor networks, internet-of-things cyber-physical systems, impose high requirements resource efficiency. advanced preprocessing classification data network edge help decrease communication demand reduce amount data processed centrally. area distributed acoustic sensing, combination algorithms high classification rate resource-constraint embedded systems essential. unfortunately, algorithms acoustic event detection high memory computational demand suited execution network edge. paper addresses aspects applying structural optimizations convolutional neural network audio event detection reduce memory requirement factor 500 computational effort factor 2.1 performing 9.2% better.",3
"predictive modelling bone age classification regression bone shapes. bone age assessment task performed daily hospitals worldwide. involves clinician estimating age patient radiograph non-dominant hand. approach automated bone age assessment modularise algorithm following three stages: segment verify hand outline; segment verify bones; use bone outlines construct models age. paper address final question: given outlines bones, learn predict bone age patient? examine two alternative approaches. firstly, attempt train classifiers individual bones predict bone stage categories commonly used bone ageing. secondly, construct regression models directly predict patient age. demonstrate models built summary features bone outline perform better built using one dimensional representation outline, also least well automated systems. show models constructed three bones accurate predicting age expert human assessors using standard technique. also demonstrate utility model quantifying importance ethnicity sex age development. conclusion feature based system separating image processing age modelling best approach automated bone ageing, since offers flexibility transparency produces accurate estimates.",3
"positivity transportation. prove paper weighted volume set integral transportation matrices two integral histograms r c equal sum positive definite kernel r c set considered weights forms positive definite matrix. computation quantity, despite subject significant research effort algebraic statistics, remains intractable challenge histograms even modest dimensions. propose alternative kernel which, rather considering matrices transportation polytope, focuses sub-sample vertices known northwestern corner solutions. resulting kernel positive definite computed number operations o(r^2d) grows linearly complexity dimension d, r^2, total amount sampled vertices, parameter controls complexity kernel.",18
"improving facial attribute prediction using semantic segmentation. attributes semantically meaningful characteristics whose applicability widely crosses category boundaries. particularly important describing recognizing concepts explicit training example given, \textit{e.g., zero-shot learning}. additionally, since attributes human describable, used efficient human-computer interaction. paper, propose employ semantic segmentation improve facial attribute prediction. core idea lies fact many facial attributes describe local properties. words, probability attribute appear face image far uniform spatial domain. build facial attribute prediction model jointly deep semantic segmentation network. harnesses localization cues learned semantic segmentation guide attention attribute prediction regions different attributes naturally show up. result approach, addition recognition, able localize attributes, despite merely access image level labels (weak supervision) training. evaluate proposed method celeba lfwa datasets achieve superior results prior arts. furthermore, show reverse problem, semantic face parsing improves facial attributes available. reaffirms need jointly model two interconnected tasks.",3
"prototype matching networks large-scale multi-label genomic sequence classification. one fundamental tasks understanding genomics problem predicting transcription factor binding sites (tfbss). hundreds transcription factors (tfs) labels, genomic-sequence based tfbs prediction challenging multi-label classification task. two major biological mechanisms tf binding: (1) sequence-specific binding patterns genomes known ""motifs"" (2) interactions among tfs known co-binding effects. paper, propose novel deep architecture, prototype matching network (pmn) mimic tf binding mechanisms. pmn model automatically extracts prototypes (""motif""-like features) tf novel prototype-matching loss. borrowing ideas few-shot matching models, use notion support set prototypes lstm learn tfs interact bind genomic sequences. reference tfbs dataset $2.1$ $million$ genomic sequences, pmn significantly outperforms baselines validates design choices empirically. knowledge, first deep learning architecture introduces prototype learning considers tf-tf interactions large-scale tfbs prediction. proposed architecture accurate, also models underlying biology.",3
"efficient two-stage sparse representation method. large number methods solving under-determined linear inverse problem. many high time complexity large datasets. propose new method called two-stage sparse representation (tssr) tackle problem. decompose representing space signals two parts, measurement dictionary sparsifying basis. dictionary designed approximate sub-gaussian distribution exploit concentration property. apply sparse coding signals dictionary first stage, obtain training testing coefficients respectively. design basis approach identity matrix second stage, acquire restricted isometry property (rip) universality property. testing coefficients encoded basis final representing coefficients obtained. verify projection testing coefficients onto basis good approximation signal onto representing space. since projection conducted much sparser space, runtime greatly reduced. concrete realization, provide instance proposed tssr. experiments four biometrics databases show tssr effective efficient, comparing several classical methods solving linear inverse problem.",3
"3d shape classification using collaborative representation based projections. novel 3d shape classification scheme, based collaborative representation learning, investigated work. data-driven feature-extraction procedure, taking form simple projection operator, core methodology. provided shape database, graph encapsulating structural relationships among available shapes, first constructed employed defining low-dimensional sparse projections. recently introduced method crps (collaborative representation based projections), based l2-graph, first variant included towards end. second algorithm, particularizes crps shape descriptors inherently nonnegative, also introduced potential alternative. cases, weights graph reflecting database structure calculated approximate shape sparse linear combination remaining dataset objects. way solving generalized eigenanalysis problem, linear matrix operator designed act feature extractor. two popular, inherently high dimensional descriptors, namely shapedna global point signature (gps), employed experimentations shrec10, shrec11 schrec 15 datasets, shape recognition cast multi-class classification problem tackled means svm (support vector machine) acting within reduced dimensional space crafted projections. results promising outperform state art methods, providing evidence highly discriminative nature introduced 3d shape representations.",3
"prosody-based automatic segmentation speech sentences topics. crucial step processing speech audio data information extraction, topic detection, browsing/playback segment input sentence topic units. speech segmentation challenging, since cues typically present segmenting text (headers, paragraphs, punctuation) absent spoken language. investigate use prosody (information gleaned timing melody speech) tasks. using decision tree hidden markov modeling techniques, combine prosodic cues word-based approaches, evaluate performance two speech corpora, broadcast news switchboard. results show prosodic model alone performs par with, better than, word-based statistical language models -- true automatically recognized words news speech. prosodic model achieves comparable performance significantly less training data, requires hand-labeling prosodic events. across tasks corpora, obtain significant improvement word-only models using probabilistic combination prosodic lexical information. inspection reveals prosodic models capture language-independent boundary indicators described literature. finally, cue usage task corpus dependent. example, pause pitch features highly informative segmenting news speech, whereas pause, duration word-based cues dominate natural conversation.",3
"investigating nature properties maxi j1810-222 radio x-ray observations. present results radio x-ray observations x-ray transient maxi j1810-222. nature accretor source identified. paper, show results quasi-simultaneous radio x-ray monitoring campaign taken australia telescope compact array (atca), neil gehrels swift observatory x-ray telescope (xrt), swift burst alert telescope (bat). also analyse x-ray temporal behaviour using observations neutron star interior composition explorer (nicer). results show seemingly peculiar x-ray spectral evolution maxi j1810-222 outburst, source initially detected soft x-ray band early part outburst. then, ~200 days maxi j1810-222 first detected hard x-ray emission increased source transitioned long-lived (~1.5 years) bright, harder x-ray state. hard state, maxi j1810-222 returned back softer state, fading transitioning harder state appearing follow typical outburst decay. x-ray spectral timing properties, source's radio behaviour, argue results study consistent maxi j1810-222 relatively distant ($\gtrsim$6 kpc) black hole x-ray binary. sufficiently large distance source simply explain seemingly odd outburst evolution observed, brightest portion outburst detectable all-sky x-ray telescopes.",0
"credal model averaging classification: representing prior ignorance expert opinions. bayesian model averaging (bma) state art approach overcoming model uncertainty. yet, especially small data sets, results yielded bma might sensitive prior models. credal model averaging (cma) addresses problem substituting single prior models set priors (credal set). approach solves problem choose prior models automates sensitivity analysis. discuss various cma algorithms building ensemble logistic regressors characterized different sets covariates. show cma appropriately tuned case one prior-ignorant case instead domain knowledge available. cma detects prior-dependent instances, namely instances different class probable depending prior models. instances cma suspends judgment, returning multiple classes. thoroughly compare different bma cma variants real case study, predicting presence alpine marmot burrows alpine valley. find bma almost random guesser instances recognized prior-dependent cma.",18
"unifying multi-view self-representations clustering tensor multi-rank minimization. paper, address multi-view subspace clustering problem. method utilizes circulant algebra tensor, constructed stacking subspace representation matrices different views rotating, capture low rank tensor subspace refinement view-specific subspaces achieved, well high order correlations underlying multi-view data explored.} introducing recently proposed tensor factorization, namely tensor-singular value decomposition (t-svd) \cite{kilmer13}, impose new type low-rank tensor constraint rotated tensor capture complementary information multiple views. different traditional unfolding based tensor norm, low-rank tensor constraint optimality properties similar matrix rank derived svd, complementary information among views explored efficiently thoroughly. established model, called t-svd based multi-view subspace clustering (t-svd-msc), falls applicable scope augmented lagrangian method, minimization problem efficiently solved theoretical convergence guarantee relatively low computational complexity. extensive experimental testing eight challenging image dataset shows proposed method achieved highly competent objective performance compared several state-of-the-art multi-view clustering methods.",3
"robust, deep inductive anomaly detection. pca classical statistical technique whose simplicity maturity seen find widespread use anomaly detection technique. however, limited regard sensitive gross perturbations input, seeking linear subspace captures normal behaviour. first issue dealt robust pca, variant pca explicitly allows data points arbitrarily corrupted, however, resolve second issue, indeed introduces new issue one longer inductively find anomalies test set. paper addresses issues single model, robust autoencoder. method learns nonlinear subspace captures majority data points, allowing data arbitrary corruption. model simple train leverages recent advances optimisation deep neural networks. experiments range real-world datasets highlight model's effectiveness.",3
"exact efficient parallel inference nonparametric mixture models. nonparametric mixture models based dirichlet process elegant alternative finite models number underlying components unknown, inference models slow. existing attempts parallelize inference models relied introducing approximations, lead inaccuracies posterior estimate. paper, describe auxiliary variable representations dirichlet process hierarchical dirichlet process allow us sample true posterior distributed manner. show approach allows scalable inference without deterioration estimate quality accompanies existing methods.",18
"batch reinforcement learning industrial benchmark: first experiences. particle swarm optimization policy (pso-p) recently introduced proven produce remarkable results interacting academic reinforcement learning benchmarks off-policy, batch-based setting. investigate properties feasibility real-world applications, paper investigates pso-p so-called industrial benchmark (ib), novel reinforcement learning (rl) benchmark aims realistic including variety aspects found industrial applications, like continuous state action spaces, high dimensional, partially observable state space, delayed effects, complex stochasticity. experimental results pso-p ib compared results closed-form control policies derived model-based recurrent control neural network (rcnn) model-free neural fitted q-iteration (nfq). experiments show pso-p interest academic benchmarks, also real-world industrial applications, since also yielded best performing policy ib setting. compared well established rl techniques, pso-p produced outstanding results performance robustness, requiring relatively low amount effort finding adequate parameters making complex design decisions.",3
"edge attention-based multi-relational graph convolutional networks. graph convolutional network (gcn) generalization convolutional neural network (cnn) work arbitrarily structured graphs. binary adjacency matrix commonly used training gcn. recently, attention mechanism allows network learn dynamic adaptive aggregation neighborhood. propose new gcn model graphs edges characterized multiple views precisely terms multiple relationships. instance, chemical graph theory, compound structures often represented hydrogen-depleted molecular graph nodes correspond atoms edges correspond chemical bonds. multiple attributes important characterize chemical bonds, atom pair (the types atoms bond connects), aromaticity, whether bond ring. different attributes lead different graph representations molecule. growing interests chemistry machine learning fields directly learn molecular properties compounds molecular graph, instead fingerprints predefined chemists. proposed gcn model, call edge attention-based multi-relational gcn (eagcn), jointly learns attention weights node features graph convolution. bond attribute, real-valued attention matrix used replace binary adjacency matrix. designing dictionary edge attention, forming attention matrix molecule looking dictionary, eagcn exploits correspondence bonds different molecules. prediction compound properties based aggregated node features, independent varying molecule (graph) size. demonstrate efficacy eagcn multiple chemical datasets: tox21, hiv, freesolv, lipophilicity, interpret resultant attention weights.",18
"quantum clones inside black holes. systematic procedure proposed better understanding evolution laws black holes terms pure quantum states. start two opposed regions ii penrose diagram, study evolution matter regions, using algebra derived earlier shapiro effect quantum particles. since spacetime two distinct asymptotic regions, one must assume mechanism reduces number states. earlier work proposed region ii describes angular antipodes region , 'antipodal identification', eventually leads contradictions. much simpler proposal states defined region ii exact quantum clones region . make physical states observable, but, surprisingly, suffices restore unitarity well.",6
"minimax lower bounds ridge combinations including neural nets. estimation functions $ $ variables considered using ridge combinations form $ \textstyle\sum_{k=1}^m c_{1,k} \phi(\textstyle\sum_{j=1}^d c_{0,j,k}x_j-b_k) $ activation function $ \phi $ function bounded value derivative. include single-hidden layer neural networks, polynomials, sinusoidal models. sample size $ n $ possibly noisy values random sites $ x \in b = [-1,1]^d $, minimax mean square error examined functions closure $ \ell_1 $ hull ridge functions activation $ \phi $. shown order $ d/n $ fractional power (when $ $ smaller order $ n $), order $ (\log d)/n $ fractional power (when $ $ larger order $ n $). dependence constraints $ v_0 $ $ v_1 $ $ \ell_1 $ norms inner parameter $ c_0 $ outer parameter $ c_1 $, respectively, also examined. also, lower upper bounds fractional power given. heart analysis development information-theoretic packing numbers classes functions.",18
"beating perils non-convexity: guaranteed training neural networks using tensor methods. training neural networks challenging non-convex optimization problem, backpropagation gradient descent get stuck spurious local optima. propose novel algorithm based tensor decomposition guaranteed training two-layer neural networks. provide risk bounds proposed method, polynomial sample complexity relevant parameters, input dimension number neurons. learning arbitrary target functions np-hard, provide transparent conditions function input learnability. training method based tensor decomposition, provably converges global optimum, set mild non-degeneracy conditions. consists simple embarrassingly parallel linear multi-linear operations, competitive standard stochastic gradient descent (sgd), terms computational complexity. thus, propose computationally efficient method guaranteed risk bounds training neural networks one hidden layer.",3
"optimality matched-pair designs randomized controlled trials. randomized controlled trials (rcts), treatment often assigned stratified randomization. show among stratified randomization schemes treat units probability one half, certain matched-pair design achieves maximum statistical precision estimating average treatment effect (ate). important special case, optimal design pairs units according baseline outcome. simulation study based datasets 10 rcts, design lowers standard error estimator ate 10% average, 34%, relative original designs.",4
"gradient layer: enhancing convergence adversarial training generative models. propose new technique boosts convergence training generative adversarial networks. generally, rate training deep models reduces severely multiple iterations. key reason phenomenon deep network expressed using highly non-convex finite-dimensional model, thus parameter gets stuck local optimum. this, methods often suffer degeneration convergence speed also limitations representational power trained network. overcome issue, propose additional layer called gradient layer seek descent direction infinite-dimensional space. layer constructed infinite-dimensional space, restricted specific model structure finite-dimensional models. result, get local optima finite-dimensional models move towards global optimal function directly. paper, phenomenon explained functional gradient method perspective gradient layer. interestingly, optimization procedure using gradient layer naturally constructs deep structure network. moreover, demonstrate procedure regarded discretization method gradient flow naturally reduces objective function. finally, method tested using several numerical experiments, show fast convergence.",18
"multiscale fractal descriptors applied texture classification. work proposes combination multiscale transform fractal descriptors employed classification gray-level texture images. apply space-scale transform (derivative + gaussian filter) bouligand-minkowski fractal descriptors, followed threshold filter response, aiming attenuating noise effects caused final part response. method tested classification well-known data set (brodatz) compared classical texture descriptor techniques. results demonstrate advantage proposed approach, achieving higher success rate reduced amount descriptors.",3
"quantum network tomography multi-party state distribution. fragile nature quantum information makes practically impossible completely isolate quantum state noise quantum channel transmissions. quantum networks complex systems formed interconnection quantum processing devices quantum channels. context, characterizing channels introduce noise transmitted quantum states paramount importance. precise descriptions error distributions introduced non-unitary quantum channels inform quantum error correction protocols tailor operations particular error model. addition, characterizing errors monitoring network end-to-end measurements enables end-nodes infer status network links. work, address end-to-end characterization quantum channels quantum network introducing problem quantum network tomography. solution problem estimator probabilities define kraus decomposition quantum channels network, using measurements performed exclusively end-nodes. study problem detail case arbitrary star quantum networks quantum channels described single pauli operator, like bit-flip quantum channels. provide solutions networks polynomial sample complexity. solutions provide evidence pre-shared entanglement brings advantages estimation terms identifiability parameters.",17
"human smart machine co-learning brain computer interface. machine learning become popular approach cybernetics systems, always considered important research computational intelligence area. nevertheless, comes smart machines, methodologies. need consider systems cybernetics well include human loop. purpose article follows: (1) integrate open source facebook ai research (fair) darkforest program facebook item response theory (irt), new open learning system, namely, ddf learning system; (2) integrate ddf go robot namely robotic ddf go system; (3) invite professional go players attend activity play go games site smart machine. research team apply technology education, as, playing games enhance children concentration learning mathematics, languages, topics. detected brainwaves, robot able speak words much point students assist teachers classroom future.",3
"defining relative likelihood partially-ordered preferential structures. starting likelihood preference order worlds, extend likelihood ordering sets worlds natural way, examine resulting logic. lewis (1973) earlier considered notion relative likelihood context studying counterfactuals, assumed total preference order worlds. complications arise examining partial orders present total orders. subtleties involving exact approach lifting order worlds order sets worlds. addition, axiomatization logic relative likelihood case partial orders gives insight connection relative likelihood default reasoning.",3
"accelerated reinforcement learning algorithms nonparametric function approximation opportunistic spectrum access. study problem throughput maximization predicting spectrum opportunities using reinforcement learning. kernel-based reinforcement learning approach coupled sparsification technique efficiently captures environment states control dimensionality finds best possible channel access actions based current state. approach allows learning planning intrinsic state-action space extends well large state action spaces. stationary markov environments, derive optimal policy channel access, associated limiting throughput, propose fast online algorithm achieving optimal throughput. show maximum-likelihood channel prediction access algorithm suboptimal general, derive conditions two algorithms equivalent. reactive markov environments, derive kernel variants q-learning, r-learning propose accelerated r-learning algorithm achieves faster convergence. finally test algorithms generic reactive network. simulation results shown validate theory show performance gains current state-of-the-art techniques.",3
"combinatorial topic models using small-variance asymptotics. topic models emerged fundamental tools unsupervised machine learning. modern topic modeling algorithms take probabilistic view derive inference algorithms based latent dirichlet allocation (lda) variants. contrast, study topic modeling combinatorial optimization problem, propose new objective function derived lda passing small-variance limit. minimize derived objective using ideas combinatorial optimization, results new, fast, high-quality topic modeling algorithm. particular, show results competitive popular lda-based topic modeling approaches, also discuss (dis)similarities approach probabilistic counterparts.",3
"future neural networks. paper describes recent developments neural networks discusses applicability neural networks development machine mimics human brain. paper mentions new architecture, pulsed neural network considered next generation neural networks. paper also explores use memristors development brain-like computer called moneta. new model, multi/infinite dimensional neural networks, recent development area advanced neural networks. paper concludes need neural networks development human-like technology essential may non-expendable it.",3
"gpflowopt: bayesian optimization library using tensorflow. novel python framework bayesian optimization known gpflowopt introduced. package based popular gpflow library gaussian processes, leveraging benefits tensorflow including automatic differentiation, parallelization gpu computations bayesian optimization. design goals focus framework easy extend custom acquisition functions models. framework thoroughly tested well documented, provides scalability. current released version gpflowopt includes standard single-objective acquisition functions, state-of-the-art max-value entropy search, well bayesian multi-objective approach. finally, permits easy use custom modeling strategies implemented gpflow.",18
"net-trim: convex pruning deep neural networks performance guarantee. introduce analyze new technique model reduction deep neural networks. large networks theoretically capable learning arbitrarily complex models, overfitting model redundancy negatively affects prediction accuracy model variance. net-trim algorithm prunes (sparsifies) trained network layer-wise, removing connections layer solving convex optimization program. program seeks sparse set weights layer keeps layer inputs outputs consistent originally trained model. algorithms associated analysis applicable neural networks operating rectified linear unit (relu) nonlinear activation. present parallel cascade versions algorithm. latter achieve slightly simpler models generalization performance, former computed distributed manner. cases, net-trim significantly reduces number connections network, also providing enough regularization slightly reduce generalization error. also provide mathematical analysis consistency initial network retrained model. analyze model sample complexity, derive general sufficient conditions recovery sparse transform matrix. single layer taking independent gaussian random vectors length $n$ inputs, show network response described using maximum number $s$ non-zero weights per node, weights learned $\mathcal{o}(s\log n)$ samples.",3
"classification autism spectrum disorder using supervised learning brain connectivity measures extracted synchrostates. objective. paper investigates presence autism using functional brain connectivity measures derived electro-encephalogram (eeg) children face perception tasks. approach. phase synchronized patterns 128-channel eeg signals obtained typical children children autism spectrum disorder (asd). phase synchronized states synchrostates temporally switch amongst underlying process completion particular cognitive task. used 12 subjects group (asd typical) analyzing eeg processing fearful, happy neutral faces. minimal maximally occurring synchrostates subject chosen extraction brain connectivity features, used classification two groups subjects. among different supervised learning techniques, explored discriminant analysis support vector machine polynomial kernels classification task. main results. leave one cross-validation classification algorithm gives 94.7% accuracy best performance corresponding sensitivity specificity values 85.7% 100% respectively. significance. proposed method gives high classification accuracies outperforms contemporary research results. effectiveness proposed method classification autistic typical children suggests possibility using larger population validate clinical practice.",14
"possible triply heavy tetraquark states. present work, triply heavy tetraquarks state $(qq\bar{q}\bar{q}, q=c, b, q=u, d, s)$ possible quantum numbers systematically investigated framework chiral quark model resonance ground method. two structures, including meson-meson configuration (the color-single channels hidden-color channels) diquark-antidiquark configuration (the color sextet-antisextet color triplet-antitriplet), considered. studied mass region, several bound states obtained $cc\bar{c}\bar{q^{'}}$, $bb\bar{c}\bar{q^{'}}$ $bc\bar{c}\bar{q}$ tetraquarks, respectively. results, important feature found channel coupling calculation great significance whether tetraquarks form bound states. predicted bound states, expect provide effective assistance experimental search exotic states.",8
"local nearest neighbour classification applications semi-supervised learning. derive new asymptotic expansion global excess risk local $k$-nearest neighbour classifier, choice $k$ may depend upon test point. expansion elucidates conditions dominant contribution excess risk comes locus points class label equally likely occur, also show conditions satisfied, dominant contribution may arise tails marginal distribution features. moreover, prove that, provided $d$-dimensional marginal distribution features finite $\rho$th moment $\rho > 4$ (as well regularity conditions), local choice $k$ yield rate convergence excess risk $o(n^{-4/(d+4)})$, $n$ sample size, whereas standard $k$-nearest neighbour classifier, theory would require $d \geq 5$ $\rho > 4d/(d-4)$ finite moments achieve rate. results motivate new $k$-nearest neighbour classifier semi-supervised learning problems, unlabelled data used obtain estimate marginal feature density, fewer neighbours used classification density estimate small. potential improvements standard $k$-nearest neighbour classifier illustrated theory via simulation study.",10
"use neural networks analysis sleep stages diagnosis narcolepsy. used neural networks ~3,000 sleep recordings 10 locations automate sleep stage scoring, producing probability distribution called hypnodensity graph. accuracy validated 70 subjects scored six technicians (gold standard). best model performed better individual scorer, reaching accuracy 0.87 (and 0.95 predictions weighed scorer agreement). also scores sleep stages 5-second instead conventional 30-second scoring-epochs. accuracy vary sleep disorder except narcolepsy, suggesting scoring difficulties machine and/or humans. narcolepsy biomarker extracted validated 105 type-1 narcoleptics versus 331 controls producing specificity 0.96 sensitivity 0.91. similar performances obtained high pretest probability sample type-2 narcolepsy idiopathic hypersomnia patients. addition hla-dqb1*06:02 increased specificity 0.99. method streamlines scoring diagnoses narcolepsy accurately.",3
"transverse momentum structure proton within basis light-front quantization framework. obtain leading-twist valence quark transverse-momentum-dependent parton distribution functions (tmd pdfs) proton within basis light-front quantization (blfq) framework. results consistent lattice qcd calculations previous results collinear limit. also obtain consistency soffer-type bounds. within approach, find six t-even tmds leading twist independent other, previously found model-dependent relations hold. promising sign results representative future, extensive treatments qcd. furthermore, obtain non-trivial $ x $-dependence $ \left<(p^\perp)^2\right> $ consistency gaussian ansatz small $ (p^\perp)^2 $ region. features suggest results may useful alternative future experimental extractions.",8
"comparison among dimensionality reduction techniques based random projection cancer classification. random projection (rp) technique widely applied many scenarios reduce high-dimensional features low-dimensional space within short time meet need real-time analysis massive data. urgent need dimensionality reduction fast increase big genomics data. however, performance rp usually lower. attempt improve classification accuracy rp combining reduction dimension methods principle component analysis (pca), linear discriminant analysis (lda), feature selection (fs). compared classification accuracy running time different combination methods three microarray datasets simulation dataset. experimental results show remarkable improvement 14.77% classification accuracy fs followed rp compared rp bc-tcga dataset. lda followed rp also helps rp yield discriminative subspace increase 13.65% classification accuracy dataset. fs followed rp outperforms combination methods classification accuracy datasets.",3
differentially private identity closeness testing discrete distributions. investigate problems identity closeness testing discrete population random samples. goal develop efficient testers guaranteeing differential privacy individuals population. describe approach yields sample-efficient differentially private testers problems. theoretical results show exist private identity closeness testers nearly sample-efficient non-private counterparts. perform experimental evaluation algorithms synthetic data. experiments illustrate private testers achieve small type type ii errors sample size sublinear domain size underlying distributions.,3
"multi-information source optimization. consider bayesian optimization expensive-to-evaluate black-box objective function, also access cheaper approximations objective. general, approximations arise applications reinforcement learning, engineering, natural sciences, subject inherent, unknown bias. model discrepancy caused inadequate internal model deviates reality vary domain, making utilization approximations non-trivial task. present novel algorithm provides rigorous mathematical treatment uncertainties arising model discrepancies noisy observations. optimization decisions rely value information analysis extends knowledge gradient factor setting multiple information sources vary cost: sampling decision maximizes predicted benefit per unit cost. conduct experimental evaluation demonstrates method consistently outperforms state-of-the-art techniques: finds designs considerably higher objective value additionally inflicts less cost exploration process.",18
"regression trees random forest based feature selection malaria risk exposure prediction. paper deals prediction anopheles number, main vector malaria risk, using environmental climate variables. variables selection based automatic machine learning method using regression trees, random forests combined stratified two levels cross validation. minimum threshold variables importance accessed using quadratic distance variables importance optimal subset selected variables used perform predictions. finally results revealed qualitatively better, selection, prediction , cpu time point view obtained glm-lasso method.",18
"almost-circular symplectic induced ginibre ensemble. consider symplectic induced ginibre process, pfaffian point process plane. let $n$ number points. focus almost-circular regime points lie thin annulus $\mathcal{s}_{n}$ width $o(\frac{1}{n})$ $n \to \infty$. main results scaling limits correlation functions near real axis, also away real axis. near real axis, limiting correlation functions pfaffians new correlation kernel, interpolates limiting kernels bulk symplectic ginibre ensemble anti-symmetric gaussian hermitian ensemble odd size. away real axis, limiting correlation functions determinants, kernel one appearing bulk limit almost-hermitian random matrices. furthermore, obtain precise large $n$ asymptotics probability points lie outside $\mathcal{s}_{n}$, well several ""semi-large"" gap probabilities.",11
"robust learning bayesian networks prior belief. recent reports described learning bayesian networks highly sensitive chosen equivalent sample size (ess) bayesian dirichlet equivalence uniform (bdeu). sensitivity often engenders unstable undesirable results. paper describes asymptotic analyses bdeu explain reasons sensitivity effects. furthermore, paper presents proposal robust learning score ess eliminating sensitive factors approximation log-bdeu.",3
"efficient classification metric data. recent advances large-margin classification data residing general metric spaces (rather hilbert spaces) enable classification various natural metrics, string edit earthmover distance. general framework developed purpose von luxburg bousquet [jmlr, 2004] left open questions computational efficiency providing direct bounds generalization error. design new algorithm classification general metric spaces, whose runtime accuracy depend doubling dimension data points, thus achieve superior classification performance many common scenarios. algorithmic core approach approximate (rather exact) solution classical problems lipschitz extension nearest neighbor search. algorithm's generalization performance guaranteed via fat-shattering dimension lipschitz classifiers, present experimental evidence superiority common kernel methods. by-product, offer new perspective nearest neighbor classifier, yields significantly sharper risk asymptotics classic analysis cover hart [ieee trans. info. theory, 1967].",3
"time-dependent interpretation neutrino emission tidal disruption events. three tidal disruption event (tde) candidates (at2019dsg, at2019fdr, at2019aalc) associated high energy astrophysical neutrinos multi-messenger follow-ups. cases, neutrino observation occurred o(100) days maximum optical-ultraviolet (ouv) luminosity. discuss unified fully time-dependent interpretations events, neutrino delays statistical effect, rather consequence physical scale post-disruption system. noting x-rays infrared (ir) dust echoes observed cases, consider three models quasi-isotropic neutrino emission due interactions accelerated protons moderate, medium, high energy energies x-rays, ouv, ir photons, respectively. find neutrino time delays well described x-ray model assuming magnetic confinement protons calorimetric approach, ir model, delay directly correlated time evolution echo luminosity (for model developed here). ouv model exhibits highest neutrino production efficiency. three models, highest neutrino fluence predicted at2019aalc, due high estimated smbh mass low redshift. models result diffuse neutrino fluxes consistent observations.",0
"deepgestalt - identifying rare genetic syndromes using deep learning. facial analysis technologies recently measured capabilities expert clinicians syndrome identification. date, technologies could identify phenotypes diseases, limiting role clinical settings hundreds diagnoses must considered. developed facial analysis framework, deepgestalt, using computer vision deep learning algorithms, quantifies similarities hundreds genetic syndromes based unconstrained 2d images. deepgestalt currently trained 26,000 patient cases rapidly growing phenotype-genotype database, consisting tens thousands validated clinical cases, curated community-driven platform. deepgestalt currently achieves 91% top-10-accuracy identifying 215 different genetic syndromes outperformed clinical experts three separate experiments. suggest form artificial intelligence ready support medical genetics clinical laboratory practices play key role future precision medicine.",3
"two notions total positivity partial flag varieties. given integers $1 \le k_1 < \cdots < k_l \le n-1$, let fl($k_1, \dots, k_l$; $n$) denote type partial flag variety consisting chains subspaces $(v_{k_1} \subset \cdots \subset v_{k_l})$ inside $\mathbb{r}^n$, $v_k$ dimension $k$. lusztig (1994, 1998) introduced totally positive part fl($k_1, \dots, k_l$; $n$) subset partial flags represented totally positive $n\times n$ matrix, defined totally nonnegative part closure totally positive part. hand, following postnikov (2007), define pl\""{u}cker-positive pl\""{u}cker-nonnegative parts fl($k_1, \dots, k_l$; $n$) subsets pl\""{u}cker coordinates positive nonnegative, respectively. follows definitions lusztig's total positivity implies pl\""{u}cker positivity, natural ask two notions positivity agree. rietsch (2009) proved agree case grassmannian fl($k$; $n$), chevalier (2011) showed two notions distinct fl($1,3$; $4$). show general, two notions agree $k_1, \dots, k_l$ consecutive integers. give elementary proof result (including case grassmannians) based classical results linear algebra theory total positivity. also show cell decomposition totally nonnegative part fl($k_1, \dots, k_l$; $n$) coincides matroid decomposition $k_1, \dots, k_l$ consecutive integers, previously known complete flag varieties, grassmannians, fl($1,3$; $4$). finally, determine notions positivity compatible natural action cyclic group order $n$ rotates index set.",10
"end-to-end prediction buffer overruns raw source code via neural memory networks. detecting buffer overruns source code one common yet challenging tasks program analysis. current approaches mainly relied rigid rules handcrafted features devised experts, limiting terms flexible applicability robustness due diverse bug patterns characteristics existing sophisticated real-world software programs. paper, propose novel, data-driven approach completely end-to-end without requiring hand-crafted features, thus free program language-specific structural limitations. particular, approach leverages recently proposed neural network model called memory networks shown state-of-the-art performances mainly question-answering tasks. experimental results using source codes demonstrate proposed model capable accurately detecting simple buffer overruns. also present in-depth analyses memory network learn understand semantics programming languages solely raw source codes, tracing variables interest, identifying numerical values, performing quantitative comparisons.",3
"bayesian estimation multidimensional latent variables asymptotic accuracy. hierarchical learning models, mixture models bayesian networks, widely employed unsupervised learning tasks, clustering analysis. consist observable hidden variables, represent given data hidden generation process, respectively. pointed conventional statistical analysis applicable models, redundancy latent variable produces singularities parameter space. recent years, method based algebraic geometry allowed us analyze accuracy predicting observable variables using bayesian estimation. however, analyze latent variables sufficiently studied, even though one main issues unsupervised learning determine accurately latent variable estimated. previous study proposed method used range latent variable redundant compared model generating data. present paper extends method situation latent variables redundant dimensions. formulate new error functions derive asymptotic forms. calculation error functions demonstrated two-layered bayesian networks.",18
"applying machine translation two-stage cross-language information retrieval. cross-language information retrieval (clir), queries documents different languages, needs translation queries and/or documents, standardize common representation. purpose, use machine translation effective approach. however, computational cost prohibitive translating large-scale document collections. resolve problem, propose two-stage clir method. first, translate given query document language, retrieve limited number foreign documents. second, machine translate documents user language, re-rank based translation result. also show effectiveness method way experiments using japanese queries english technical documents.",3
"incremental knowledge base construction using deepdive. populating database unstructured information long-standing problem industry research encompasses problems extraction, cleaning, integration. recent names used problem include dealing dark data knowledge base construction (kbc). work, describe deepdive, system combines database machine learning ideas help develop kbc systems, present techniques make kbc process efficient. observe kbc process iterative, develop techniques incrementally produce inference results kbc systems. propose two methods incremental inference, based respectively sampling variational techniques. also study tradeoff space methods develop simple rule-based optimizer. deepdive includes contributions, evaluate deepdive five kbc systems, showing speed kbc inference tasks two orders magnitude negligible impact quality.",3
"analysis planar ornament patterns via motif asymmetry assumption local connections. planar ornaments, a.k.a. wallpapers, regular repetitive patterns exhibit translational symmetry two independent directions. exactly $17$ distinct planar symmetry groups. present fully automatic method complete analysis planar ornaments $13$ groups, specifically, groups called $p6m, \, p6, \, p4g, \,p4m, \,p4, \, p31m, \,p3m, \, p3, \, cmm, \, pgg, \, pg, \, p2$ $p1$. given image ornament fragment, present method simultaneously classify input one $13$ groups extract called fundamental domain (fd), minimum region sufficient reconstruct entire ornament. nice feature method even given ornament image small portion contain multiple translational units, symmetry group well fundamental domain still defined. because, contrast common approach, attempt first identify global translational repetition lattice. though presented constructions work quite wide range ornament patterns, key assumption make perceivable motifs (shapes repeat) alone provide clues underlying symmetries ornament. sense, main target planar arrangements asymmetric interlocking shapes, symmetry art escher.",3
"bibliographical references intonation intonational meaning. by-no-means-complete collection references interested intonational meaning, miscellaneous references intonation included. additional references welcome, sent julia@research.att.com.",1
"sequential monte carlo bandits. paper propose flexible efficient framework handling multi-armed bandits, combining sequential monte carlo algorithms hierarchical bayesian modeling techniques. framework naturally encompasses restless bandits, contextual bandits, bandit variants single inferential model. despite model's generality, propose efficient monte carlo algorithms make inference scalable, based recent developments sequential monte carlo methods. two simulation studies, framework shown outperform empirical methods, also naturally scaling complex problems existing approaches cope. additionally, successfully apply framework online video-based advertising recommendation, show increased efficacy compared current state art bandit algorithms.",18
hybrid fuzzy-art based k-means clustering methodology cellular manufacturing using operational time. paper presents new hybrid fuzzy-art based k-means clustering technique solve part machine grouping problem cellular manufacturing systems considering operational time. performance proposed technique tested problems open literature results compared existing clustering models simple k-means algorithm modified art1 algorithm using efficient modified performance measure known modified grouping efficiency (mge) found literature. results support better performance proposed algorithm. novelty study lies simple efficient methodology produce quick solutions shop floor managers least computational efforts time.,3
"automatic response assessment regions language cortex epilepsy patients using ecog-based functional mapping machine learning. accurate localization brain regions responsible language cognitive functions epilepsy patients carefully determined prior surgery. electrocorticography (ecog)-based real time functional mapping (rtfm) shown safer alternative electrical cortical stimulation mapping (esm), currently clinical/gold standard. conventional methods analyzing rtfm signals based statistical comparison signal power certain frequency bands. compared gold standard (esm), limited accuracies assessing channel responses. study, address accuracy limitation current rtfm signal estimation methods analyzing full frequency spectrum signal replacing signal power estimation methods machine learning algorithms, specifically random forest (rf), proof concept. train rf power spectral density time-series rtfm signal supervised learning framework ground truth labels obtained esm. results obtained rtfm six adult patients strictly controlled experimental setup reveal state art detection accuracy $\approx 78\%$ language comprehension task, improvement $23\%$ conventional rtfm estimation method. best knowledge, first study exploring use machine learning approaches determining rtfm signal characteristics, using whole-frequency band better region localization. results demonstrate feasibility machine learning based rtfm signal analysis method full spectrum clinical routine near future.",15
"understanding task design trade-offs crowdsourced paraphrase collection. linguistically diverse datasets critical training evaluating robust machine learning systems, data collection costly process often requires experts. crowdsourcing process paraphrase generation effective means expanding natural language datasets, limited analysis trade-offs arise designing tasks. paper, present first systematic study key factors crowdsourcing paraphrase collection. consider variations instructions, incentives, data domains, workflows. manually analyzed paraphrases correctness, grammaticality, linguistic diversity. observations provide new insight trade-offs accuracy diversity crowd responses arise result task design, providing guidance future paraphrase generation procedures.",3
"semi-parametric topological memory navigation. introduce new memory architecture navigation previously unseen environments, inspired landmark-based navigation animals. proposed semi-parametric topological memory (sptm) consists (non-parametric) graph nodes corresponding locations environment (parametric) deep network capable retrieving nodes graph based observations. graph stores metric information, connectivity locations corresponding nodes. use sptm planning module navigation system. given 5 minutes footage previously unseen maze, sptm-based navigation agent build topological map environment use confidently navigate towards goals. average success rate sptm agent goal-directed navigation across test environments higher best-performing baseline factor three. video agent available https://youtu.be/vrf7f4lhswo",3
"multi-view bayesian correlated component analysis. correlated component analysis proposed dmochowski et al. (2012) tool investigating brain process similarity responses multiple views given stimulus. correlated components identified assumption involved spatial networks identical. propose hierarchical probabilistic model infer level universality multi-view data, completely unrelated representations, corresponding canonical correlation analysis, identical representations correlated component analysis. new model, denote bayesian correlated component analysis, evaluates favourably three relevant algorithms simulated data. well-established benchmark eeg dataset used validate new model infer variability spatial representations across multiple subjects.",18
"improving tagging performance using voting taggers. present bootstrapping method develop annotated corpus, specially useful languages available resources. method applied develop corpus spanish 5mw. method consists taking advantage collaboration two different pos taggers. cases taggers agree present higher accuracy used retrain taggers.",3
"information-maximization clustering based squared-loss mutual information. information-maximization clustering learns probabilistic classifier unsupervised manner mutual information feature vectors cluster assignments maximized. notable advantage approach involves continuous optimization model parameters, substantially easier solve discrete optimization cluster assignments. however, existing methods still involve non-convex optimization problems, therefore finding good local optimal solution straightforward practice. paper, propose alternative information-maximization clustering method based squared-loss variant mutual information. novel approach gives clustering solution analytically computationally efficient way via kernel eigenvalue decomposition. furthermore, provide practical model selection procedure allows us objectively optimize tuning parameters included kernel function. experiments, demonstrate usefulness proposed approach.",18
"asymptotic preserving energy stable scheme barotropic euler system incompressible limit. asymptotic preserving energy stable scheme barotropic euler system low mach number scaling designed analysed. velocity shift proportional pressure gradient introduced convective fluxes, leads dissipation mechanical energy entropy stability mach numbers. resolution semi-implicit time upwind space fully-discrete scheme involves two steps: solution elliptic problem density explicit evaluation velocity. proposed scheme possess several physically relevant attributes, positivity density, entropy stability consistency weak formulation continuous euler system. ap property scheme, i.e.\ boundedness mesh parameters respect mach number consistency incompressible limit system, shown rigorously. results extensive case studies presented substantiate robustness efficacy proposed scheme well theoretical claims.",10
"deepapt: nation-state apt attribution using end-to-end deep neural networks. recent years numerous advanced malware, aka advanced persistent threats (apt) allegedly developed nation-states. task attributing apt specific nation-state extremely challenging several reasons. nation-state usually single cyber unit develops advanced malware, rendering traditional authorship attribution algorithms useless. furthermore, apts use state-of-the-art evasion techniques, making feature extraction challenging. finally, dataset available apts extremely small. paper describe deep neural networks (dnn) could successfully employed nation-state apt attribution. use sandbox reports (recording behavior apt run dynamically) raw input neural network, allowing dnn learn high level feature abstractions apts itself. using test set 1,000 chinese russian developed apts, achieved accuracy rate 94.6%.",3
"boosting cooperative coevolution large scale optimization fine-grained computation resource allocation strategy. cooperative coevolution (cc) shown great potential solving large scale optimization problems (lsops). however, traditional cc algorithms often waste part computation resource (cr) equally allocate cr among subproblems. recently developed contribution-based cc (cbcc) algorithms improve traditional ones certain extent adaptively allocating cr according heuristic rules. different existing works, study explicitly constructs mathematical model cr allocation (cra) problem cc proposes novel fine-grained cra (fcra) strategy fully considering theoretically optimal solution cra model evolution characteristics cc. fcra takes single iteration basic cra unit always selects subproblem likely make largest contribution total fitness improvement undergo new iteration, contribution subproblem new iteration estimated according current contribution, current evolution status well estimation current contribution. verified efficiency fcra combining shade excellent differential evolution variant never employed cc framework. experimental results two benchmark suites lsops demonstrate fcra significantly outperforms existing cra strategies resultant cc algorithm highly competitive solving lsops.",3
"human activity learning using object affordances rgb-d videos. human activities comprise several sub-activities performed sequence involve interactions various objects. makes reasoning object affordances central task activity recognition. work, consider problem jointly labeling object affordances human activities rgb-d videos. frame problem markov random field nodes represent objects sub-activities, edges represent relationships object affordances, relations sub-activities, evolution time. formulate learning problem using structural svm approach, labeling various alternate temporal segmentations considered latent variables. tested method dataset comprising 120 activity videos collected four subjects, obtained end-to-end precision 81.8% recall 80.0% labeling activities.",3
"painter: spatio-spectral image reconstruction algorithm optical interferometry. astronomical optical interferometers sample fourier transform intensity distribution source observation wavelength. rapid perturbations caused atmospheric turbulence, phases complex fourier samples (visibilities) cannot directly exploited. consequently, specific image reconstruction methods devised last decades. modern polychromatic optical interferometric instruments paving way multiwavelength imaging. paper devoted derivation spatio-spectral (3d) image reconstruction algorithm, coined painter (polychromatic optical interferometric reconstruction software). algorithm relies iterative process, alternates estimation polychromatic images complex visibilities. complex visibilities estimated squared moduli closure phases, also differential phases, helps better constrain polychromatic reconstruction. simulations synthetic data illustrate efficiency algorithm particular relevance injecting differential phases model reconstruction.",0
"tableaux dynamic logic propositional assignments. dynamic logic propositional assignments (dl-pa) recently studied alternative propositional dynamic logic (pdl). dl-pa, abstract atomic programs pdl replaced assignments propositional variables truth values. makes dl-pa enjoy interesting meta-logical properties pdl not, eliminability kleene star, compactness interpolation. define analytic tableaux calculus dl-pa show matches known complexity results.",3
"soft rule ensembles statistical learning. article supervised learning problems solved using soft rule ensembles. first review importance sampling learning ensembles (isle) approach useful generating hard rules. soft rules obtained logistic regression corresponding hard rules. order deal perfect separation problem related logistic regression, firth's bias corrected likelihood used. various examples simulation results show soft rule ensembles improve predictive performance hard rule ensembles.",18
"deterministic conditions subspace clustering missing data. paper present deterministic conditions success sparse subspace clustering (ssc) missing data, data assumed come union subspaces (uos) model. consider two algorithms, variants ssc entry-wise zero-filling differ terms optimization problems used find affinity matrix spectral clustering. algorithms, provide deterministic conditions pattern missing data perfect clustering achieved. provide extensive sets simulation results clustering well completion data missing entries, uos model. experimental results indicate contrast full data case, accurate clustering imply accurate subspace identification completion, indicating natural order relative hardness problems.",3
"scalable adaptive stochastic optimization using random projections. adaptive stochastic gradient methods adagrad gained popularity particular training deep neural networks. commonly used studied variant maintains diagonal matrix approximation second order information accumulating past gradients used tune step size adaptively. certain situations full-matrix variant adagrad expected attain better performance, however high dimensions computationally impractical. present ada-lr radagrad two computationally efficient approximations full-matrix adagrad based randomized dimensionality reduction. able capture dependencies features achieve similar performance full-matrix adagrad much smaller computational cost. show regret ada-lr close regret full-matrix adagrad up-to exponentially smaller dependence dimension diagonal variant. empirically, show ada-lr radagrad perform similarly full-matrix adagrad. task training convolutional neural networks well recurrent neural networks, radagrad achieves faster convergence diagonal adagrad.",18
"dropping convexity faster semi-definite optimization. study minimization convex function $f(x)$ set $n\times n$ positive semi-definite matrices, problem recast $\min_u g(u) := f(uu^\top)$, $u \in \mathbb{r}^{n \times r}$ $r \leq n$. study performance gradient descent $g$---which refer factored gradient descent (fgd)---under standard assumptions original function $f$. provide rule selecting step size and, choice, show local convergence rate fgd mirrors standard gradient descent original $f$: i.e., $k$ steps, error $o(1/k)$ smooth $f$, exponentially small $k$ $f$ (restricted) strongly convex. addition, provide procedure initialize fgd (restricted) strongly convex objectives one access $f$ via first-order oracle; several problem instances, proper initialization leads global convergence guarantees. fgd similar procedures widely used practice problems posed matrix factorization. best knowledge, first paper provide precise convergence rate guarantees general convex functions standard convex assumptions.",18
"improved k-means clustering based approach detect dna structure h&e image mouse tissue reacted cd4-green antigen. manuscript present technique detect analyze dna rich structure haemotoxylin & eosin (h&e) image tissue treated anti cd4 green antigen. detection dna rich structure considered detection blue nuclei present biomedical signal/image processing technique performed image tissue obtained scanning electron microscope(sem). earlier tissue treated anti cd4 green antigen, stained h&e staining solution.",3
"automated correction syntax errors programming assignments using recurrent neural networks. present method automatically generating repair feedback syntax errors introductory programming problems. syntax errors constitute one largest classes errors (34%) dataset student submissions obtained mooc course edx. previous techniques generating automated feed- back programming assignments focused functional correctness style considerations student programs. techniques analyze program ast program perform dynamic symbolic analyses compute repair feedback. unfortunately, possible generate asts student pro- grams syntax errors therefore previous feedback techniques applicable repairing syntax errors. present technique providing feedback syntax errors uses recurrent neural networks (rnns) model syntactically valid token sequences. approach inspired recent work learning language models big code (large code corpus). given programming assignment, first learn rnn model valid token sequences using set syntactically correct student submissions. then, student submission syntax errors, query learnt rnn model prefix to- ken sequence predict token sequences fix error either replacing inserting predicted token sequence error location. evaluate technique 14, 000 student submissions syntax errors. technique completely re- pair 31.69% (4501/14203) submissions syntax errors addition partially correct 6.39% (908/14203) submissions.",3
"character composition model convolutional neural networks dependency parsing morphologically rich languages. present transition-based dependency parser uses convolutional neural network compose word representations characters. character composition model shows great improvement word-lookup model, especially parsing agglutinative languages. improvements even better using pre-trained word embeddings extra data. spmrl data sets, system outperforms previous best greedy parser (ballesteros et al., 2015) margin 3% average.",3
"neuroprosthetic decoder training imitation learning. neuroprosthetic brain-computer interfaces function via algorithm decodes neural activity user movements end effector, cursor robotic arm. practice, decoder often learned updating parameters user performs task. user's intention directly observable, recent methods demonstrated value training decoder surrogate user's intended movement. describe training decoder way novel variant imitation learning problem, oracle expert employed supervised training lieu direct observations, available. specifically, describe generic imitation learning meta-algorithm, dataset aggregation (dagger, [1]), adapted train generic brain-computer interface. deriving existing learning algorithms brain-computer interfaces framework, provide novel analysis regret (an important metric learning efficacy) brain-computer interfaces. analysis allows us characterize space algorithmic variants bounds regret rates. existing approaches decoder learning performed cursor control setting, available design principles decoders impossible scale naturalistic settings. leveraging findings, offer algorithm combines imitation learning optimal control, allow training arbitrary effectors optimal control generate goal-oriented control. demonstrate novel general bci algorithm simulated neuroprosthetic control 26 degree-of-freedom model arm, sophisticated realistic end effector.",18
"wavelet fast fourier transform based analysis solar image. wavelet fast fourier transform strong signal processing tools field data analysis. paper fast fourier transform (fft) wavelet transform employed observe important features solar image (december, 2004). tried find periodicity coherence different sections solar image. plotted distribution energy solar surface analyzing solar image scalograms 3d-coefficient plots.",3
"dna2vec: consistent vector representations variable-length k-mers. one ubiquitous representation long dna sequence dividing shorter k-mer components. unfortunately, straightforward vector encoding k-mer one-hot vector vulnerable curse dimensionality. worse yet, distance pair one-hot vectors equidistant. particularly problematic applying latest machine learning algorithms solve problems biological sequence analysis. paper, propose novel method train distributed representations variable-length k-mers. method based popular word embedding model word2vec, trained shallow two-layer neural network. experiments provide evidence summing dna2vec vectors akin nucleotides concatenation. also demonstrate correlation needleman-wunsch similarity score cosine similarity dna2vec vectors.",15
"hierarchical embeddings hypernymy detection directionality. present novel neural model hypervec learn hierarchical embeddings hypernymy detection directionality. previous embeddings shown limitations prototypical hypernyms, hypervec represents unsupervised measure embeddings learned specific order capture hypernym$-$hyponym distributional hierarchy. moreover, model able generalize unseen hypernymy pairs, using small sets training data, mapping languages. results benchmark datasets show hypervec outperforms state$-$of$-$the$-$art unsupervised measures embedding models hypernymy detection directionality, predicting graded lexical entailment.",3
"redundancy, deduction schemes, minimum-size bases association rules. association rules among widely employed data analysis methods field data mining. association rule form partial implication two sets binary variables. common approach, association rules parameterized lower bound confidence, empirical conditional probability consequent given antecedent, and/or parameter bounds ""support"" deviation independence. study notions redundancy among association rules fundamental perspective. see transaction dataset interpretation (or model) propositional logic sense, consider existing notions redundancy, is, logical entailment, among association rules, form ""any dataset first rule holds must obey also second rule, therefore second redundant"". discuss several existing alternative definitions redundancy association rules provide new characterizations relationships among them. show main alternatives discuss correspond actually two variants, differ treatment full-confidence implications. two notions redundancy, provide sound complete deduction calculus, show construct complete bases (that is, axiomatizations) absolutely minimum size terms number rules. explore finally approach redundancy respect several association rules, fully characterize simplest case two partial premises.",3
"pixelnet: representation pixels, pixels, pixels. explore design principles general pixel-level prediction problems, low-level edge detection mid-level surface normal estimation high-level semantic segmentation. convolutional predictors, fully-convolutional network (fcn), achieved remarkable success exploiting spatial redundancy neighboring pixels convolutional processing. though computationally efficient, point approaches statistically efficient learning precisely spatial redundancy limits information learned neighboring pixels. demonstrate stratified sampling pixels allows one (1) add diversity batch updates, speeding learning; (2) explore complex nonlinear predictors, improving accuracy; (3) efficiently train state-of-the-art models tabula rasa (i.e., ""from scratch"") diverse pixel-labeling tasks. single architecture produces state-of-the-art results semantic segmentation pascal-context dataset, surface normal estimation nyudv2 depth dataset, edge detection bsds.",3
"kernel bayes' rule. nonparametric kernel-based method realizing bayes' rule proposed, based representations probabilities reproducing kernel hilbert spaces. probabilities uniquely characterized mean canonical map rkhs. prior conditional probabilities expressed terms rkhs functions empirical sample: explicit parametric model needed quantities. posterior likewise rkhs mean weighted sample. estimator expectation function posterior derived, rates consistency shown. representative applications kernel bayes' rule presented, including baysian computation without likelihood filtering nonparametric state-space model.",18
"mobirnn: efficient recurrent neural network execution mobile gpu. paper, explore optimizations run recurrent neural network (rnn) models locally mobile devices. rnn models widely used natural language processing, machine translation, tasks. however, existing mobile applications use rnn models cloud. address privacy efficiency concerns, show rnn models run locally mobile devices. existing work porting deep learning models mobile devices focus convolution neural networks (cnns) cannot applied directly rnn models. response, present mobirnn, mobile-specific optimization framework implements gpu offloading specifically mobile gpus. evaluations using rnn model activity recognition shows mobirnn significantly decrease latency running rnn models phones.",3
"dynamic policy programming. paper, propose novel policy iteration method, called dynamic policy programming (dpp), estimate optimal policy infinite-horizon markov decision processes. prove finite-iteration asymptotic l\infty-norm performance-loss bounds dpp presence approximation/estimation error. bounds expressed terms l\infty-norm average accumulated error opposed l\infty-norm error case standard approximate value iteration (avi) approximate policy iteration (api). suggests dpp achieve better performance avi api since averages simulation noise caused monte-carlo sampling throughout learning process. examine theoretical results numerically com- paring performance approximate variants dpp existing reinforcement learning (rl) methods different problem domains. results show that, cases, dpp-based algorithms outperform rl methods wide margin.",3
linear learning sparse data. linear predictors especially useful data high-dimensional sparse. one standard techniques used train linear predictor averaged stochastic gradient descent (asgd) algorithm. present efficient implementation asgd avoids dense vector operations. also describe translation invariant extension called centered averaged stochastic gradient descent (casgd).,3
"deep structured output learning unconstrained text recognition. develop representation suitable unconstrained recognition words natural images: general case fixed lexicon unknown length. end propose convolutional neural network (cnn) based architecture incorporates conditional random field (crf) graphical model, taking whole word image single input. unaries crf provided cnn predicts characters position output, higher order terms provided another cnn detects presence n-grams. show entire model (crf, character predictor, n-gram predictor) jointly optimised back-propagating structured output loss, essentially requiring system perform multi-task learning, training uses purely synthetically generated data. resulting model accurate system standard real-world text recognition benchmarks character prediction alone, setting benchmark systems trained particular lexicon. addition, model achieves state-of-the-art accuracy lexicon-constrained scenarios, without specifically modelled constrained recognition. test generalisation model, also perform experiments random alpha-numeric strings evaluate method visual language model applicable.",3
mining relevant interval rules. article extends method garriga et al. mining relevant rules numerical attributes extracting interval-based pattern rules. propose algorithm extracts rules numerical datasets using interval-pattern approach kaytoue et al. algorithm implemented evaluated real datasets.,3
"interaction histories short term memory: enactive development turn-taking behaviors childlike humanoid robot. article, enactive architecture described allows humanoid robot learn compose simple actions turn-taking behaviors playing interaction games human partner. robot's action choices reinforced social feedback human form visual attention measures behavioral synchronization. demonstrate system acquire switch behaviors learned interaction based social feedback human partner. role reinforcement based short term memory interaction experimentally investigated. results indicate feedback based immediate state insufficient learn certain turn-taking behaviors. therefore history interaction must considered acquisition turn-taking, efficiently handled use short term memory.",3
"tensor decomposition signal processing machine learning. tensors {\em multi-way arrays} functions three indices $(i,j,k,\cdots)$ -- similar matrices (two-way arrays), functions two indices $(r,c)$ (row,column). tensors rich history, stretching almost century, touching upon numerous disciplines; recently become ubiquitous signal data analytics confluence signal processing, statistics, data mining machine learning. overview article aims provide good starting point researchers practitioners interested learning working tensors. such, focuses fundamentals motivation (using various application examples), aiming strike appropriate balance breadth {\em depth} enable someone taken first graduate courses matrix algebra probability get started research and/or developing tensor algorithms software. background applied optimization useful strictly required. material covered includes tensor rank rank decomposition; basic tensor factorization models relationships properties (including fairly good coverage identifiability); broad coverage algorithms ranging alternating optimization stochastic gradient; statistical performance analysis; applications ranging source separation collaborative filtering, mixture topic modeling, classification, multilinear subspace learning.",18
"practical ontology large-scale modeling scholarly artifacts usage. large-scale analysis scholarly artifact usage constrained primarily current practices usage data archiving, privacy issues concerned dissemination usage data, lack practical ontology modeling usage domain. remedy third constraint, article presents scholarly ontology engineered represent classes large-scale bibliographic usage data exists, supports usage research, whose instantiation scalable order 50 million articles along associated artifacts (e.g. authors journals) accompanying 1 billion usage events. real world instantiation presented abstract ontology semantic network model scholarly community lends scholarly process statistical analysis computational support. present ontology, discuss instantiation, provide example inference rules calculating various scholarly artifact metrics.",3
"big data sp theory intelligence. article ""sp theory intelligence"" realisation ""sp machine"" may, advantage, applied management analysis big data. sp system -- introduced article fully described elsewhere -- may help overcome problem variety big data: potential ""a universal framework representation processing diverse kinds knowledge"" (ufk), helping reduce diversity formalisms formats knowledge different ways processed. strengths unsupervised learning discovery structure data, pattern recognition, parsing production natural language, several kinds reasoning, more. lends analysis streaming data, helping overcome problem velocity big data. central workings system lossless compression information: making big data smaller reducing problems storage management. potential substantial economies transmission data, big cuts use energy computing, faster processing, smaller lighter computers. system provides handle problem veracity big data, potential assist management errors uncertainties data. lends visualisation knowledge structures inferential processes. high-parallel, open-source version sp machine would provide means researchers everywhere explore done system create new versions it.",3
"symmetry breaking using value precedence. present comprehensive study use value precedence constraints break value symmetry. first give simple encoding value precedence ternary constraints efficient effective breaking symmetry. extend value precedence deal number generalizations like wreath value partial interchangeability. also show value precedence closely related lexicographical ordering. finally, consider interaction value precedence symmetry breaking constraints variable symmetries.",3
"measurement $z$ boson production cross-section proton-lead collisions $\sqrt{s_\mathrm{nn}}=8.16\,\mathrm{tev}$. article presents first measurement differential $z$-boson production cross-section forward region using proton-lead collisions lhcb detector. dataset collected nucleon-nucleon centre-of-mass energy $\sqrt{s_\mathrm{nn}}=8.16\,\mathrm{tev}$ 2016, corresponding integrated luminosity $30.8\,\mathrm{nb}^{-1}$. forward-backward ratio nuclear modification factors measured together differential cross-section functions $z$ boson rapidity centre-of-mass frame, transverse momentum $z$ boson geometric variable $\phi^{*}$. results good agreement predictions nuclear parton distribution functions, providing strong constraining power small bjorken-$x$.",7
"pav algorithm optimizes binary proper scoring rules. much recent interest application pool-adjacent-violators (pav) algorithm purpose calibrating probabilistic outputs automatic pattern recognition machine learning algorithms. special cost functions, known proper scoring rules form natural objective functions judge goodness calibration. show binary pattern classifiers, non-parametric optimization calibration, subject monotonicity constraint, solved pav solution optimal regular binary proper scoring rules. extends previous results limited convex binary proper scoring rules. show result holds calibration probabilities, also calibration log-likelihood-ratios, case optimality holds independently prior probabilities pattern classes.",18
"semantic autoencoder zero-shot learning. existing zero-shot learning (zsl) models typically learn projection function feature space semantic embedding space (e.g.~attribute space). however, projection function concerned predicting training seen class semantic representation (e.g.~attribute prediction) classification. applied test data, context zsl contains different (unseen) classes without training data, zsl model typically suffers project domain shift problem. work, present novel solution zsl based learning semantic autoencoder (sae). taking encoder-decoder paradigm, encoder aims project visual feature vector semantic space existing zsl models. however, decoder exerts additional constraint, is, projection/code must able reconstruct original visual feature. show additional reconstruction constraint, learned projection function seen classes able generalise better new unseen classes. importantly, encoder decoder linear symmetric enable us develop extremely efficient learning algorithm. extensive experiments six benchmark datasets demonstrate proposed sae outperforms significantly existing zsl models additional benefit lower computational cost. furthermore, sae applied supervised clustering problem, also beats state-of-the-art.",3
"feasible roadmap developing volumetric probability atlas localized prostate cancer. statistical volumetric model, showing probability map localized prostate cancer within host anatomical structure, developed 90 optically-imaged surgical specimens. master model permits accurate characterization prostate cancer distribution patterns atlas-informed biopsy sampling strategy. model constructed mapping individual prostate models onto site model, together localized tumors. accurate multi-object non-rigid warping scheme developed based mixture principal-axis registrations. report evaluation pilot studies effectiveness method application optimizing needle biopsy strategies.",15
"stochastic semismooth newton method nonsmooth nonconvex optimization. work, present globalized stochastic semismooth newton method solving stochastic optimization problems involving smooth nonconvex nonsmooth convex terms objective function. assume noisy gradient hessian information smooth part objective function available via calling stochastic first second order oracles. proposed method seen hybrid approach combining stochastic semismooth newton steps stochastic proximal gradient steps. two inexact growth conditions incorporated monitor convergence acceptance semismooth newton steps shown algorithm converges globally stationary points expectation. moreover, standard assumptions utilizing random matrix concentration inequalities, prove proposed approach locally turns pure stochastic semismooth newton method converges r-superlinearly high probability. present numerical results comparisons $\ell_1$-regularized logistic regression nonconvex binary classification demonstrate efficiency algorithm.",10
"majorana-weyl cones ferroelectric superconductors. topological superconductors predicted exhibit outstanding phenomena, including non-abelian anyon excitations, heat-carrying edge states, topological nodes bogoliubov spectra. nonetheless, despite major experimental efforts, still lacking unambiguous signatures exotic phenomena. context, recent discovery coexisting superconductivity ferroelectricity lightly doped ultra clean srtio$_3$ opens new opportunities. indeed, promising route engineer topological superconductivity combination strong spin-orbit coupling inversion-symmetry breaking. study three-dimensional parabolic band minimum rashba spin-orbit coupling, whose axis aligned direction ferroelectric moment. show aforementioned phenomena naturally emerge model magnetic field applied. critical zeeman field, majorana-weyl cones emerge regardless electronic density. cones manifest majorana arcs states appearing surfaces tetragonal domain walls. rotating magnetic field respect direction ferroelectric moment tilts majorana-weyl cones, eventually driving type-ii state bogoliubov fermi surfaces. consider consequences orbital magnetic field. first, single vortex found surrounded topological halo, characterized two majorana zero modes: one localized vortex core boundary topological halo. finite density vortices forming close enough upper critical field, halos overlap eventually percolate system, causing bulk topological transition always precedes normal state. finally, propose concrete experiments test predictions.",2
"training restricted boltzmann machines via thouless-anderson-palmer free energy. restricted boltzmann machines undirected neural networks shown effective many applications, including serving initializations training deep multi-layer neural networks. one main reasons success existence efficient practical stochastic algorithms, contrastive divergence, unsupervised training. propose alternative deterministic iterative procedure based improved mean field method statistical physics known thouless-anderson-palmer approach. demonstrate algorithm provides performance equal to, sometimes superior to, persistent contrastive divergence, also providing clear easy evaluate objective function. believe strategy easily generalized models well accurate higher-order approximations, paving way systematic improvements training boltzmann machines hidden units.",2
"reasoning agent programs using atl-like logics. propose variant alternating-time temporal logic (atl) grounded agents' operational know-how, defined libraries abstract plans. inspired atles, variant atl, possible logic explicitly refer ""rational"" strategies agents developed belief-desire-intention agent programming paradigm. allows us express verify properties bdi systems using atl-type logical frameworks.",3
"cognitive dynamic systems: technical review cognitive radar. start history cognitive radar, origins pac, fuster research cognition principals cognition provided. fuster describes five cognitive functions: perception, memory, attention, language, intelligence. describe perception-action cyclec applies cognitive radar, discuss long-term memory, memory storage, memory retrieval working memory. comparison memory human cognition cognitive radar given well. attention another function described fuster, given comparison attention human cognition cognitive radar. talk four functional blocks pac: bayesian filter, feedback information, dynamic programming state-space model radar environment. then, show pac improves tracking accuracy cognitive radar traditional active radar, provided simulation results. simulation, three nonlinear filters: cubature kalman filter, unscented kalman filter extended kalman filter compared. based results, radars implemented ckf perform better radars implemented ukf radars implemented ekf. further, radar ekf worst accuracy biggest computation load derivation evaluation jacobian matrices. suggest using concept risk management better control parameters improve performance cognitive radar. believe, spectrum sensing seen potential interest used cognitive radar propose new approach probabilistic ica presumably reduce noise based estimation error cognitive radar. parallel computing concept based divide conquers mechanism, suggest using parallel computing approach cognitive radar complicated calculations tasks reduce processing time.",3
"hardware design binarization thinning fingerprint images. two critical steps fingerprint recognition binarization thinning image. need real time processing motivates us select local adaptive thresholding approach binarization step. introduce new hardware purpose based pipeline architecture. propose formula selecting optimal block size thresholding purpose. decrease minutiae false detection, binarized image dilated. also present paper new pipeline structure implementing thinning algorithm",3
"multi-target tracking algorithms 3d. ladars provide unique capability identification objects motions scenes fixed 3d field view (fov). paper describes algorithms multi-target tracking 3d scenes including preprocessing (mathematical morphology parzen windows), labeling connected components, sorting targets selectable attributes (size, length track, velocity), handling target states (acquired, coasting, re-acquired tracked) order assemble target trajectories. paper derived working algorithms coded matlab, tested reviewed others, speculate usage general formulas frameworks.",3
"copula-based kernel dependency measures. paper presents new copula based method measuring dependence random variables. approach extends maximum mean discrepancy copula joint distribution. prove approach several advantageous properties. similarly shannon mutual information, proposed dependence measure invariant strictly increasing transformation marginal variables. important many applications, example feature selection. estimator consistent, robust outliers, uses rank statistics only. derive upper bounds convergence rate propose independence tests too. illustrate theoretical contributions series experiments feature selection low-dimensional embedding distributions.",3
"generators elliptic curve $e_{(p,q)} : y^2 = x^3 - p^2x + q^2$. let $\{e_{(p,q)}\}$ family elliptic curves rational field $e_{(p,q)} : y^2 = x^3 - p^2x + q^2$, $p$ $q$ prime numbers greater five. earlier work showed elliptic curve $e_{(p,q)}$ ranked least two $p, q > 5$ two independent points. paper shows two points extended basis $e_{(p,q)}$ conditions confident fully recover.",10
"quasi-monte carlo feature maps shift-invariant kernels. consider problem improving efficiency randomized fourier feature maps accelerate training testing speed kernel methods large datasets. approximate feature maps arise monte carlo approximations integral representations shift-invariant kernel functions (e.g., gaussian kernel). paper, propose use quasi-monte carlo (qmc) approximations instead, relevant integrands evaluated low-discrepancy sequence points opposed random point sets monte carlo approach. derive new discrepancy measure called box discrepancy based theoretical characterizations integration error respect given sequence. propose learn qmc sequences adapted setting based explicit box discrepancy minimization. theoretical analyses complemented empirical results demonstrate effectiveness classical adaptive qmc techniques problem.",18
"building morphological chains agglutinative languages. paper, build morphological chains agglutinative languages using log-linear model morphological segmentation task. model based unsupervised morphological segmentation system called morphochains. extend morphochains log linear model expanding candidate space recursively cover split points agglutinative languages turkish, whereas original model candidates generated considering binary segmentation word. results show improve state-of-art turkish scores 12% f-measure 72% improve english scores 3% f-measure 74%. eventually, system outperforms morphochains well-known unsupervised morphological segmentation systems. results indicate candidate generation plays important role unsupervised log-linear model learned using contrastive estimation negative samples.",3
"identification conditional interventional distributions. subject paper elucidation effects actions causal assumptions represented directed graph, statistical knowledge given probability distribution. particular, interested predicting conditional distributions resulting performing action set variables and, subsequently, taking measurements another set. provide necessary sufficient graphical condition cases distributions uniquely computed available information, well algorithm performs computation whenever condition holds. furthermore, use results prove completeness do-calculus [pearl, 1995] identification problem.",3
"mass testing characterization 20-inch pmts juno. main goal juno experiment determine neutrino mass ordering using 20kt liquid-scintillator detector. key feature excellent energy resolution least 3 % 1 mev, instruments need meet certain quality thus fully characterized. 20,000 20-inch pmts received assessed juno detailed testing program began 2017 elapsed four years. based mass characterization set specific requirements, good quality accepted pmts could ascertained. paper presents performed testing procedure designed testing systems well statistical characteristics 20-inch pmts intended used juno experiment, covering fifteen performance parameters including photocathode uniformity. constitutes largest sample 20-inch pmts ever produced studied detail date, i.e. 15,000 newly developed 20-inch mcp-pmts northern night vision technology co. (nnvt) 5,000 dynode pmts hamamatsu photonics k. k.(hpk).",14
"regularization nonlinearities neural language models: needed?. neural language models (lms) based recurrent neural networks (rnn) successful word character-level lms. work well, particular better linear neural lms? possible explanations rnns implicitly better regularization rnns higher capacity storing patterns due nonlinearities both. argue first explanation limit little training data second explanation large amounts text data. show state-of-the-art performance popular small penn dataset rnn lms regularized random dropout. nonetheless, show even better performance simplified, much less expressive linear rnn model without off-diagonal entries recurrent matrix. call model impulse-response lm (irlm). using random dropout, column normalization annealed learning rates, irlms develop neurons keep memory 50 words past achieve perplexity 102.5 penn dataset. two large datasets however, regularization methods unsuccessful models rnn's expressivity allows overtake irlm 10 20 percent perplexity, respectively. despite perplexity gap, irlms still outperform rnns microsoft research sentence completion (mrsc) task. develop slightly modified irlm separates long-context units (lcus) short-context units show lcus alone achieve state-of-the-art performance mrsc task 60.8%. analysis indicates fruitful direction research neural lms lies developing accessible internal representations, suggests optimization regime high momentum terms effectively training models.",18
"equivalence martingale tail bounds deterministic regret inequalities. study equivalence (i) deterministic pathwise statements appearing online learning literature (termed \emph{regret bounds}), (ii) high-probability tail bounds supremum collection martingales (of specific form arising uniform laws large numbers martingales), (iii) in-expectation bounds supremum. virtue equivalence, prove exponential tail bounds norms banach space valued martingales via deterministic regret bounds online mirror descent algorithm adaptive step size. extend results beyond linear structure banach space: define notion \emph{martingale type} general classes real-valued functions show equivalence (up logarithmic factor) various sequential complexities class (in particular, sequential rademacher complexity offset version). classes general martingale type 2, exhibit finer notion variation allows partial adaptation function indexing martingale. proof technique rests sequential symmetrization certifying \emph{existence} regret minimization strategies certain online prediction problems.",10
"joint epipolar tracking (jet): simultaneous optimization epipolar geometry feature correspondences. traditionally, pose estimation considered two step problem. first, feature correspondences determined direct comparison image patches, associating feature descriptors. second step, relative pose coordinates corresponding points estimated, often minimizing reprojection error (rpe). rpe optimization based loss function merely aware feature pixel positions underlying image intensities. paper, propose sparse direct method introduces loss function allows simultaneously optimize unscaled relative pose, well set feature correspondences directly considering image intensity values. furthermore, show integrate statistical prior information motion optimization process. constructive inclusion bayesian bias term particularly efficient application cases strongly predictable (short term) dynamic, e.g. driving scenario. experiments, demonstrate jet algorithm propose outperforms classical reprojection error optimization two synthetic datasets kitti dataset. jet algorithm runs real-time single cpu thread.",3
"linear framework region-based image segmentation inpainting involving curvature penalization. present first method handle curvature regularity region-based image segmentation inpainting independent initialization. end start new formulation length-based optimization schemes, based surface continuation constraints, discuss connections existing schemes. formulation based \emph{cell complex} considers basic regions boundary elements. corresponding optimization problem cast integer linear program. show method extended include curvature regularity, cast integer linear program. here, considering pairs boundary elements reflect curvature. moreover, constraint set derived ensure boundary variables indeed reflect boundary regions described region variables. show solving linear programming relaxation one gets quite close global optimum, curvature regularity indeed much better suited presence long thin objects compared standard length regularity.",3
"n-dimensional approach towards object based classification remotely sensed imagery. remote sensing techniques widely used land cover classification urban analysis. availability high resolution remote sensing imagery limits level classification accuracy attainable pixel-based approach. paper object-based classification scheme based hierarchical support vector machine introduced. combining spatial spectral information, amount overlap classes decreased; thereby yielding higher classification accuracy accurate land cover maps. adopted certain automatic approaches based advanced techniques cellular automata genetic algorithm kernel tuning parameter selection. performance evaluation proposed methodology comparison existing approaches performed reference bhopal city study area.",3
"algorithm aligning sentences bilingual corpora using lexical information. paper describe algorithm aligning sentences translations bilingual corpus using lexical information languages. existing efficient algorithms ignore word identities consider sentence lengths (brown, 1991; gale church, 1993). sentence source language text, proposed algorithm picks likely translation target language text using lexical information certain heuristics. statistical analysis using sentence lengths. algorithm language independent. also aids detecting addition deletion text translations. algorithm gives comparable results existing algorithms cases better cases statistical algorithms give good results.",3
"enriched degree wronski. given $mp$ different $p$-planes general position $(m+p)$-dimensional space, classical problem ask many $p$-planes intersect them. example $m = p = 2$, precisely question ""lines meeting four lines 3-space"" projectivizing. brouwer degree wronski map provides answer general question, first computed schubert complex numbers eremenko gabrielov reals. provide enriched degree wronski $m$ $p$ even, valued grothendieck-witt ring field, using machinery $\mathbf{a}^1$-homotopy theory. demonstrate parities local contribution $m$-plane determinantal relationship certain pl\""ucker coordinates $p$-planes intersects.",10
"stochastic inference spiking neurons high-conductance state. highly variable dynamics neocortical circuits observed vivo hypothesized represent signature ongoing stochastic inference stand apparent contrast deterministic response neurons measured vitro. based propagation membrane autocorrelation across spike bursts, provide analytical derivation neural activation function holds large parameter space, including high-conductance state. basis, show ensemble leaky integrate-and-fire neurons conductance-based synapses embedded spiking environment attain correct firing statistics sampling well-defined target distribution. recurrent networks, examine convergence toward stationarity computer simulations demonstrate sample-based bayesian inference mixed graphical model. points new computational role high-conductance states establishes rigorous link deterministic neuron models functional stochastic dynamics network level.",15
"sensor validation using dynamic belief networks. trajectory robot monitored restricted dynamic environment using light beam sensor data. dynamic belief network (dbn), based discrete model domain, provides discrete monitoring analogous conventional quantitative filter techniques. sensor observations added basic dbn form specific evidence. however, sensor data often partially totally incorrect. show basic dbn, infers impossible combination evidence, may modified handle specific types incorrect data may occur domain. present extension dbn, addition invalidating node, models status sensor working defective. node provides qualitative explanation inconsistent data: caused defective sensor. connection successive instances invalidating node models status sensor time, allowing dbn handle persistent intermittent faults.",3
"convex parametrization new class universal kernel functions use kernel learning. propose new class universal kernel functions admit linear parametrization using positive semidefinite matrices. kernels generalizations sobolev kernel defined piecewise-polynomial functions. class kernels termed ""tessellated"" resulting discriminant defined piecewise hyper-rectangular domains whose corners determined training data. kernels scalable complexity, instance universal sense hypothesis space dense $l_2$. using numerical testing, show soft margin svm, class eliminate need gaussian kernels. furthermore, demonstrate ratio number training data features high, method significantly outperform kernel learning algorithms. finally, reduce complexity associated sdp-based kernel learning methods, use randomized basis positive matrices integrate existing multiple kernel learning algorithms simplemkl.",18
"stag: stable fiducial marker system. paper, propose stag, fiducial marker system provides stable pose estimation. outer square border marker used detection pose estimation. followed novel pose refinement step using inner circular border. refined pose stable robust across viewing conditions compared state art. addition, lexicographic generation algorithm adapted fiducial markers, libraries various sizes created. makes system suitable applications require many unique markers, unique markers high occlusion resistance. edge segment-based detection algorithm low complexity, returns false candidates. features demonstrated experiments real images, including comparisons state art fiducial marker systems.",3
"ramified covers abelian varieties torsion fields. study rational points ramified covers abelian varieties certain infinite galois extensions $\mathbb{q}$. particular, prove every elliptic curve $e$ $\mathbb{q}$ weak hilbert property corvaja-zannier maximal abelian extension $\mathbb{q}^{\rm ab}$ $\mathbb{q}$, field $\mathbb{q}(a_{\rm tor})$ obtained adjoining $\mathbb{q}$ torsion points abelian variety $a$ $\mathbb{q}$.",10
"markov decision processes continuous side information. consider reinforcement learning (rl) setting agent interacts sequence episodic mdps. start episode agent access side-information context determines dynamics mdp episode. setting motivated applications healthcare baseline measurements patient start treatment episode form context may provide information patient might respond treatment decisions. propose algorithms learning contextual markov decision processes (cmdps) assumption unobserved mdp parameters vary smoothly observed context. also give lower upper pac bounds smoothness assumption. lower bound exponential dependence dimension, consider tractable linear setting context used create linear combinations finite set mdps. linear setting, give pac learning algorithm based kwik learning techniques.",18
"double null data characteristic problem general relativity. general hypersurfaces causal character studied abstractly using hypersurface data formalism. null case, write tangential components ambient ricci tensor terms abstract data thus generalizing null structure equations (which recovered appropriate gauge). using formalism, formulate solve completely abstract way characteristic cauchy problem einstein vacuum field equations. initial data detached spacetime notion, fully diffeomorphism gauge covariant. results paper put characteristic problem similar footing standard cauchy problem general relativity.",6
"handwriting trajectory recovery using end-to-end deep encoder-decoder network. paper, introduce novel technique recover pen trajectory offline characters crucial step handwritten character recognition. generally, online acquisition approach advantage offline counterpart online technique keeps track pen movement. hence, pen tip trajectory retrieval offline text bridge gap online offline methods. proposed framework employs sequence sequence model consists encoder-decoder lstm module. encoder module consists convolutional lstm network, takes offline character image input encodes feature sequence hidden representation. output encoder fed decoder lstm get successive coordinate points every time step decoder lstm. although sequence sequence model popular paradigm various computer vision language translation tasks, main contribution work lies designing end-to-end network decade old popular problem document image analysis community. tamil, telugu devanagari characters lipi toolkit dataset used experiments. proposed method achieved superior performance compared conventional approaches.",3
"novel genetic algorithm using helper objectives 0-1 knapsack problem. 0-1 knapsack problem well-known combinatorial optimisation problem. approximation algorithms designed solving return provably good solutions within polynomial time. hand, genetic algorithms well suited solving knapsack problem find reasonably good solutions quickly. naturally arising question whether genetic algorithms able find solutions good approximation algorithms do. paper presents novel multi-objective optimisation genetic algorithm solving 0-1 knapsack problem. experiment results show new algorithm outperforms rivals, greedy algorithm, mixed strategy genetic algorithm, greedy algorithm + mixed strategy genetic algorithm.",3
"versatile? applications sgnmt decoder machine translation. sgnmt decoding platform machine translation allows paring various modern neural models translation different kinds constraints symbolic models. paper, describe three use cases sgnmt currently playing active role: (1) teaching sgnmt used course work student theses mphil machine learning, speech language technology university cambridge, (2) research research work cambridge mt group based sgnmt, (3) technology transfer show sgnmt helping transfer research findings laboratory industry, eg. product sdl plc.",3
parameter-less self-organizing map algorithm. parameter-less self-organizing map (plsom) new neural network algorithm based self-organizing map (som). eliminates need learning rate annealing schemes learning rate neighbourhood size. discuss relative performance plsom som demonstrate tasks som fails plsom performs satisfactory. finally discuss example applications plsom present proof ordering certain limited conditions.,3
"cavs: vertex-centric programming interface dynamic neural networks. recent deep learning (dl) models moved beyond static network architectures dynamic ones, handling data network structure changes every example, sequences variable lengths, trees, graphs. existing dataflow-based programming models dl---both static dynamic declaration---either cannot readily express dynamic models, inefficient due repeated dataflow graph construction processing, difficulties batched execution. present cavs, vertex-centric programming interface optimized system implementation dynamic dl models. cavs represents dynamic network structure static vertex function $\mathcal{f}$ dynamic instance-specific graph $\mathcal{g}$, performs backpropagation scheduling execution $\mathcal{f}$ following dependencies $\mathcal{g}$. cavs bypasses expensive graph construction preprocessing overhead, allows use static graph optimization techniques pre-defined operations $\mathcal{f}$, naturally exposes batched execution opportunities different graphs. experiments comparing cavs two state-of-the-art frameworks dynamic nns (tensorflow fold dynet) demonstrate efficacy approach: cavs achieves near one order magnitude speedup training various dynamic nn architectures, ablations demonstrate contribution proposed batching memory management strategies.",3
"darn: deep adversial residual network intrinsic image decomposition. present new deep supervised learning method intrinsic decomposition single image albedo shading components. contributions based new fully convolutional neural network estimates absolute albedo shading jointly. solution relies single end-to-end deep sequence residual blocks perceptually-motivated metric formed two adversarially trained discriminators. opposed classical intrinsic image decomposition work, fully data-driven, hence require physical priors like shading smoothness albedo sparsity, rely geometric information depth. compared recent deep learning techniques, simplify architecture, making easier build train, constrain generate valid reversible decomposition. rediscuss augment set quantitative metrics account challenging recovery non scale-invariant quantities. train demonstrate architecture publicly available mpi sintel dataset intrinsic image decomposition, show attenuated overfitting issues discuss generalizability data. results show work outperforms state art deep algorithms qualitative quantitative aspect.",3
"collaborative low-rank subspace clustering. paper present collaborative low-rank subspace clustering. given multiple observations phenomenon learn unified representation matrix. unified matrix incorporates features observations, thus increasing discriminative power compared learning representation matrix observation separately. experimental evaluation shows method outperforms subspace clustering separate observations state art collaborative learning algorithm.",3
"causal reasoning graphical time series models. propose definition causality time series terms effect intervention one component multivariate time series another component later point time. conditions identifiability, comparable back-door front-door criteria, presented also verified graphically. computation causal effect derived illustrated linear case.",18
"detecting egregious conversations customers virtual agents. virtual agents becoming prominent channel interaction customer service. customer interactions smooth, however, become almost comically bad. instances, human agent might need step salvage conversation. detecting bad conversations important since disappointing customer service may threaten customer loyalty impact revenue. paper, outline approach detecting egregious conversations, using behavioral cues user, patterns agent responses, user-agent interaction. using logs two commercial systems, show using features improves detection f1-score around 20% using textual features alone. addition, show features common across two quite different domains and, arguably, universal.",3
"new approach quantum computing using magnetic resonance imaging. explosive increase interest quantum computing resulted various proposals generation quantum bits qubits, basic quantum computing unit. superconducting qubits josephson junction widely accepted currently used, ion-trap similar molecule-based qubits proposed recently. methods, qubit generated individually great effort. proposed new technique using magnetic resonance imaging (mri)-based qubit generation, multiple qubits generated. simultaneously provides complete qubit platform quantum computing. central proposed method simultaneous generation multiple qubits using 'gradient' concept together multiple radiofrequency coils, one qubits, others individual qubits small q-coil. another key concept time-encoded probability amplitude (tepa) technique, using individual q-coils together spin-echo series qubit incorporating readout gating time-encoding. mri-based qubit-generation qubit-encoding technique allowed us develop entirely new class quantum computing platform. newly proposed mri-based qubits well-suited currently available electronics, superconducting semiconductor technologies, well nuclear magnetic resonance mri physics technology.",17
"detecting off-topic responses visual prompts. automated methods essay scoring made great progress recent years, achieving accuracies close human annotators. however, known weakness automated scorers taking account semantic relevance submitted text. existing work detecting answer relevance given textual prompt, little previous research done incorporate visual writing prompts. propose neural architecture several extensions detecting off-topic responses visual prompts evaluate dataset texts written language learners.",3
"color constancy derivative colors. information illuminant color well contained achromatic regions specular components highlight regions. paper, propose novel way achieve color constancy exploiting clues. key approach lies use suitably extracted derivative colors, able compute illuminant color robustly kernel density estimation. extracting derivative colors achromatic regions approximate illuminant color well basically straightforward, success extraction highlight regions attributed different rates variation diffuse specular magnitudes dichromatic reflection model. proposed approach requires training phase simple implement. significantly, performs quite satisfactorily inter-database parameter settings. experiments three standard databases demonstrate effectiveness fine performance comparison state-of-the-art methods.",3
"hyper-hue emap hyperspectral images supervised layer decomposition old master drawings. old master drawings mostly created step step several layers using different materials. art historians restorers, examination layers brings various insights artistic work process helps answer questions object, attribution authenticity. however, layers typically overlap oftentimes difficult differentiate unaided eye. example, common layer combination red chalk ink. work, propose image processing pipeline operates hyperspectral images separate layers. using pipeline, show hyperspectral images enable better layer separation rgb images, spectral focus stacking aids layer separation. particular, propose use two descriptors hyperspectral historical document analysis, namely hyper-hue extended multi-attribute profile (emap). comparative results features underline efficacy three proposed improvements.",3
"satellite imagery analysis operational damage assessment emergency situations. major disaster occurs questions raised estimate damage time support decision making process relief efforts local authorities humanitarian teams. paper consider use machine learning computer vision remote sensing imagery improve time efficiency assessment damaged buildings disaster affected area. propose general workflow useful various disaster management applications, demonstrate use proposed workflow assessment damage caused wildfires california 2017.",3
"fast methods denoising matrix completion formulations, applications robust seismic data interpolation. recent svd-free matrix factorization formulations enabled rank minimization systems millions rows columns, paving way matrix completion extremely large-scale applications, seismic data interpolation. paper, consider matrix completion formulations designed hit target data-fitting error level provided user, propose algorithm called lr-bpdn able exploit factorized formulations solve corresponding optimization problem. since practitioners typically strong prior knowledge target error level, innovation makes easy apply algorithm practice, leaving factor rank determined. within established framework, propose two extensions highly relevant solving practical challenges data interpolation. first, propose weighted extension allows known subspace information improve results matrix completion formulations. show weighting used context frequency continuation, essential aspect seismic data interpolation. second, propose matrix completion formulations robust large measurement errors available data. illustrate advantages lr-bpdn collaborative filtering problem using movielens 1m, 10m, netflix 100m datasets. then, use new method, along robust subspace re-weighted extensions, obtain high-quality reconstructions large scale seismic interpolation problems real data, even presence data contamination.",18
"all-optical tunability metalenses permeated liquid crystals. metasurfaces extensively engineered produce wide range optical phenomena, allowing unprecedented control propagation light. however, generally designed single-purpose devices without modifiable post-fabrication optical response, limitation real-world applications. work, report nanostructured planar fused silica metalens permeated nematic liquid crystal (nlc) gold nanoparticle solution. physical properties embedded nlcs manipulated application external stimuli, enabling reconfigurable optical metasurfaces. report all-optical, dynamic control metalens optical response resulting thermo-plasmonic induced changes nlc solution associated nematic-isotropic phase transition. continuous reversible tuning metalens focal length experimentally demonstrated, variation 80 um (0.16% 5 cm nominal focal length) along optical axis. achieved without direct mechanical electrical manipulation device. reconfigurable properties compared corroborating numerical simulations focal length shift exhibit close correspondence.",14
"synthcam3d: semantic understanding synthetic indoor scenes. interested automatic scene understanding geometric cues. end, aim bring semantic segmentation loop real-time reconstruction. semantic segmentation built deep autoencoder stack trained exclusively synthetic depth data generated novel 3d scene library, synthcam3d. importantly, network able segment real world scenes without noise modelling. present encouraging preliminary results.",3
"(semi)-exact hamiltonian curvature perturbation $ζ$. total hamiltonian general relativity, involves first class hamiltonian momentum constraints, weakly vanishes. however, action expanded around classical solution case single scalar field inflationary model, appears non-vanishing hamiltonian additional first class constraints; time theory becomes perturbative number fluctuation fields. show one reorganize expansion solve hamiltonian constraint exactly, yield explicit order action. hand, momentum constraint solved perturbatively tensor modes $\gamma_{ij}$ still keeping curvature perturbation $\zeta$ dependence exact. way, gauge fixing, one obtain semi-exact hamiltonian $\zeta$ gets corrections interactions tensor modes (hence hamiltonian becomes exact tensor perturbations set zero). equations motion clearly exhibit evolution $\zeta$ involves logarithmic time dependence, subtle point debated literature. discuss long wavelength late time limits, obtain simple non-trivial classical solutions $\zeta$ zero-mode.",9
"kernel regression, minimax rates effective dimensionality: beyond regular case. investigate kernel regularization methods achieve minimax convergence rates source condition regularity assumption target function. questions considered past literature, specific assumptions decay, typically polynomial, spectrum kernel mapping covariance operator. perspective distribution-free results, investigate issue much weaker assumption eigenvalue decay, allowing complex behavior reflect different structure data different scales.",18
"applying neuro - computing e-com domain. prior studies generally suggested artificial neural networks (anns) superior conventional statistical models predicting consumer buying behavior. are, however, contradicting findings raise question usefulness anns. paper discusses development three neural networks modeling consumer e-commerce behavior compares findings equivalent logistic regression models. results showed anns predict e-commerce adoption slightly accurately logistic models hardly justifiable given added complexity. further, anns seem highly adaptive, particularly small sample coupled large number nodes hidden layers which, turn, limits neural networks' generalisability.",3
"modelling data dispersion degree automatic robust estimation multivariate gaussian mixture models application noisy speech processing. trimming scheme prefixed cutoff portion known method improving robustness statistical models multivariate gaussian mixture models (mg- mms) small scale tests alleviating impacts outliers. however, method applied real- world data, noisy speech processing, hard know optimal cut-off portion remove outliers sometimes removes useful data samples well. paper, propose new method based measuring dispersion degree (dd) training data avoid problem, realise automatic robust estimation mgmms. dd model studied using two different measures. one, theoretically prove dd data samples context mgmms approximately obeys specific (chi chi-square) distribution. proposed method evaluated real-world application moderately-sized speaker recognition task. experiments show proposed method significantly improve robustness conventional training method gmms speaker recognition.",3
"warm starting bayesian optimization. develop framework warm-starting bayesian optimization, reduces solution time required solve optimization problem one sequence related problems. useful optimizing output stochastic simulator fails provide derivative information, bayesian optimization methods well-suited. solving sequences related optimization problems arises making several business decisions using one optimization model input data collected different time periods markets. many gradient-based methods warm started initiating optimization solution previous problem, warm start approach apply bayesian optimization methods, carry full metamodel objective function iteration iteration. approach builds joint statistical model entire collection related objective functions, uses value information calculation recommend points evaluate.",18
"clifford fourier-mellin transform two real square roots -1 cl(p,q), p+q=2. describe non-commutative generalization complex fourier-mellin transform clifford algebra valued signal functions domain $\r^{p,q}$ taking values cl(p,q), p+q=2. keywords: algebra, fourier transforms; logic, set theory, algebra, fourier analysis, integral transforms",10
"gaia edr3 comparative study protoplanetary disk fractions young stellar clusters. (abridged) lifetime protoplanetary disks around young stars limits timescale planets form. disk dissipation timescale < 10 myr inferred surveys providing fraction stars disks young stellar clusters different ages. however, previous surveys focused compact region within ~ 2 pc clusters' centers, disk fraction information considering outer part practically absent. aim test disk fraction estimates change inferred extended region around clusters' centers. gaia edr3 data best-suited, virtual observatory (vo)-based tool -clusterix-, used identify member stars representative sample 19 young clusters considering two concentric fields view (fov) radii ~ 20 pc ~ 2 pc. analysis reveals inner disk fractions inferred compact extended regions equal within ~ 10%, support previous hypothesis proposing disk fractions significantly larger considering extended regions. list member disk stars cluster provided stored vo-compliant archive. averaged values plots characterizing whole clusters also provided, including hr diagrams based gaia colors absolute magnitudes. results cover largest fields ever probed dealing disk fractions clusters analysed, imply complete characterization requires use wide fovs. resulting database benchmark future detailed studies young clusters, whose disk fractions must accurately determined using multi-wavelength analysis potentially combined data coming gaia releases.",0
"interstitial content detection. interstitial content online content grays out, otherwise obscures main page content. technical report, discuss exploratory research detecting presence interstitial content web pages. discuss use computer vision techniques detect interstitials, potential use techniques provide labelled dataset machine learning.",3
"generalized beta mixtures gaussians. recent years, rich variety shrinkage priors proposed great promise addressing massive regression problems. general, new priors expressed scale mixtures normals, complex forms better properties traditional cauchy double exponential priors. first propose new class normal scale mixtures novel generalized beta distribution encompasses many interesting priors special cases. encompassing framework prove useful comparing competing priors, considering properties revealing close connections. develop class variational bayes approximations new hierarchy presented scale efficiently types truly massive data sets encountered routinely.",18
"local minimax complexity stochastic convex optimization. extend traditional worst-case, minimax analysis stochastic convex optimization introducing localized form minimax complexity individual functions. main result gives function-specific lower upper bounds number stochastic subgradient evaluations needed optimize either function ""hardest local alternative"" given numerical precision. bounds expressed terms localized computational analogue modulus continuity central statistical minimax analysis. show computational modulus continuity explicitly calculated concrete cases, relates curvature function optimum. also prove superefficiency result demonstrates meaningful benchmark, acting computational analogue fisher information statistical estimation. nature practical implications results demonstrated simulations.",18
"natural language inference external knowledge. modeling informal inference natural language challenging. recent availability large annotated data, become feasible train complex models neural networks perform natural language inference (nli), achieved state-of-the-art performance. although exist relatively large annotated data, machines learn knowledge needed perform nli data? not, nli models benefit external knowledge build nli models leverage it? paper, aim answer questions enriching state-of-the-art neural natural language inference models external knowledge. demonstrate proposed models external knowledge improve state art stanford natural language inference (snli) dataset.",3
"comparing hierarchy keywords on-line news portals. tagging on-line content informative keywords widespread phenomenon scientific article repositories blogs on-line news portals. cases, tags given item free words chosen authors independently. therefore, relations among keywords collection news items unknown. however, cases topics concepts described keywords forming latent hierarchy, general topics categories top, specialised ones bottom. apply recent, cooccurrence-based tag hierarchy extraction method sets keywords obtained four different on-line news portals. resulting hierarchies show substantial differences topics rendered important (being top hierarchy) less interest (categorised low hierarchy), also underlying network structure. reveals discrepancies plausible keyword association frameworks studied news portals.",14
"characterising equilibrium logic nested logic programs: reductions complexity. equilibrium logic approach nonmonotonic reasoning extends stable-model answer-set semantics logic programs. particular, includes general case nested logic programs, arbitrary boolean combinations permitted heads bodies rules, special kinds theories. paper, present polynomial reductions main reasoning tasks associated equilibrium logic nested logic programs quantified propositional logic, extension classical propositional logic quantifications atomic formulas permitted. provide reductions decision problems, also central semantical concepts equilibrium logic nested logic programs. particular, encodings map given decision problem formula latter valid precisely case former holds. basic tasks deal consistency problem, brave reasoning, skeptical reasoning. additionally, also provide encodings testing equivalence theories programs different notions equivalence, viz. ordinary, strong, uniform equivalence. considered reasoning tasks, analyse computational complexity give strict complexity bounds.",3
logic programming paradigm prolog. tutorial logic programming prolog appropriate course programming languages students familiar imperative programming.,3
"learning non-maximum suppression. object detectors hugely profited moving towards end-to-end learning paradigm: proposals, features, classifier becoming one neural network improved results two-fold general object detection. one indispensable component non-maximum suppression (nms), post-processing algorithm responsible merging detections belong object. de facto standard nms algorithm still fully hand-crafted, suspiciously simple, -- based greedy clustering fixed distance threshold -- forces trade-off recall precision. propose new network architecture designed perform nms, using boxes score. report experiments person detection pets general object categories coco dataset. approach shows promise providing improved localization occlusion handling.",3
"stackgan: text photo-realistic image synthesis stacked generative adversarial networks. synthesizing high-quality images text descriptions challenging problem computer vision many practical applications. samples generated existing text-to-image approaches roughly reflect meaning given descriptions, fail contain necessary details vivid object parts. paper, propose stacked generative adversarial networks (stackgan) generate 256x256 photo-realistic images conditioned text descriptions. decompose hard problem manageable sub-problems sketch-refinement process. stage-i gan sketches primitive shape colors object based given text description, yielding stage-i low-resolution images. stage-ii gan takes stage-i results text descriptions inputs, generates high-resolution images photo-realistic details. able rectify defects stage-i results add compelling details refinement process. improve diversity synthesized images stabilize training conditional-gan, introduce novel conditioning augmentation technique encourages smoothness latent conditioning manifold. extensive experiments comparisons state-of-the-arts benchmark datasets demonstrate proposed method achieves significant improvements generating photo-realistic images conditioned text descriptions.",3
"reward function survival. obtaining survival strategy (policy) one fundamental problems biological agents. paper, generalize formulation previous research related survival agent formulate survival problem maximization multi-step survival probability future time steps. introduce method converting maximization multi-step survival probability classical reinforcement learning problem. using conversion, reward function (negative temporal cost function) expressed log temporal survival probability. show objective function reinforcement learning sense proportional variational lower bound original problem. finally, empirically demonstrate agent learns survival behavior using reward function introduced paper.",3
"benchmarking quality diffusion-weighted images. present novel method allows measuring quality diffusion-weighted mr images dependent image resolution image noise. purpose, introduce new thresholding technique noise signal automatically estimated single data set. thus, user interaction well double acquisition technique, requires time-consuming proper geometrical registration, needed. coarser image resolution slice thickness leads higher signal-to-noise ratio (snr), benchmark determines resolution-independent quality measure images different resolutions adequately compared. evaluate method, set diffusion-weighted images different vendors used. shown quality efficiently determined automatically computed snr comparable snr measured manually manually selected region interest.",3
"improving ecological niche models data mining large environmental datasets surrogate models. whywhere new ecological niche modeling (enm) algorithm mapping explaining distribution species. algorithm uses image processing methods efficiently sift large amounts data find variables best predict species occurrence. purpose paper describe justify main parameterizations show preliminary success rapidly providing accurate, scalable, simple enms. preliminary results 6 species plants animals different regions indicate significant (p<0.01) 14% increase accuracy garp algorithm using models few, typically two, variables. increase attributed access additional data, particularly monthly vs. annual climate averages. whywhere also 6 times faster garp large data sets. data mining based approach transparent access remote data archives new paradigm enm, particularly suited finding correlates large databases fine resolution surfaces. software whywhere freely available, service desktop downloadable form web site http://biodi.sdsc.edu/ww_home.html.",15
"automated scalable bayesian inference via hilbert coresets. automation posterior inference bayesian data analysis enabled experts nonexperts alike use sophisticated models, engage faster exploratory modeling analysis, ensure experimental reproducibility. however, standard automated posterior inference algorithms tractable scale massive modern datasets, modifications make typically model-specific, require expert tuning, break theoretical guarantees inferential quality. building bayesian coresets framework, work instead takes advantage data redundancy shrink dataset preprocessing step, providing fully-automated, scalable bayesian inference theoretical guarantees. begin intuitive reformulation bayesian coreset construction sparse vector sum approximation, demonstrate automation performance-based shortcomings arise use supremum norm. address shortcomings develop hilbert coresets, i.e., bayesian coresets constructed norm induced inner-product log-likelihood function space. propose two hilbert coreset construction algorithms---one based importance sampling, one based frank-wolfe algorithm---along theoretical guarantees approximation quality function coreset size. since exact computation proposed inner-products model-specific, automate construction random finite-dimensional projection log-likelihood functions. resulting automated coreset construction algorithm simple implement, experiments variety models real synthetic datasets show provides high-quality posterior approximations significant reduction computational cost inference.",18
"linear readout object manifolds. objects represented sensory systems continuous manifolds due sensitivity neuronal responses changes physical features location, orientation, intensity. makes certain sensory representations better suited invariant decoding objects downstream networks? present theory characterizes ability linear readout network, perceptron, classify objects variable neural responses. show readout perceptron capacity depends dimensionality, size, shape object manifolds input neural representation.",2
"predictive-state decoders: encoding future recurrent networks. recurrent neural networks (rnns) vital modeling technique rely internal states learned indirectly optimization supervised, unsupervised, reinforcement training loss. rnns used model dynamic processes characterized underlying latent states whose form often unknown, precluding analytic representation inside rnn. predictive-state representation (psr) literature, latent state processes modeled internal state representation directly models distribution future observations, recent work area relied explicitly representing targeting sufficient statistics probability distribution. seek combine advantages rnns psrs augmenting existing state-of-the-art recurrent neural networks predictive-state decoders (psds), add supervision network's internal state representation target predicting future observations. predictive-state decoders simple implement easily incorporated existing training pipelines via additional loss regularization. demonstrate effectiveness psds experimental results three different domains: probabilistic filtering, imitation learning, reinforcement learning. each, method improves statistical performance state-of-the-art recurrent baselines fewer iterations less data.",18
"nonparametric nearest neighbor descent clustering based delaunay triangulation. physically inspired in-tree (it) based clustering algorithm series it, one free parameter involved computing potential value point. work, based delaunay triangulation dual voronoi tessellation, propose nonparametric process compute potential values local information. computation, though nonparametric, relatively rough, consequently, many local extreme points generated. however, unlike gradient-based methods, it-based methods generally insensitive local extremes. positively demonstrates superiority parametric (previous) nonparametric (in work) it-based methods.",18
"new probabilistic interest measures association rules. mining association rules important technique discovering meaningful patterns transaction databases. many different measures interestingness proposed association rules. however, measures fail take probabilistic properties mined data account. paper, start presenting simple probabilistic framework transaction data used simulate transaction data associations present. use data real-world database grocery outlet explore behavior confidence lift, two popular interest measures used rule mining. results show confidence systematically influenced frequency items left hand side rules lift performs poorly filter random noise transaction data. based probabilistic framework develop two new interest measures, hyper-lift hyper-confidence, used filter order mined association rules. new measures show significantly better performance lift applications spurious rules problematic.",3
classification conic bundles rational elliptic surface characteristic. let $x$ rational elliptic surface elliptic fibration $\pi:x\to\bbb{p}^1$ algebraically closed field $k$ characteristic. given conic bundle $\varphi:x\to\bbb{p}^1$ use numerical arguments classify possible fibers $\varphi$ study interplay singular fibers $\pi$ $\varphi$.,10
"electronic dictionary basis nlp tools: greek case. existence dictionary electronic form modern greek (mg) mandatory one process mg morphological syntactic levels since mg highly inflectional language marked stress spelling system many characteristics carried ancient greek. moreover, tool becomes necessary one create efficient sophisticated nlp applications substantial linguistic backing coverage. present paper focus deployment electronic dictionary modern greek, built two phases: first constructed basis spelling correction schema reconstructed order become platform deployment wider spectrum nlp tools.",3
"interpretable graph-based semi-supervised learning via flows. paper, consider interpretability foundational laplacian-based semi-supervised learning approaches graphs. introduce novel flow-based learning framework subsumes foundational approaches additionally provides detailed, transparent, easily understood expression learning process terms graph flows. result, one visualize interactively explore precise subgraph along information labeled nodes flows unlabeled node interest. surprisingly, proposed framework avoids trading accuracy interpretability, fact leads improved prediction accuracy, supported theoretical considerations empirical results. flow-based framework guarantees maximum principle construction handle directed graphs out-of-the-box manner.",18
"optimal approximation piecewise smooth functions using deep relu neural networks. study necessary sufficient complexity relu neural networks-in terms depth number weights-required approximating classifier functions $l^2$-sense. model, consider set $\mathcal{e}^\beta (\mathbb{r}^d)$ possibly discontinuous piecewise $c^\beta$ functions $f : [-1/2, 1/2]^d \to \mathbb{r}$, different 'smooth regions' $f$ separated $c^\beta$ hypersurfaces. given dimension $d \geq 2$, regularity $\beta > 0$, accuracy $\varepsilon > 0$, construct relu neural networks approximate functions $\mathcal{e}^\beta(\mathbb{r}^d)$ $l^2$ error $\varepsilon$. constructed networks fixed number layers, depending $d$ $\beta$ $o(\varepsilon^{-2(d-1)/\beta})$ many nonzero weights, prove optimal. addition optimality terms number weights, show order achieve optimal approximation rate, one needs relu networks certain minimal depth. precisely, piecewise $c^\beta(\mathbb{r}^d)$ functions, minimal depth given-up multiplicative constant-by $\beta/d$. log factor, constructed networks match bound. partly explains benefits depth relu networks showing deep networks necessary achieve efficient approximation (piecewise) smooth functions. finally, analyze approximation high-dimensional spaces function $f$ approximated factorized smooth dimension reducing feature map $\tau$ classifier function $g$-defined low-dimensional feature space-as $f = g \circ \tau$. show case approximation rate depends dimension feature space input dimension.",10
"unified vector space mapping knowledge representation systems. one significant problems inhibits developments areas knowledge representation artificial intelligence problem semantic alignment knowledge mapping. progress solution greatly beneficial advances information retrieval, ontology alignment, relevance calculation, text mining, natural language processing etc. paper concept multidimensional global knowledge map, elaborated unsupervised extraction dependencies large documents corpus, proposed. addition, problem direct human - knowledge representation system interface addressed concept adaptive decoder proposed purpose interaction previously described unified mapping model. combination two approaches suggested basis development new generation knowledge representation systems.",3
3d convolutional neural networks brain tumor segmentation: comparison multi-resolution architectures. paper analyzes use 3d convolutional neural networks brain tumor segmentation mr images. address problem using three different architectures combine fine coarse features obtain final segmentation. compare three different networks use multi-resolution features terms design performance show improve single-resolution counterparts.,18
"anveshak - groundtruth generation tool foreground regions document images. propose graphical user interface based groundtruth generation tool paper. here, annotation input document image done based foreground pixels. foreground pixels grouped together user interaction form labeling units. units labeled user user defined labels. output produced tool image xml file containing metadata information. annotated data used different applications document image analysis.",3
"common graphs arbitrary chromatic number. ramsey's theorem guarantees every graph h 2-edge-coloring sufficiently large complete graph contains monochromatic copy h. 1962, erdos conjectured random 2-edge-coloring minimizes number monochromatic copies k_k, conjecture extended burr rosta graphs. late 1980s, conjectures disproved thomason sidorenko, respectively. classification graphs whose number monochromatic copies minimized random 2-edge-coloring, referred common graphs, remains challenging open problem. sidorenko's conjecture, one significant open problems extremal graph theory, true, every 2-chromatic graph common, fact, 2-chromatic common graph unsettled sidorenko's conjecture known. examples 3-chromatic common graphs known long time, existence 4-chromatic common graph open 2012, common graph larger chromatic number known. construct connected k-chromatic common graphs every k. answers question posed hatami, hladky, kral, norine razborov [combin. probab. comput. 21 (2012), 734-742], problem listed conlon, fox sudakov [london math. soc. lecture note ser. 424 (2015), 49-118, problem 2.28]. also answers stronger form question raised jagger, stovicek thomason [combinatorica 16, (1996), 123-131] whether exists common graph chromatic number least four.",10
"distributional perspective reinforcement learning. paper argue fundamental importance value distribution: distribution random return received reinforcement learning agent. contrast common approach reinforcement learning models expectation return, value. although established body literature studying value distribution, thus far always used specific purpose implementing risk-aware behaviour. begin theoretical results policy evaluation control settings, exposing significant distributional instability latter. use distributional perspective design new algorithm applies bellman's equation learning approximate value distributions. evaluate algorithm using suite games arcade learning environment. obtain state-of-the-art results anecdotal evidence demonstrating importance value distribution approximate reinforcement learning. finally, combine theoretical empirical evidence highlight ways value distribution impacts learning approximate setting.",3
"gaia dr3: specific processing validation all-sky rr lyrae cepheid stars -- cepheid sample. context. cepheids pulsating stars play crucial role several astrophysical contexts. among different types, classical cepheids fundamental tools calibration extragalactic distance ladder. also powerful stellar population tracers context galactic studies. gaia third data release (dr3) publishes improved data cepheids collected initial 34 months operations. aims. present gaia dr3 catalogue cepheids types, obtained analysis carried specific object study (sos) cep&rrl pipeline. methods. discuss procedures adopted clean cepheid sample spurious objects, validate results, re-classify sources wrong outcome sos cep&rrl pipeline. results. gaia dr3 includes multi-band time-series photometry characterisation sos cep&rrl pipeline sample 15,006 cepheids types. sample includes 4,663, 4,616, 321 185 pulsators, distributed lmc, smc, m31 m33, respectively, well 5 221 objects remaining sky sub-region includes stars mw field/clusters number small satellites galaxy. among sample, 327 objects known variable stars literature different classification, while, best knowledge, 474 stars reported variable stars therefore likely new cepheids discovered gaia.",0
"randomized approximation gram matrix: exact computation probabilistic bounds. given real matrix n columns, problem approximate gram product aa^t c << n weighted outer products columns a. necessary sufficient conditions exact computation aa^t (in exact arithmetic) c >= rank(a) columns depend right singular vector matrix a. monte-carlo matrix multiplication algorithm drineas et al. samples outer products, present probabilistic bounds 2-norm relative error due randomization. bounds depend stable rank rank a, matrix dimensions. numerical experiments illustrate bounds informative, even stringent success probabilities matrices small dimension. also derive bounds smallest singular value condition number matrices obtained sampling rows orthonormal matrices.",10
"situationally aware options. hierarchical abstractions, also known options -- type temporally extended action (sutton et. al. 1999) enables reinforcement learning agent plan higher level, abstracting away lower-level details. work, learn reusable options whose parameters vary, encouraging different behaviors, based current situation. principle, behaviors include vigor, defence even risk-averseness. examples refer broader context situational awareness (sa). incorporate sa, form vigor, hierarchical rl defining learning situationally aware options probabilistic goal semi-markov decision process (pg-smdp). achieved using situationally aware options (sap) policy gradient algorithm comes theoretical convergence guarantee. learn reusable options different scenarios robocup soccer domain (i.e., winning/losing). options learn execute different levels vigor resulting human-like behaviours `time-wasting' winning scenario. show potential agent exit bad local optima using reusable options robocup. finally, using sap, agent mitigates feature-based model misspecification bottomless pit death domain.",3
"random walks, equidistribution graphical designs. let $g=(v,e)$ $d$-regular graph $n$ vertices let $\mu_0$ probability measure $v$. act moving randomly chosen neighbor leads sequence probability measures supported $v$ given $\mu_{k+1} = d^{-1} \mu_k$, $a$ adjacency matrix $d$ diagonal matrix vertex degrees $g$. ordering eigenvalues $ d^{-1}$ $1 = \lambda_1 \geq |\lambda_2| \geq \dots \geq |\lambda_n| \geq 0$, well-known graphs $|\lambda_2|$ small random walk process converges quickly uniform distribution: initial probability measures $\mu_0$ $k \geq 0$, $$ \sum_{v \in v} \left| \mu_k(v) - \frac{1}{n} \right|^2 \leq \lambda_2^{2k}.$$ one could wonder whether rate improved specific initial probability measures $\mu_0$. show $g$ regular, $1 \leq \ell \leq n$, exists probability measure $\mu_0$ supported $\ell$ vertices $$ \sum_{v \in v} \left| \mu_k(v) - \frac{1}{n} \right|^2 \leq \lambda_{\ell+1}^{2k}.$$ result applications graph sampling problem: show measures good sampling properties reconstructing global averages.",10
"co-pacrr: context-aware neural ir model ad-hoc retrieval. neural ir models, drmm pacrr, achieved strong results successfully capturing relevance matching signals. argue context matching signals also important. intuitively, extracting, modeling, combining matching signals, one would like consider surrounding text (local context) well signals document contribute overall relevance score. work, highlight three potential shortcomings caused considering context information propose three neural ingredients address them: disambiguation component, cascade k-max pooling, shuffling combination layer. incorporating components pacrr model yields co-pacrr, novel context-aware neural ir model. extensive comparisons established models trec web track data confirm proposed model achieve superior search results. addition, ablation analysis conducted gain insights impact interactions different components. release code enable future comparisons.",3
"discovering bayesian market views intelligent asset allocation. along advance opinion mining techniques, public mood found key element stock market prediction. however, manner market participants affected public mood rarely discussed. result, little progress leveraging public mood asset allocation problem, application preferred trusted interpretable way. order address issue incorporating public mood analyzed social media, propose formalize market views integrated modern portfolio theory. framework, optimal market views maximize returns period bayesian asset allocation model. train two neural models generate market views, benchmark performance model using market views popular asset allocation strategies. experimental results suggest formalization market views significantly increases profitability (5% 10%) simulated portfolio given risk level.",16
"learning compose domain-specific transformations data augmentation. data augmentation ubiquitous technique increasing size labeled training sets leveraging task-specific data transformations preserve class labels. often easy domain experts specify individual transformations, constructing tuning sophisticated compositions typically needed achieve state-of-the-art results time-consuming manual task practice. propose method automating process learning generative sequence model user-specified transformation functions using generative adversarial approach. method make use arbitrary, non-deterministic transformation functions, robust misspecified user input, trained unlabeled data. learned transformation model used perform data augmentation end discriminative model. experiments, show efficacy approach image text datasets, achieving improvements 4.0 accuracy points cifar-10, 1.4 f1 points ace relation extraction task, 3.4 accuracy points using domain-specific transformation operations medical imaging dataset compared standard heuristic augmentation approaches.",18
"writing positive/negative-conditional equations conveniently. present convenient notation positive/negative-conditional equations. idea merge rules specifying function using case-, if-, match-, let-expressions. based presented macro-rule-construct, positive/negative-conditional equational specifications written higher level. rewrite system translates macro-rule-constructs positive/negative-conditional equations.",3
"olive oil made olives, baby oil made babies: interpreting noun compounds using paraphrases neural model. automatic interpretation relation constituents noun compound, e.g. olive oil (source) baby oil (purpose) important task many nlp applications. recent approaches typically based either noun-compound representations paraphrases. former initially shown promising results, recent work suggests success stems memorizing single prototypical words relation. explore neural paraphrasing approach demonstrates superior performance memorization possible.",3
"quasi-homography warps image stitching. naturalness warps gaining extensive attentions image stitching. recent warps sphp aanap, use global similarity warps mitigate projective distortion (which enlarges regions), however, necessarily bring perspective distortion (which generates inconsistencies). paper, propose novel quasi-homography warp, effectively balances perspective distortion projective distortion non-overlapping region create natural-looking panorama. approach formulates warp solution bivariate system, perspective distortion projective distortion characterized slope preservation scale linearization respectively. proposed warp relies global homography, thus totally parameter-free. comprehensive experiment shows quasi-homography warp outperforms state-of-the-art warps urban scenes, including homography, autostitch sphp. user study demonstrates wins users' favor, comparing homography sphp.",3
"robust convolutional neural networks adversarial noise. recent studies shown convolutional neural networks (cnns) vulnerable small perturbation input called ""adversarial examples"". work, propose new feedforward cnn improves robustness presence adversarial noise. model uses stochastic additive noise added input image cnn models. proposed model operates conjunction cnn trained either standard adversarial objective function. particular, convolution, max-pooling, relu layers modified benefit noise model. feedforward model parameterized mean variance per pixel simplifies computations makes method scalable deep architecture. cifar-10 imagenet test, proposed model outperforms methods improvement evident difficult classification tasks stronger adversarial noise.",3
"period changes mira variables m16-m17 region. analyzed light curves 165 agb variables, mostly miras, sky area centered m16 m17 (l=16, b=0), using ogle gvs database i_c band. comparison published light curves, derived 50 years earlier p.~maffei using kodak i-n photographic plates, allowed us find significant period changes star. remarkably, stars sample appear substantially changed average luminosity, striking case kz ser. provide better identification three stars: ix ser, nsv 10522, nsv 10326, miras. classify light curves 6 stars, discovered classified maffei, (gl ser, nsv 10271, nsv 10326, nsv 10522, nsv 10677, nsv 10772) five new miras, confirm r crb nature v391 sct. magnitude scale used maffei compared modern i_c one.",0
"frozen-density embedding including environmental effects dirac-kohn-sham theory: implementation based density fitting prototyping techniques. frozen density embedding scheme represents embedding method environmental effects onto given subsystem included representing subsystems making surroundings quantum mechanically, means electron densities. present paper, extend full 4-component relativistic dirac-kohn-sham method, implemented bertha code, include environmental confinement effects fde scheme. implementation enormously facilitated bertha's python api (pybertha), provides flexible framework development using python advantages terms code re-usability, portability facilitating interoperability fde implementations available pyadf framework. computational performance evaluated series gold clusters (au$_n$, n=2,4,8) embedded increasing number water molecules (5, 10, 20, 40 80 water molecules). found procedure scales approximately linearly size frozen surrounding environment (in line underpinnings fde approach) size active system (in line use density fitting). finally, applied code series heavy (rn) super-heavy elements (cn, fl, og) embedded c_60 cage explore confinement effect induced c_60 electronic structure. compare results simulations approximate models employed atomic physics literature, confinement represented radial potential slightly affected nature central atom. results indicate specific interactions described fde able improve upon cruder approximations currently employed, thus provide basis generate realistic radial potentials confined atoms.",14
"inconsistency indices inconsistency axioms pairwise comparisons. pairwise comparisons important tool modern (multiple criteria) decision making. since human judgments often inconsistent, many studies focused ways express measure inconsistency, several inconsistency indices proposed alternative saaty inconsistency index inconsistency ratio reciprocal pairwise comparisons matrices. paper aims to: firstly, introduce new measure inconsistency pairwise comparisons prove basic properties; secondly, postulate additional axiom, upper boundary axiom, existing set axioms; last, least, paper provides proofs satisfaction additional axiom selected inconsistency indices well provides numerical comparison.",3
"efficient delivery policy minimize user traffic consumption guaranteed advertising. work, study guaranteed delivery model widely used online display advertising. guaranteed delivery scenario, ad exposures (which also called impressions works) users guaranteed contracts signed advance advertisers publishers. crucial problem advertising platform fully utilize valuable user traffic generate much possible revenue. different previous works usually minimize penalty unsatisfied contracts cost (e.g. representativeness), propose novel consumption minimization model, primary objective minimize user traffic consumed satisfy contracts. model, develop near optimal method deliver ads users. main advantage method lies consumes nearly least possible user traffic satisfy contracts, therefore contracts accepted produce revenue. also enables publishers estimate much user traffic redundant short sell buy part traffic bulk exchange market. furthermore, robust regard priori knowledge user type distribution. finally, simulation shows method outperforms traditional state-of-the-art methods.",3
"experimental multi-arm pendulum cart: benchmark system chaos, learning, control. single, double, triple pendulum served illustrative experimental benchmark system scientists study dynamical behavior four centuries. pendulum system exhibits wide range interesting behaviors, simple harmonic motion single pendulum chaotic dynamics multi-arm pendulums. forcing, even single pendulum may exhibit chaos, providing simple example damped-driven system. multi-armed pendulums characterized existence index-one saddle points, mediate transport trajectories system, providing simple mechanical analog various complex transport phenomena, biolocomotion transport within solar system. further, pendulum systems long used design test linear nonlinear control strategies, addition arms making problem challenging. work, provide extensive designs construction operation high-performance, multi-link pendulum cart system. although many experimental setups built study behavior pendulum systems, extensive documentation design, construction, operation missing literature. resulting experimental system highly flexible, enabling wide range benchmark problems dynamical systems modeling, system identification learning, control. promote reproducible research, made entire system open-source, including 3d cad drawings, basic tutorial code, data. moreover, discuss possibility extending system capability operated remotely enable researchers around world use it, thus increasing access.",12
"upper bound knots neural networks. neural networks rectified linear unit activations essentially multivariate linear splines. such, one many ways measure ""complexity"" ""expressivity"" neural network count number knots spline model. study number knots fully-connected feedforward neural networks rectified linear unit activation functions. intentionally keep neural networks simple, make theoretical analyses approachable. induction number layers $l$ reveals tight upper bound number knots $\mathbb{r} \to \mathbb{r}^p$ deep neural networks. $n_i \gg 1$ neurons layer $i = 1, \dots, l$, upper bound approximately $n_1 \dots n_l$. show exact upper bound tight, demonstrate upper bound example. purpose analyses pave path understanding behavior general $\mathbb{r}^q \to \mathbb{r}^p$ neural networks.",18
"her2 challenge contest: detailed assessment automated her2 scoring algorithms whole slide images breast cancer tissues. evaluating expression human epidermal growth factor receptor 2 (her2) visual examination immunohistochemistry (ihc) invasive breast cancer (bca) key part diagnostic assessment bca due recognised importance predictive prognostic marker clinical practice. however, visual scoring her2 subjective consequently prone inter-observer variability. given prognostic therapeutic implications her2 scoring, objective method required. paper, report recent automated her2 scoring contest, held conjunction annual pathsoc meeting held nottingham june 2016, aimed systematically comparing advancing state-of-the-art artificial intelligence (ai) based automated methods her2 scoring. contest dataset comprised digitised whole slide images (wsi) sections 86 cases invasive breast carcinoma stained haematoxylin & eosin (h&e) ihc her2. contesting algorithms automatically predicted scores ihc slides unseen subset dataset predicted scores compared 'ground truth' (a consensus score least two experts). also report simple man vs machine contest scoring her2 show automated methods could beat pathology experts contest dataset. paper presents benchmark comparing performance automated algorithms scoring her2. also demonstrates enormous potential automated algorithms assisting pathologist objective ihc scoring.",3
"fairness testing: testing software discrimination. paper defines software fairness discrimination develops testing-based method measuring much software discriminates, focusing causality discriminatory behavior. evidence software discrimination found modern software systems recommend criminal sentences, grant access financial products, determine allowed participate promotions. approach, themis, generates efficient test suites measure discrimination. given schema describing valid system inputs, themis generates discrimination tests automatically require oracle. evaluate themis 20 software systems, 12 come prior work explicit focus avoiding discrimination. find (1) themis effective discovering software discrimination, (2) state-of-the-art techniques removing discrimination algorithms fail many situations, times discriminating much 98% input subdomain, (3) themis optimizations effective producing efficient test suites measuring discrimination, (4) themis efficient systems exhibit discrimination. thus demonstrate fairness testing critical aspect software development cycle domains possible discrimination provide initial tools measuring software discrimination.",3
"efficient supervised learning networks binary synapses. recent experimental studies indicate synaptic changes induced neuronal activity discrete jumps small number stable states. learning systems discrete synapses known computationally hard problem. here, study neurobiologically plausible on-line learning algorithm derives belief propagation algorithms. show performs remarkably well model neuron binary synapses, finite number `hidden' states per synapse, learn random classification task. system able learn number associations close theoretical limit, time sublinear system size. knowledge first on-line algorithm able achieve efficiently finite number patterns learned per binary synapse. furthermore, show performance optimal finite number hidden states becomes small sparse coding. algorithm similar standard `perceptron' learning algorithm, additional rule synaptic transitions occur currently presented pattern `barely correct'. case, synaptic changes meta-plastic (change hidden states actual synaptic state), stabilizing synapse current state. finally, show system two visible states k hidden states much robust noise system k visible states. suggest rule sufficiently simple easily implemented neurobiological systems hardware.",15
multitask learning fine-grained twitter sentiment analysis. traditional sentiment analysis approaches tackle problems like ternary (3-category) fine-grained (5-category) classification learning tasks separately. argue classification tasks correlated propose multitask approach based recurrent neural network benefits jointly learning them. study demonstrates potential multitask models type problems improves state-of-the-art results fine-grained sentiment classification problem.,3
"word sense disambiguation using wsd specific wordnet polysemy words. paper presents new model wordnet used disambiguate correct sense polysemy word based clue words. related words sense polysemy word well single sense word referred clue words. conventional wordnet organizes nouns, verbs, adjectives adverbs together sets synonyms called synsets expressing different concept. contrast structure wordnet, developed new model wordnet organizes different senses polysemy words well single sense words based clue words. clue words sense polysemy word well single sense word used disambiguate correct meaning polysemy word given context using knowledge based word sense disambiguation (wsd) algorithms. clue word noun, verb, adjective adverb.",3
"decomposition uncertainty bayesian deep learning efficient risk-sensitive learning. bayesian neural networks latent variables (bnns+lvs) scalable flexible probabilistic models: account uncertainty estimation network weights and, making use latent variables, capture complex noise patterns data. work, show separate two forms uncertainty decision-making purposes. decomposition allows us successfully identify informative points active learning functions heteroskedastic bimodal noise. also demonstrate decomposition allows us define novel risk-sensitive reinforcement learning criterion identify policies balance expected cost, model-bias noise averseness.",18
"multilingual part-of-speech tagging bidirectional long short-term memory models auxiliary loss. bidirectional long short-term memory (bi-lstm) networks recently proven successful various nlp sequence modeling tasks, little known reliance input representations, target languages, data set size, label noise. address issues evaluate bi-lstms word, character, unicode byte embeddings pos tagging. compare bi-lstms traditional pos taggers across languages data sizes. also present novel bi-lstm model, combines pos tagging loss function auxiliary loss function accounts rare words. model obtains state-of-the-art performance across 22 languages, works especially well morphologically complex languages. analysis suggests bi-lstms less sensitive training data size label corruptions (at small noise levels) previously assumed.",3
"much chemistry deep neural network need know make accurate predictions?. meteoric rise deep learning models computer vision research, achieved human-level accuracy image recognition tasks firm evidence impact representation learning deep neural networks. chemistry domain, recent advances also led development similar cnn models, chemception, trained predict chemical properties using images molecular drawings. work, investigate effects systematically removing adding localized domain-specific information image channels training data. augmenting images 3 additional basic information, without introducing architectural changes, demonstrate augmented chemception (augchemception) outperforms original model prediction toxicity, activity, solvation free energy. then, altering information content images, examining resulting model's performance, also identify two distinct learning patterns predicting toxicity/activity compared solvation free energy. patterns suggest chemception learning tasks manner consistent established knowledge. thus, work demonstrates advanced chemical knowledge pre-requisite deep learning models accurately predict complex chemical properties.",18
"makes place? building bespoke place dependent object detectors robotics. paper enabling robots improve perceptual performance repeated use operating environment, creating local expert detectors fitted places robot moves. leverage concept 'experiences' visual perception robotics, accounting bias data robot sees fitting object detector models particular place. key question seek answer paper simply: define place? build bespoke pedestrian detector models autonomous driving, highlighting necessary trade generalisation model capacity vary extent place fit to. demonstrate sizeable performance gain current state-of-the-art detector using computationally lightweight bespoke place-fitted detector models.",3
"robust spectral detection global structures data learning regularization. spectral methods popular detecting global structures given data represented matrix. however data matrix sparse noisy, classic spectral methods usually fail work, due localization eigenvectors (or singular vectors) induced sparsity noise. work, propose general method solve localization problem learning regularization matrix localized eigenvectors. using matrix perturbation analysis, demonstrate learned regularizations suppress eigenvalues associated localized eigenvectors enable us recover informative eigenvectors representing global structure. show applications method several inference problems: community detection networks, clustering pairwise similarities, rank estimation matrix completion problems. using extensive experiments, illustrate method solves localization problem works theoretical detectability limits different kinds synthetic data. contrast existing spectral algorithms based data matrix, non-backtracking matrix, laplacians rank-one regularizations, perform poorly sparse case noise.",18
"quadratic penalties elastic weight consolidation. elastic weight consolidation (ewc, kirkpatrick et al, 2017) novel algorithm designed safeguard catastrophic forgetting neural networks. ewc seen approximation laplace propagation (eskin et al, 2004), view consistent motivation given kirkpatrick et al (2017). note, present extended derivation covers case two tasks. show quadratic penalties ewc inconsistent derivation might lead double-counting data earlier tasks.",18
"machine learning approximation algorithms high-dimensional fully nonlinear partial differential equations second-order backward stochastic differential equations. high-dimensional partial differential equations (pde) appear number models financial industry, derivative pricing models, credit valuation adjustment (cva) models, portfolio optimization models. pdes applications high-dimensional dimension corresponds number financial assets portfolio. moreover, pdes often fully nonlinear due need incorporate certain nonlinear phenomena model default risks, transaction costs, volatility uncertainty (knightian uncertainty), trading constraints model. high-dimensional fully nonlinear pdes exceedingly difficult solve computational effort standard approximation methods grows exponentially dimension. work propose new method solving high-dimensional fully nonlinear second-order pdes. method particular used sample high-dimensional nonlinear expectations. method based (i) connection fully nonlinear second-order pdes second-order backward stochastic differential equations (2bsdes), (ii) merged formulation pde 2bsde problem, (iii) temporal forward discretization 2bsde spatial approximation via deep neural nets, (iv) stochastic gradient descent-type optimization procedure. numerical results obtained using ${\rm t{\small ensor}f{\small low}}$ ${\rm p{\small ython}}$ illustrate efficiency accuracy method cases $100$-dimensional black-scholes-barenblatt equation, $100$-dimensional hamilton-jacobi-bellman equation, nonlinear expectation $ 100 $-dimensional $ g $-brownian motion.",10
"transfer view-manifold learning similarity perception novel objects. develop model perceptual similarity judgment based re-training deep convolution neural network (dcnn) learns associate different views 3d object capture notion object persistence continuity visual experience. re-training process effectively performs distance metric learning object persistency constraints, modify view-manifold object representations. reduces effective distance representations different views object without compromising distance views different objects, resulting untangling view-manifolds individual objects within category across categories. untangling enables model discriminate recognize objects within category, independent viewpoints. found ability limited trained objects, transfers novel objects trained untrained categories, well variety completely novel artificial synthetic objects. transfer learning suggests modification distance metrics view- manifolds general abstract, likely levels parts, independent specific objects categories experienced training. interestingly, resulting transformation feature representation deep networks found significantly better match human perceptual similarity judgment alexnet, suggesting object persistence could important constraint development perceptual similarity judgment biological neural networks.",3
"peculiar thermal states first-order thermodynamics gravity. context recently proposed first-order thermodynamics scalar-tensor gravity, discuss possibility zero-temperature states equilibrium einstein gravity, including pathological brans-dicke theory, palatini $f(r)$ gravity, cuscuton gravity, non-dynamical scalar fields. formalism extended nordstr\""om gravity, contains scalar degree freedom negative temperature relative general relativity.",6
"nearest prime simplicial complex object recognition. structure representation data distribution plays important role understanding underlying mechanism generating data. paper, propose nearest prime simplicial complex approaches (nsc) utilizing persistent homology capture structures. assuming class represented prime simplicial complex, classify unlabeled samples based nearest projection distances samples simplicial complexes. also extend extrapolation ability complexes projection constraint term. experiments simulated practical datasets indicate compared several published algorithms, proposed nsc approaches achieve promising performance without losing structure representation.",3
"comparison aerosol emissions specific speech tasks. study aerosols droplets emitted oral cavity become increasingly important throughout covid-19 pandemic. studies show particulates emitted speaking generally much smaller compared coughing sneezing. however, recent investigations revealed large enough carry respiratory contagions. although studies shown particulate emissions indeed occur speech, date, little information relative contribution different speech sounds producing particle emissions. study compares airborne aerosol generation participants producing isolated speech sounds: fricative consonants, plosive consonants, vowel sounds. participants produced isolated speech tasks, planar beam laser light, high-speed camera, image software calculated number particulates detected overtime. study compares airborne aerosols emitted human participants distance 2.54 cm laser sheet mouth reveals statistically significant increases particulate counts ambient dust distribution speech sounds. vowel sounds statistically greater consonants, suggesting mouth opening, opposed place vocal tract constriction manner sound production, might primary influence degree particulates become aerosolized speech. results research inform boundary conditions computation models aerosolized particulates speech.",15
"jordan recoverability subcategories modules gentle algebras. gentle algebras form class finite-dimensional algebras introduced i. assem a. skowro\'{n}ski 1980s. modules algebra described string band combinatorics associated gentle quiver work m.c.r. butler c.m. ringel. module naturally associated quiver representation. nilpotent endomorphism quiver representation induces linear transformations vector spaces vertex. generically among nilpotent endomorphisms, well-defined jordan form exists representations. $\mathcal{q} = (q, i)$ finite connected gentle quiver, show characterization vertices $m$ representations consisting direct sums indecomposable representations including $m$ support, subcategory denote $\mathscr{c}_{\mathcal{q},m}$, determined isomorphism invariant.",10
"dataset movie description. descriptive video service (dvs) provides linguistic descriptions movies allows visually impaired people follow movie along peers. descriptions design mainly visual thus naturally form interesting data source computer vision computational linguistics. work propose novel dataset contains transcribed dvs, temporally aligned full length hd movies. addition also collected aligned movie scripts used prior work compare two different sources descriptions. total movie description dataset contains parallel corpus 54,000 sentences video snippets 72 hd movies. characterize dataset benchmarking different approaches generating video descriptions. comparing dvs scripts, find dvs far visual describes precisely shown rather happen according scripts created prior movie production.",3
"efficiently using second order information large l1 regularization problems. propose novel general algorithm lhac efficiently uses second-order information train class large-scale l1-regularized problems. method executes cheap iterations achieving fast local convergence rate exploiting special structure low-rank matrix, constructed via quasi-newton approximation hessian smooth loss function. greedy active-set strategy, based largest violations dual constraints, employed maintain working set iteratively estimates complement optimal active set. allows smaller size subproblems eventually identifies optimal active set. empirical comparisons confirm lhac highly competitive several recently proposed state-of-the-art specialized solvers sparse logistic regression sparse inverse covariance matrix selection.",18
"sufficient necessary causation dual. causation issue philosophic debate since hippocrates. recent work defines actual causation terms pearl/halpern's causality framework, formalizing necessary causes (ijcai'15). inspired causality notions security domain (csf'15), which, perhaps surprisingly, formalize sufficient causes instead. provide explicit relation necessary sufficient causes.",3
"face verification using boosted cross-image features. paper proposes new approach face verification, pair images needs classified belonging person not. problem relatively new well-explored literature. current methods mostly adopt techniques borrowed face recognition, process images pair independently, counter intuitive. contrast, propose extract cross-image features, i.e. features across pair images, which, demonstrate, discriminative similarity dissimilarity faces. features derived popular haar-like features, however, extended handle face verification problem instead face detection. collect large bank cross-image features using filters different sizes, locations, orientations. consequently, use adaboost select weight discriminative features. carried extensive experiments proposed ideas using three standard face verification datasets, obtained promising results outperforming state-of-the-art.",3
"discriminative relational topic models. many scientific engineering fields involve analyzing network data. document networks, relational topic models (rtms) provide probabilistic generative process describe link structure document contents, shown promise predicting network structures discovering latent topic representations. however, existing rtms limitations restricted model expressiveness incapability dealing imbalanced network data. expand scope improve inference accuracy rtms, paper presents three extensions: 1) unlike common link likelihood diagonal weight matrix allows the-same-topic interactions only, generalize use full weight matrix captures pairwise topic interactions applicable asymmetric networks; 2) instead standard bayesian inference, perform regularized bayesian inference (regbayes) regularization parameter deal imbalanced link structure issue common real networks improve discriminative ability learned latent representations; 3) instead variational approximation strict mean-field assumptions, present collapsed gibbs sampling algorithms generalized relational topic models exploring data augmentation without making restricting assumptions. generic regbayes framework, carefully investigate two popular discriminative loss functions, namely, logistic log-loss max-margin hinge loss. experimental results several real network datasets demonstrate significance extensions improving prediction performance, time efficiency dramatically improved simple fast approximation method.",3
"""the sum parts"": joint learning word phrase representations autoencoders. recently, lot effort represent words continuous vector spaces. representations shown capture semantic syntactic information words. however, distributed representations phrases remain challenge. introduce novel model jointly learns word vector representations summation. word representations learnt using word co-occurrence statistical information. embed sequences words (i.e. phrases) different sizes common semantic space, propose average word vector representations. contrast previous methods reported posteriori compositionality aspects simple summation, simultaneously train words sum, keeping maximum information original vectors. evaluate quality word representations several classical word evaluation tasks, introduce novel task evaluate quality phrase representations. distributed representations compete methods learning word representations word evaluations, show give better performance phrase evaluation. representations phrases could interesting many tasks natural language processing.",3
"denoising autoencoders trained minimise binary cross-entropy. denoising autoencoders (daes) powerful deep learning models used feature extraction, data generation network pre-training. daes consist encoder decoder may trained simultaneously minimise loss (function) input reconstruction corrupted version input. two common loss functions used training autoencoders, include mean-squared error (mse) binary cross-entropy (bce). training autoencoders image data natural choice loss function bce, since pixel values may normalised take values [0,1] decoder model may designed generate samples take values (0,1). show theoretically daes trained minimise bce may used take gradient steps data space towards regions high probability data-generating distribution. previously shown daes trained using mse. consequence theory, iterative application trained dae moves data sample regions low probability regions higher probability data-generating distribution. firstly, validate theory showing novel data samples, consistent training data, may synthesised initial data samples random noise. secondly, motivate theory showing initial data samples synthesised via methods may improved via iterative application trained dae initial samples.",3
"linear convergence primal-dual gradient method convex-concave saddle point problems without strong convexity. consider convex-concave saddle point problem $\min_{x}\max_{y} f(x)+y^\top x-g(y)$ $f$ smooth convex $g$ smooth strongly convex. prove coupling matrix $a$ full column rank, vanilla primal-dual gradient method achieve linear convergence even $f$ strongly convex. result generalizes previous work either requires $f$ $g$ quadratic functions requires proximal mappings $f$ $g$. adopt novel analysis technique iteration uses ""ghost"" update reference, show iterates primal-dual gradient method converge ""ghost"" sequence. using technique give analysis primal-dual stochastic variance reduced gradient (svrg) method convex-concave saddle point problems finite-sum structure.",10
"deep reinforcement learning human preferences. sophisticated reinforcement learning (rl) systems interact usefully real-world environments, need communicate complex goals systems. work, explore goals defined terms (non-expert) human preferences pairs trajectory segments. show approach effectively solve complex rl tasks without access reward function, including atari games simulated robot locomotion, providing feedback less one percent agent's interactions environment. reduces cost human oversight far enough practically applied state-of-the-art rl systems. demonstrate flexibility approach, show successfully train complex novel behaviors hour human time. behaviors environments considerably complex previously learned human feedback.",18
"knowledge base relation detection via multi-view matching. relation detection core component knowledge base question answering (kbqa). paper, propose kb relation detection model via multi-view matching utilizes useful information extracted question kb. matching inside view multiple perspectives compare two input texts thoroughly. components designed end-to-end trainable neural network model. experiments simplequestions webqsp yield state-of-the-art results.",3
"scaling multiple-source entity resolution using statistically efficient transfer learning. consider serious, previously-unexplored challenge facing almost approaches scaling entity resolution (er) multiple data sources: prohibitive cost labeling training data supervised learning similarity scores pair sources. exists rich literature describing almost aspects pairwise er, new challenge arising due unprecedented ability acquire store data online sources, features driven er enriched search verticals, uniqueness noisy missing data characteristics source. show real-world synthetic data state-of-the-art techniques, reality heterogeneous sources means number labeled training data must scale quadratically number sources, maintain constant precision/recall. address challenge brand new transfer learning algorithm requires far less training data (or equivalently, achieves superior accuracy data) trained using fast convex optimization. intuition behind approach adaptively share structure learned one scoring problem scoring problems sharing data source common. demonstrate theoretically motivated approach incurs runtime cost maintain constant precision/recall cost labeling increasing linearly number sources.",3
parameter-free model rank polysemantic distribution. model rank polysemantic distribution minimal number fitting parameters offered. ideal case parameter-free description dependence basis one several immediate features distribution possible.,3
"effects intra-detector compton scatter zero-frequency dqe photon-counting ct using edge-on-irradiated silicon detectors. background: edge-on-irradiated silicon detectors currently investigated use photon-counting ct applications. low atomic number silicon leads significant number incident photons compton scattered detector, depositing part energy potentially counted multiple times. although physics compton scatter well established, effects compton interactions detector image quality edge-on-irradiated silicon detector still thoroughly investigated. purpose: investigate explain effects compton scatter zero-frequency dqe photon-counting ct using edge-on-irradiated silicon detectors. methods: extend existing monte carlo model edge-on-irradiated silicon detector develop projection image domain performance metrics pure density pure spectral imaging tasks. show lowest energy threshold detector used effective discriminator primary counts cross-talk caused compton scatter. study developed metrics functions lowest threshold energy. results: density imaging performance decreases monotonically function lowest threshold projection image domains. spectral imaging performance plateau 0 10 kev decreases monotonically thereafter, projection image domain. conclusions: compton interactions contribute significantly density imaging performance edge-on-irradiated silicon detectors. studied detector topology, benefit counting primary compton interactions outweighs penalty multiple counting lower threshold energies. compton interactions also contribute significantly spectral imaging performance measured energies 10 kev.",14
"two timescale stochastic approximation controlled markov noise off-policy temporal difference learning. present first time asymptotic convergence analysis two time-scale stochastic approximation driven `controlled' markov noise. particular, faster slower recursions non-additive controlled markov noise components addition martingale difference noise. analyze asymptotic behavior framework relating limiting differential inclusions time-scales defined terms ergodic occupation measures associated controlled markov processes. finally, present solution off-policy convergence problem temporal difference learning linear function approximation, using results.",10
"partially observable markov decision process recommender systems. report ""recurrent deterioration"" (rd) phenomenon observed online recommender systems. rd phenomenon reflected trend performance degradation recommendation model always trained based users' feedbacks previous recommendations. several reasons recommender systems encounter rd phenomenon, including lack negative training data evolution users' interests, etc. motivated tackle problems causing rd phenomenon, propose pomdp-rec framework, neural-optimized partially observable markov decision process algorithm recommender systems. show pomdp-rec framework effectively uses accumulated historical data real-world recommender systems automatically achieves comparable results models fine-tuned exhaustively domain exports public datasets.",3
"transformation-based learning fast lane. transformation-based learning successfully employed solve many natural language processing problems. achieves state-of-the-art performance many natural language processing tasks overtrain easily. however, serious drawback: training time often intorelably long, especially large corpora often used nlp. paper, present novel realistic method speeding training time transformation-based learner without sacrificing performance. paper compares contrasts training time needed performance achieved modified learner two systems: standard transformation-based learner, ica system \cite{hepple00:tbl}. results experiments show system able achieve significant improvement training time still achieving performance standard transformation-based learner. valuable contribution systems algorithms utilize transformation-based learning part execution.",3
"parameterized complexity analysis bi-level optimisation evolutionary algorithms. bi-level optimisation problems gained increasing interest field combinatorial optimisation recent years. paper, start runtime analysis evolutionary algorithms bi-level optimisation problems. examine two np-hard problems, generalised minimum spanning tree problem (gmst), generalised travelling salesman problem (gtsp) context parameterised complexity. generalised minimum spanning tree problem, analyse two approaches presented hu raidl (2012) respect number clusters distinguish chosen representation possible solutions. results show (1+1) ea working spanning nodes representation fixed-parameter evolutionary algorithm problem, whereas global structure representation enables solve problem fixed-parameter time. present hard instances approach show two approaches highly complementary proving solve other's hard instances efficiently. generalised travelling salesman problem, analyse problem respect number clusters problem instance. results show (1+1) ea working global structure representation fixed-parameter evolutionary algorithm problem.",3
"ligand-based benchmarks measure overfitting rather accuracy. undetected overfitting occur significant redundancies training validation data. describe ave, new measure training-validation redundancy ligand-based classification problems accounts similarity amongst inactive molecules well active. investigated nine widely-used benchmarks virtual screening qsar, show amount ave bias strongly correlates performance ligand-based predictive methods irrespective predicted property, chemical fingerprint, similarity measure, previously-applied unbiasing techniques. therefore, likely previously-reported performance ligand-based methods explained overfitting benchmarks rather good prospective accuracy.",15
"consistency spectral hypergraph partitioning planted partition model. hypergraph partitioning lies heart number problems machine learning network sciences. many algorithms hypergraph partitioning proposed extend standard approaches graph partitioning case hypergraphs. however, theoretical aspects methods seldom received attention literature compared extensive studies guarantees graph partitioning. instance, consistency results spectral graph partitioning stochastic block model well known. paper, present planted partition model sparse random non-uniform hypergraphs generalizes stochastic block model. derive error bound spectral hypergraph partitioning algorithm model using matrix concentration inequalities. best knowledge, first consistency result related partitioning non-uniform hypergraphs.",18
"comparing fifty natural languages twelve genetic languages using word embedding language divergence (weld) quantitative measure language distance. introduce new measure distance languages based word embedding, called word embedding language divergence (weld). weld defined divergence unified similarity distribution words languages. using measure, perform language comparison fifty natural languages twelve genetic languages. natural language dataset collection sentence-aligned parallel corpora bible translations fifty languages spanning variety language families. although use parallel corpora, guarantees content languages, interestingly many cases languages within family cluster together. addition natural languages, perform language comparison coding regions genomes 12 different organisms (4 plants, 6 animals, two human subjects). result confirms significant high-level difference genetic language model humans/animals versus plants. proposed method step toward defining quantitative measure similarity languages, applications languages classification, genre identification, dialect identification, evaluation translations.",3
"qbf solving counterexample-guided expansion. introduce novel generalization counterexample-guided inductive synthesis (cegis) instantiate yield novel, competitive algorithm solving quantified boolean formulas (qbf). current qbf solvers based counterexample-guided expansion use recursive approach scales poorly number quantifier alternations. generalization cegis removes need recursive approach, instantiate yield simple efficient algorithm qbf solving. lastly, research supported competitive, though straightforward, implementation algorithm, making possible study practical impact algorithm design decisions, along various optimizations.",3
"electromagnetic measurement temperature quark-gluon plasma produced central ultrarelativistic nuclear collisions. ultrarelativistic collisions large nuclei produce short-lived plasma deconfined quarks gluons. gev-energy photons detected nuclear collisions, small number emitted directly quark-gluon plasma. use characteristic near-exponential energy spectrum photons estimate maximum temperature plasma, assuming undergoes hydrodynamic expansion. quantify non-exponential corrections photon spectrum resulting characteristic equation state photon emission rate nuclear matter.",13
"panorama viable $f(r)$ gravity dark energy models. work shall study late-time dynamics several $f(r)$ gravity models. appropriately expressing field equations terms redshift statefinder function, shall solve numerically field equations using appropriate physical motivated initial conditions. consider models which, construction, described nearly $r^2$-model early epochs fine tune parameters achieve viability compatibility latest planck constraints late times. models provide unified description inflation dark energy era notably common feature models presence dark energy oscillations. furthermore, show that, contrast general relativistic fluids scalar field descriptions, large spectrum different dark energy physics generated simple $f(r)$ gravity models, varying phantom, nearly de sitter quintessential dark energy eras.",6
"one-shot coresets: case k-clustering. scaling clustering algorithms massive data sets challenging task. recently, several successful approaches based data summarization methods, coresets sketches, proposed. techniques provide provably good small summaries, inherently problem dependent - practitioner commit fixed clustering objective even exploring data. however, one construct small data summaries wide range clustering problems simultaneously? work, affirmatively answer question proposing efficient algorithm constructs one-shot summaries k-clustering problems retaining strong theoretical guarantees.",18
"image denoising using optimally weighted bilateral filters: sure fast approach. bilateral filter known quite effective denoising images corrupted small dosages additive gaussian noise. denoising performance filter, however, known degrade quickly increase noise level. several adaptations filter proposed literature address shortcoming, often substantial computational overhead. paper, report simple pre-processing step substantially improve denoising performance bilateral filter, almost additional cost. modified filter designed robust large noise levels, often tends perform poorly certain noise threshold. get best original modified filter, propose combine weighted fashion, weights chosen minimize (a surrogate of) oracle mean-squared-error (mse). optimally-weighted filter thus guaranteed perform better either component filters terms mse, noise levels. also provide fast algorithm weighted filtering. visual quantitative denoising results standard test images reported demonstrate improvement original filter significant visually terms psnr. moreover, denoising performance optimally-weighted bilateral filter competitive computation-intensive non-local means filter.",3
"universal consistency robustness localized support vector machines. massive amount available data potentially used discover patters machine learning challenge kernel based algorithms respect runtime storage capacities. local approaches might help relieve issues. statistical point view local approaches allow additionally deal different structures data different ways. paper analyses properties localized kernel based, non-parametric statistical machine learning methods, particular support vector machines (svms) methods close them. show locally learnt kernel methods universal consistent. furthermore, give upper bound maxbias order show statistical robustness proposed method.",18
"evolution analysis embodied spiking neural networks reveals task-specific clusters effective networks. elucidating principles underlie computation neural networks currently major research topic interest neuroscience. transfer entropy (te) increasingly used tool bridge gap network structure, function, behavior fmri studies. computational models allow us bridge gap even directly associating individual neuron activity behavior. however, computational models analyzed embodied behaviors employed non-spiking neurons. hand, computational models employ spiking neural networks tend restricted disembodied tasks. show first time artificial evolution te-analysis embodied spiking neural networks perform cognitively-interesting behavior. specifically, evolved agent controlled izhikevich neural network perform visual categorization task. smallest networks capable performing task found repeating evolutionary runs different network sizes. informational analysis best solution revealed task-specific te-network clusters, suggesting within-task homogeneity across-task heterogeneity key behavioral success. moreover, analysis ensemble solutions revealed task-specificity te-network clusters correlated fitness. provides empirically testable hypothesis links network structure behavior.",15
"large eddy simulations reacting non-reacting transcritical fuel sprays using multiphase thermodynamics. accurate simulations high-pressure transcritical fuel sprays essential design optimization next-generation gas turbines, internal combustion engines, liquid propellant rocket engines. important challenging accurate modelling complex real-gas effects high-pressure environments, especially hybrid subcritical-to-supercritical mode evaporation mixing fuel oxidizer. paper, present novel modeling framework high-fidelity simulations reacting non-reacting transcritical fuel sprays. method, high-pressure jet disintegration modeled using diffuse interface method multiphase thermodynamics, combines multi-component real-fluid kinetic caloric state equations vapor-liquid equilibrium calculations order compute thermodynamic properties mixture transcritical pressures. multiphase thermodynamic formulations presented general cubic state equations coupled rapid phase-equilibrium calculation method. proposed method represents multiphase turbulent fluid flows transcritical pressures without relying semi-empirical break-up evaporation models. combustion source terms evaluated using finite-rate chemistry model, including real-gas effects based fugacity species mixture. adaptive local deconvolution method (aldm) used physically consistent turbulence model large-eddy simulation (les). les results show good agreement available experimental data reacting non-reacting ecn spray transcritical operating conditions.",14
"complete recursive feature theory. various feature descriptions employed logic programming languages constrained-based grammar formalisms. common notational primitive descriptions functional attributes called features. descriptions considered paper possibly quantified first-order formulae obtained signature binary unary predicates called features sorts, respectively. establish first-order theory ft means three axiom schemes, show completeness, construct three elementarily equivalent models. one models consists so-called feature graphs, data structure common computational linguistics. two models consist so-called feature trees, record-like data structure generalizing trees corresponding first-order terms. completeness proof exhibits terminating simplification system deciding validity satisfiability possibly quantified feature descriptions.",1
"adaptive stochastic resource control: machine learning approach. paper investigates stochastic resource allocation problems scarce, reusable resources non-preemtive, time-dependent, interconnected tasks. approach natural generalization several standard resource management problems, scheduling transportation problems. first, reactive solutions considered defined control policies suitably reformulated markov decision processes (mdps). argue reformulation several favorable properties, finite state action spaces, aperiodic, hence policies proper space control policies safely restricted. next, approximate dynamic programming (adp) methods, fitted q-learning, suggested computing efficient control policy. order compactly maintain cost-to-go function, two representations studied: hash tables support vector regression (svr), particularly, nu-svrs. several additional improvements, application limited-lookahead rollout algorithms initial phases, action space decomposition, task clustering distributed sampling investigated, too. finally, experimental results benchmark industry-related data presented.",3
"cross-correlation thermal sunyaev-zel'dovich effect projected galaxy density field. present joint analysis power spectra planck compton $y$-parameter map projected galaxy density field using wise all-sky survey. detect statistical correlation wise planck data (g$y$) significance $21.8\,\sigma$. also measure auto-correlation spectrum tsz ($yy$) galaxy density field maps (gg) significance $150\,\sigma$ $88\,\sigma$, respectively. construct halo model use measured correlations $c^{\rm gg}_{\ell}$, $c^{yy}_{\ell}$ $c^{{\rm g}y}_{\ell}$ constrain tsz mass bias $b\equiv m_{500}/m^{\rm tsz}_{500}$. also fit galaxy bias, included explicit redshift multipole dependencies $b_{\rm g}(z,\ell)=b_{\rm g}^0(1+z)^{\alpha}(\ell/\ell_0)^{\beta}$, $\ell_0=117$. obtain constraints $b =1.50{\pm 0.07}\,(\textrm{stat}) \pm{0.34}\,(\textrm{sys})$, i.e. $1-b_{\rm h}=0.67\pm 0.03\,({\rm stat})\pm 0.16\,({\rm sys})$ (68% confidence level) hydrostatic mass bias, $b_{\rm g}^0=1.28^{+0.03}_{-0.04}\,(\textrm{stat}) \pm{0.11}\,(\textrm{sys})$, $\alpha=0.20^{+0.11}_{-0.07}\,(\textrm{stat}) \pm{0.10}\,(\textrm{sys})$ $\beta=0.45{\pm 0.01}\,(\textrm{stat}) \pm{0.02}\,(\textrm{sys})$ galaxy bias. incoming data sets future cmb galaxy surveys (e.g. rubin observatory) allow probing large-scale gas distribution detail.",0
"regression model hidden logistic process feature extraction time series. new approach feature extraction time series proposed paper. approach consists specific regression model incorporating discrete hidden logistic process. model parameters estimated maximum likelihood method performed dedicated expectation maximization (em) algorithm. parameters hidden logistic process, inner loop em algorithm, estimated using multi-class iterative reweighted least-squares (irls) algorithm. piecewise regression algorithm iterative variant also considered comparisons. experimental study using simulated real data reveals good performances proposed approach.",18
"product version hilton-milner theorem. two families $\mathcal{f},\mathcal{g}$ $k$-subsets $\{1,2,\ldots,n\}$ called non-trivial cross-intersecting $f\cap g\neq \emptyset$ $f\in \mathcal{f}, g\in \mathcal{g}$ $\cap \{f\colon f\in \mathcal{f}\}=\emptyset=\cap \{g\colon g\in \mathcal{g}\}$. present paper, determine maximum product sizes two non-trivial cross-intersecting families $k$-subsets $\{1,2,\ldots,n\}$ $n\geq 9k$, $k\geq 6$, product version classical hilton-milner theorem.",10
"three faces $u(3)$. $su(3)\times u(1)$ gauge symmetry known phenomenologically problematic unless $u(1)$ factor lives ""dark"" ""hidden"" sector. contrast, $u(3)$ semi-direct product group characterized non-trivial homomorphisms mapping $u(1)$ automorphism group $su(3)$. since semi-direct product includes mutual coupling among gauge field mass-energy, $u(3)$ least chance phenomenologically sensible gauge symmetry outside dark hidden sector. $u(3)$, three different non-trivial homomorphisms induce three separate defining representations. toy model $u(3)$ yang-mills (endowed suitable inner product) coupled massive fermions, renders three distinct covariant derivatives acting single matter field. employing $\mathrm{mod}\,3$ permutation vector space carrying defining representation induced ""large"" gauge transformation, three covariant derivatives one matter field alternatively expressed single covariant derivative acting three distinct species matter fields possessing $u(3)$ quantum numbers different renormalized masses. one interpret three generations matter fields.",14
"first order methods robust non-negative matrix factorization large scale noisy data. nonnegative matrix factorization (nmf) shown identifiable separability assumption, columns(or rows) input data matrix belong convex cone generated columns(or rows) [1]. real applications, however, separability assumption hard satisfy. following [4] [5], paper, look linear programming (lp) based reformulation locate extreme rays convex cone noisy setting. furthermore, order deal large scale data, employ first-order methods (fom) mitigate computational complexity lp, primarily results large number constraints. show performance algorithm real synthetic data sets.",18
"cost-minimising strategies data labelling : optimal stopping active learning. supervised learning deals inference distribution output label space $\cy$ conditioned points observation space $\cx$, given training dataset $d$ pairs $\cx \times \cy$. however, lot applications interest, acquisition large amounts observations easy, process generating labels time-consuming costly. one way deal problem {\em active} learning, points labelled selected aim creating model better performance model trained equal number randomly sampled points. paper, instead propose deal labelling cost directly: learning goal defined minimisation cost function expected model performance total cost labels used. allows development general strategies specific algorithms (a) optimal stopping, expected cost dictates whether label acquisition continue (b) empirical evaluation, cost used performance metric given combination inference, stopping sampling methods. though main focus paper optimal stopping, also aim provide background developments discussion related field active learning.",3
"sum us: strategyproof selection selectors. consider directed graphs set n agents, edge (i,j) taken mean agent supports trusts agent j. given graph integer k\leq n, wish select subset k agents maximizes sum indegrees, i.e., subset k popular trusted agents. time assume individual agent interested selected, may misreport outgoing edges end. problem formulation captures realistic scenarios agents choose among themselves, found context internet search, social networks like twitter, reputation systems like epinions. goal design mechanisms without payments map graph k-subset agents selected satisfy following two constraints: strategyproofness, i.e., agents cannot benefit misreporting outgoing edges, approximate optimality, i.e., sum indegrees selected subset agents always close optimal. first main result surprising impossibility: k \in {1,...,n-1}, deterministic strategyproof mechanism provide finite approximation ratio. second main result randomized strategyproof mechanism approximation ratio bounded four value k, approaches one k grows.",3
"incremental classifier learning generative adversarial networks. paper, address incremental classifier learning problem, suffers catastrophic forgetting. main reason catastrophic forgetting past data available learning. typical approaches keep exemplars past classes use distillation regularization retain classification capability past classes balance past new classes. however, four main problems approaches. first, loss function efficient classification. second, unbalance problem past new classes. third, size pre-decided exemplars usually limited might distinguishable unseen new classes. forth, exemplars may allowed kept long time due privacy regulations. address problems, propose (a) new loss function combine cross-entropy loss distillation loss, (b) simple way estimate remove unbalance old new classes , (c) using generative adversarial networks (gans) generate historical data select representative exemplars generation. believe data generated gans much less privacy issues real images gans directly copy real image patches. evaluate proposed method cifar-100, flower-102, ms-celeb-1m-base datasets extensive experiments demonstrate effectiveness method.",3
"one-shot imitation learning. imitation learning commonly applied solve different tasks isolation. usually requires either careful feature engineering, significant number samples. far desire: ideally, robots able learn demonstrations given task, instantly generalize new situations task, without requiring task-specific engineering. paper, propose meta-learning framework achieving capability, call one-shot imitation learning. specifically, consider setting large set tasks, task many instantiations. example, task could stack blocks table single tower, another task could place blocks table two-block towers, etc. case, different instances task would consist different sets blocks different initial states. training time, algorithm presented pairs demonstrations subset tasks. neural net trained takes input one demonstration current state (which initially initial state demonstration pair), outputs action goal resulting sequence states actions matches closely possible second demonstration. test time, demonstration single instance new task presented, neural net expected perform well new instances new task. use soft attention allows model generalize conditions tasks unseen training data. anticipate training model much greater variety tasks settings, obtain general system turn demonstrations robust policies accomplish overwhelming variety tasks. videos available https://bit.ly/nips2017-oneshot .",3
"neural probabilistic model non-projective mst parsing. paper, propose probabilistic parsing model, defines proper conditional probability distribution non-projective dependency trees given sentence, using neural representations inputs. neural network architecture based bi-directional lstm-cnns benefits word- character-level representations automatically, using combination bidirectional lstm cnn. top neural network, introduce probabilistic structured layer, defining conditional log-linear model non-projective trees. evaluate model 17 different datasets, across 14 different languages. exploiting kirchhoff's matrix-tree theorem (tutte, 1984), partition functions marginals computed efficiently, leading straight-forward end-to-end model training procedure via back-propagation. parser achieves state-of-the-art parsing performance nine datasets.",3
"adaptive statistical non-uniform quantizer detail wavelet components lossy jpeg2000 image compression. paper presents non-uniform quantization method detail components jpeg2000 standard. incorporating fact coefficients lying towards ends histogram plot detail component represent structural information image, quantization step sizes become smaller approach ends histogram plot. variable quantization step sizes determined actual statistics wavelet coefficients. mean standard deviation two statistical parameters used iteratively obtain variable step sizes. moreover, mean coefficients lying within step size chosen quantized value, contrary deadzone uniform quantizer selects midpoint quantization step size quantized value. experimental results deadzone uniform quantizer proposed non-uniform quantizer objectively compared using mean-squared error (mse) mean structural similarity index measure (mssim), evaluate quantization error reconstructed image quality, respectively. subjective analysis reconstructed images also carried out. objective subjective assessments, shown non-uniform quantizer performs better deadzone uniform quantizer perceptual quality reconstructed image, especially low bitrates. importantly, unlike deadzone uniform quantizer, non-uniform quantizer accomplishes better visual quality quantized values.",3
"deep learning real-time gravitational wave detection parameter estimation: results advanced ligo data. recent nobel-prize-winning detections gravitational waves merging black holes subsequent detection collision two neutron stars coincidence electromagnetic observations inaugurated new era multimessenger astrophysics. enhance scope emergent field science, pioneered use deep learning convolutional neural networks, take time-series inputs, rapid detection characterization gravitational wave signals. approach, deep filtering, initially demonstrated using simulated ligo noise. article, present extension deep filtering using real data ligo, detection parameter estimation gravitational waves binary black hole mergers using continuous data streams multiple ligo detectors. demonstrate first time machine learning detect estimate true parameters real events observed ligo. results show deep filtering achieves similar sensitivities lower errors compared matched-filtering far computationally efficient resilient glitches, allowing real-time processing weak time-series signals non-stationary non-gaussian noise minimal resources, also enables detection new classes gravitational wave sources may go unnoticed existing detection algorithms. unified framework data analysis ideally suited enable coincident detection campaigns gravitational waves multimessenger counterparts real-time.",6
"adacomp : adaptive residual gradient compression data-parallel distributed training. highly distributed training deep neural networks (dnns) future compute platforms (offering 100 teraops/s computational capacity) expected severely communication constrained. overcome limitation, new gradient compression techniques needed computationally friendly, applicable wide variety layers seen deep neural networks adaptable variations network architectures well hyper-parameters. paper introduce novel technique - adaptive residual gradient compression (adacomp) scheme. adacomp based localized selection gradient residues automatically tunes compression rate depending local activity. show excellent results wide spectrum state art deep learning models multiple domains (vision, speech, language), datasets (mnist, cifar10, imagenet, bn50, shakespeare), optimizers (sgd momentum, adam) network parameters (number learners, minibatch-size etc.). exploiting sparsity quantization, demonstrate end-to-end compression rates ~200x fully-connected recurrent layers, ~40x convolutional layers, without noticeable degradation model accuracies.",3
"wasserstein distance guided representation learning domain adaptation. domain adaptation aims generalizing high-performance learner target domain via utilizing knowledge distilled source domain different related data distribution. one solution domain adaptation learn domain invariant feature representations learned representations also discriminative prediction. learn representations, domain adaptation frameworks usually include domain invariant representation learning approach measure reduce domain discrepancy, well discriminator classification. inspired wasserstein gan, paper propose novel approach learn domain invariant feature representations, namely wasserstein distance guided representation learning (wdgrl). wdgrl utilizes neural network, denoted domain critic, estimate empirical wasserstein distance source target samples optimizes feature extractor network minimize estimated wasserstein distance adversarial manner. theoretical advantages wasserstein distance domain adaptation lie gradient property promising generalization bound. empirical studies common sentiment image classification adaptation datasets demonstrate proposed wdgrl outperforms state-of-the-art domain invariant representation learning approaches.",18
"adaptive neighborhood graph construction inference multi-relational networks. neighborhood graph, represents instances vertices relations weighted edges, basis many semi-supervised relational models node labeling link prediction. methods employ sequential process construct neighborhood graph. process often consists generating candidate graph, pruning candidate graph make neighborhood graph, performing inference variables (i.e., nodes) neighborhood graph. paper, propose framework dynamically adapt neighborhood graph based states variables intermediate inference results, well structural properties relations connecting them. key strength framework ability handle multi-relational data employ varying amounts relations instance based intermediate inference results. formulate link prediction task inference neighborhood graphs, include preliminary results illustrating effects different strategies proposed framework.",3
"accelerated stochastic power iteration. principal component analysis (pca) one powerful tools machine learning. simplest method pca, power iteration, requires $\mathcal o(1/\delta)$ full-data passes recover principal component matrix eigen-gap $\delta$. lanczos, significantly complex method, achieves accelerated rate $\mathcal o(1/\sqrt{\delta})$ passes. modern applications, however, motivate methods ingest subset available data, known stochastic setting. online stochastic setting, simple algorithms like oja's iteration achieve optimal sample complexity $\mathcal o(\sigma^2/\delta^2)$. unfortunately, fully sequential, also require $\mathcal o(\sigma^2/\delta^2)$ iterations, far $\mathcal o(1/\sqrt{\delta})$ rate lanczos. propose simple variant power iteration added momentum term, achieves optimal sample iteration complexity. full-pass setting, standard analysis shows momentum achieves accelerated rate, $\mathcal o(1/\sqrt{\delta})$. demonstrate empirically naively applying momentum stochastic method, result acceleration. perform novel, tight variance analysis reveals ""breaking-point variance"" beyond acceleration occur. combining insight modern variance reduction techniques, construct stochastic pca algorithms, online offline setting, achieve accelerated iteration complexity $\mathcal o(1/\sqrt{\delta})$. due embarassingly parallel nature methods, acceleration translates directly wall-clock time deployed parallel environment. approach general, applies many non-convex optimization problems accelerated using technique.",10
"mass volume curves anomaly ranking. paper aims formulating issue ranking multivariate unlabeled observations depending degree abnormality unsupervised statistical learning task. 1-d situation, problem usually tackled means tail estimation techniques: univariate observations viewed 'abnormal' located far tail(s) underlying probability distribution. would desirable well dispose scalar valued 'scoring' function allowing comparing degree abnormality mul-tivariate observations. formulate issue scoring anomalies m-estimation problem means novel functional performance criterion, referred mass volume curve (mv curve short), whose optimal elements strictly increasing transforms density. first study statistical estimation mv curve given scoring function provide strategy build confidence regions using smoothed bootstrap approach. optimization functional criterion set piecewise constant scoring functions next tackled. boils estimating sequence empirical minimum volume sets whose levels chosen adaptively data, adjust variations optimal mv curve, controling bias approximation stepwise curve. generalization bounds established difference sup norm mv curve empirical scoring function thus obtained optimal mv curve.",18
"recovery missing samples using sparse approximation via convex similarity measure. paper, study missing sample recovery problem using methods based sparse approximation. regard, investigate algorithms used solving inverse problem associated restoration missed samples image signal. problem also known inpainting context image processing purpose, suggest iterative sparse recovery algorithm based constrained $l_1$-norm minimization new fidelity metric. proposed metric called convex similarity (csim) index, simplified version structural similarity (ssim) index, convex error-sensitive. optimization problem incorporating criterion, solved via alternating direction method multipliers (admm). simulation results show efficiency proposed method missing sample recovery 1d patch vectors inpainting 2d image signals.",18
"statistical deep learning spatial spatio-temporal data. deep neural network models become ubiquitous recent years, applied nearly areas science, engineering, industry. models particularly useful data strong dependencies space (e.g., images) time (e.g., sequences). indeed, deep models also extensively used statistical community model spatial spatio-temporal data through, example, use multi-level bayesian hierarchical models deep gaussian processes. review, first present overview traditional statistical machine learning perspectives modeling spatial spatio-temporal data, focus variety hybrid models recently developed latent process, data, parameter specifications. hybrid models integrate statistical modeling ideas deep neural network models order take advantage strengths modeling paradigm. conclude giving overview computational technologies proven useful hybrid models, brief discussion future research directions.",18
"phase-sensitive method filtering sphere. paper examines filtering sphere, first examining roles spherical harmonic magnitude phase. show phase important magnitude determining structure spherical function. examine properties linear phase shifts spherical harmonic domain, suggest mechanism constructing finite-impulse-response (fir) filters. show filters desirable properties, associative, mapping spherical functions spherical functions, allowing directional filtering, defined relatively simple equations. provide examples filters spherical manifold data.",10
"compiling language models linguistically motivated unification grammar. systems exist able compile unification grammars language models included speech recognizer, far unclear whether non-trivial linguistically principled grammars used purpose. describe series experiments investigate question empirically, incrementally constructing grammar discovering problems emerge successively larger versions compiled finite state graph representations used language models medium-vocabulary recognition task.",3
"sistem pendukung keputusan kelayakan tki ke luar negeri menggunakan fmadm. bp3tki palembang government agencies coordinate, execute selection prospective migrants registration placement. simplify existing procedures improve decision-making necessary build decision support system (dss) determine eligibility employment abroad applying fuzzy multiple attribute decision making (fmadm), using linear sequential systems development methods. system built using microsoft visual basic. net 2010 sql server 2008 database. design system using use case diagrams class diagrams identify needs users systems well systems implementation guidelines. decision support system able rank produce prospective migrants, making easier parties take decision bp3tki workers working country.",3
"autoencoder forest. auto-encoding important task typically realized deep neural networks (dnns) convolutional neural networks (cnn). paper, propose encoderforest (abbrv. eforest), first tree ensemble based auto-encoder. present procedure enabling forests backward reconstruction utilizing equivalent classes defined decision paths trees, demonstrate usage supervised unsupervised setting. experiments show that, compared dnn autoencoders, eforest able obtain lower reconstruction error fast training speed, model reusable damage-tolerable.",3
"labeled directed acyclic graphs: generalization context-specific independence directed graphical models. introduce novel class labeled directed acyclic graph (ldag) models finite sets discrete variables. ldags generalize earlier proposals allowing local structures conditional probability distribution node, unrestricted label sets determine edges deleted underlying directed acyclic graph (dag) given context. several properties models derived, including generalization concept markov equivalence classes. efficient bayesian learning ldags enabled introducing ldag-based factorization dirichlet prior model parameters, marginal likelihood calculated analytically. addition, develop novel prior distribution model structures appropriately penalize model labeling complexity. non-reversible markov chain monte carlo algorithm combined greedy hill climbing approach used illustrating useful properties ldag models real synthetic data sets.",18
"paraboost stereoscopic image quality assessment (pbsiqa) system. problem stereoscopic image quality assessment, finds applications 3d visual content delivery 3dtv, investigated work. specifically, propose new paraboost (parallel-boosting) stereoscopic image quality assessment (pbsiqa) system. system consists two stages. first stage, various distortions classified types, individual quality scorers targeting specific distortion type developed. scorers offer complementary performance face database consisting heterogeneous distortion types. second stage, scores multiple quality scorers fused achieve best overall performance, fuser designed based parallel boosting idea borrowed machine learning. extensive experimental results conducted compare performance proposed pbsiqa system existing stereo image quality assessment (siqa) metrics. developed quality metric serve objective function optimize performance 3d content delivery system.",3
"permutation approach selecting penalty parameter penalized model selection. describe simple, efficient, permutation based procedure selecting penalty parameter lasso. procedure, intended applications variable selection primary focus, applied variety structural settings, including generalized linear models. briefly discuss connections permutation selection existing theory lasso. addition, present simulation study analysis three real data sets permutation selection compared cross-validation (cv), bayesian information criterion (bic), selection method based recently developed testing procedures lasso.",18
"fruit quantity quality estimation using robotic vision system. accurate localisation crop remains highly challenging unstructured environments farms. many developed systems still rely use hand selected features crop identification often neglect estimation crop quantity quality, key assigning labor farming processes. alleviate limitations present robotic vision system accurately estimate quantity quality sweet pepper (capsicum annuum l), key horticultural crop. system consists three parts: detection, quality estimation, tracking. efficient detection achieved using fasterrcnn framework. quality estimated framework learning parallel layer show experimentally results superior performance treating quality extra classes traditional faster-rcnn framework. evaluation two techniques outlines improved performance parallel layer, achieve f1 score 77.3 parallel technique yet 72.5 best scoring (red) multi-class implementation. track crop present tracking via detection approach, uses fasterrcnn parallel layers, also vision-only solution. approach cheap implement requires camera experiments across 2 days show proposed system accurately estimate number sweet pepper present, within 4.1% ground truth.",3
"analysis dropout online learning. deep learning state-of-the-art fields visual object recognition speech recognition. learning uses large number layers huge number units connections. therefore, overfitting serious problem it, dropout kind regularization tool used. however, online learning, effect dropout well known. paper presents investigation effect dropout online learning. analyzed effect dropout convergence speed near singular point. results indicated dropout effective online learning. dropout tends avoid singular point convergence speed near point.",3
"concept drift detection adaptation hierarchical hypothesis testing. streaming environment, often need statistical prediction models detect adapt concept drifts (i.e., changes joint distribution predictor response variables) mitigate deteriorating predictive performance time. various concept drift detection approaches proposed past decades. however, perform well across different concept drift types (e.g., gradual abrupt, recurrent irregular) different data stream distributions (e.g., balanced imbalanced labels). paper presents novel framework detect also adapt various concept drift types, even presence imbalanced data labels. framework leverages hierarchical set hypothesis tests online fashion detect concept drifts employs adaptive training strategy significantly boost adaptation capability. performance proposed framework compared benchmark approaches using simulated real-world datasets spanning breadth concept drift types. proposed approach significantly outperforms benchmark solutions terms precision, delay detection well adaptability across different concepts.",18
"assessment consumers awareness nutrition, food safety hygiene life cycle sustainable development bangladesh- study sylhet city corporation. objective study explore knowledge consumers nutrition, food safety hygiene sampled respondents sylhet city corporation bangladesh terms different demographic characteristics. study includes different types consumers life cycle viz. baby, child, adolescent, young old overall awareness level measured urban area bangladesh. study confined scc area types respondents included findings study used generally bangladesh making policy conducting study population divided six group as, baby, child, adolescent, parental, unmarried adult young married adult matured. find average score awareness food nutrition hygiene unmarried adult higher married adults. study suggested needed increase awareness parents feeding babies. average awareness parents childs eating behavior 5 9 years 3.36 5. awareness around 67 percent careful regard. average awareness adolescent food habit 1.89 three points scales 63 percent only. therefore, consciousness adolescent increase taking food. average feeding styles parents 4.24 5 children 9 years percentage 84 percent.",4
"lcnn: low-level feature embedded cnn salient object detection. paper, propose novel deep neural network framework embedded low-level features (lcnn) salient object detection complex images. utilise advantage convolutional neural networks automatically learn high-level features capture structured information semantic context image. order better adapt cnn model saliency task, redesign network architecture based small-scale datasets. several low-level features extracted, effectively capture contrast spatial information salient regions, incorporated compensate learned high-level features output last fully connected layer. concatenated feature vector fed hinge-loss svm detector joint discriminative learning manner final saliency score region within bounding box obtained linear combination detector's weights. experiments three challenging benchmark (msra-5000, pascal-s, eccsd) demonstrate algorithm effective superior low-level oriented state-of-the-arts terms p-r curves, f-measure mean absolute errors.",3
"piercing boson star black hole. new light fundamental fields natural candidates fraction dark matter. self-gravitating structures fields might common objects universe, could comprise even galactic haloes. structures would interact gravitationally black holes, process utmost importance, since dictates lifetime, black hole motion possible gravitational radiation emission. here, study dynamics black hole piercing much larger fully relativistic boson star, made complex minimally coupled massive scalar without self-interactions. black hole pierces bosonic structure, slowed accretion dynamical friction, giving rise gravitational wave emission. since interested studying interaction large heavy scalar structures, consider mass ratios $q\sim 10$ length ratios ${\cal l} \sim 62$. somewhat surprisingly, simulations, black hole accretes 95% boson star material, even initially small black hole collides large velocity. consequence extreme ""tidal capture"" process, binds black hole boson star together, mass ratios. find evidence ""gravitational atom"" left behind product process.",6
"adaptive algorithms online convex optimization long-term constraints. present adaptive online gradient descent algorithm solve online convex optimization problems long-term constraints , constraints need satisfied accumulated finite number rounds , violated intermediate rounds. user-defined trade-off parameter $\beta$ $\in$ (0, 1), proposed algorithm achieves cumulative regret bounds o(t^max{$\beta$,1--$\beta$}) o(t^(1--$\beta$/2)) loss constraint violations respectively. results hold convex losses handle arbitrary convex constraints without requiring knowledge number rounds advance. contributions improve best known cumulative regret bounds mahdavi, et al. (2012) respectively o(t^1/2) o(t^3/4) general convex domains, respectively o(t^2/3) o(t^2/3) restricting polyhedral domains. supplement analysis experiments validating performance algorithm practice.",18
"methods mapping forest disturbance degradation optical earth observation data: review. purpose review: paper presents review current state art remote sensing based monitoring forest disturbances forest degradation optical earth observation data. part one comprises overview currently available optical remote sensing sensors, used forest disturbance degradation mapping. part two reviews two main categories existing approaches: classical image-to-image change detection time series analysis. recent findings: launch sentinel-2a satellite available landsat imagery, time series analysis become promising also demanding category degradation mapping approaches. four time series classification methods distinguished. methods explained benefits drawbacks discussed. separate chapter presents number recent forest degradation mapping studies two different ecosystems: temperate forests geographical focus europe tropical forests geographical focus africa. summary: review revealed wide variety methods detection forest degradation already available. today, main challenge transfer approaches high resolution time series data multiple sensors. future research also focus classification disturbance types development robust up-scalable methods enable near real time disturbance mapping support operational reactive measures.",3
"generalised object detection semantic analysis: casino example using matlab. matlab version 7.1 used detect playing cards casino table suits ranks cards identified. process gives example application computer vision problem rectangular objects detected information content objects extracted out. case playing cards, suit rank card. image processing system done two passes. pass 1 detects rectangular shapes template matched template left right edges cards. pass 2 extracts suit rank cards matching top left portion card contains rank suit information, stored templates ranks suits playing cards using series if-then statements.",3
"lsh microbatches stochastic gradients: value rearrangement. metric embeddings immensely useful representation interacting entities videos, users, search queries, online resources, words, more. embeddings computed optimizing loss function form sum provided associations relation embedding vectors reflects strength association. moreover, resulting embeddings allow us predict strength unobserved associations. typically, optimization performs stochastic gradient updates minibatches associations arranged independently random. propose study antithesis {\em coordinated} arrangements, obtain efficiently {\em lsh microbatching}, similar associations grouped together. coordinated arrangements leverage similarity entities evident association vectors. experimentally study benefit tunable minibatch arrangements, demonstrating consistent reductions 3-15\% training. arrangement emerges powerful performance knob sgd orthogonal compatible tuning methods, thus candidate wide deployment.",3
"deeper look sarcastic tweets using deep convolutional neural networks. sarcasm detection key task many natural language processing tasks. sentiment analysis, example, sarcasm flip polarity ""apparently positive"" sentence and, hence, negatively affect polarity detection performance. date, approaches sarcasm detection treated task primarily text categorization problem. sarcasm, however, expressed subtle ways requires deeper understanding natural language standard text categorization techniques cannot grasp. work, develop models based pre-trained convolutional neural network extracting sentiment, emotion personality features sarcasm detection. features, along network's baseline features, allow proposed models outperform state art benchmark datasets. also address often ignored generalizability issue classifying data seen models learning phase.",3
"solving l1 regularized least square problem via box-constrained smooth minimization. paper, equivalent smooth minimization l1 regularized least square problem proposed. proposed problem convex box-constrained smooth minimization allows applying fast optimization methods find solution. further, investigated property ""the dual dual primal"" holds l1 regularized least square problem. solver smooth problem proposed, affinity proximal gradient shown. finally, experiments l1 total variation regularized problems performed, corresponding results reported.",10
"fermi two-atom problem: non-perturbative approach via relativistic quantum information algebraic quantum field theory. work revisit famous fermi two-atom problem, concerns relativistic causality impacts atomic transition probabilities, using tools relativistic quantum information (rqi) algebraic quantum field theory (aqft). problem sparked different analyses many directions angles since proposed solution buchholz yngvason (1994). analyses employ various approximations, heuristics, perturbative methods, tends render otherwise useful insights somewhat obscured. also noted studied flat spacetime. show current tools relativistic quantum information, combined algebraic approach quantum field theory, powerful enough provide fuller cleaner analysis fermi two-atom problem arbitrary curved spacetimes completely non-perturbative manner. result gives original solution buchholz yngvason operational reinterpretation terms qubits interacting quantum field, allows various natural generalizations inclusion detector-based local measurement quantum field (phys. rev. 105, 065003).",17
"fast r-cnn. paper proposes fast region-based convolutional network method (fast r-cnn) object detection. fast r-cnn builds previous work efficiently classify object proposals using deep convolutional networks. compared previous work, fast r-cnn employs several innovations improve training testing speed also increasing detection accuracy. fast r-cnn trains deep vgg16 network 9x faster r-cnn, 213x faster test-time, achieves higher map pascal voc 2012. compared sppnet, fast r-cnn trains vgg16 3x faster, tests 10x faster, accurate. fast r-cnn implemented python c++ (using caffe) available open-source mit license https://github.com/rbgirshick/fast-rcnn.",3
"spatial phase-sweep: increasing temporal resolution transient imaging using light source array. transient imaging light-in-flight techniques capture propagation ultra-short pulse light scene, effect captures optical impulse response scene. recently, shown capture transient images using commercially available time-of-flight (tof) systems photonic mixer devices (pmd). paper, propose `spatial phase-sweep', technique exploits speed light increase temporal resolution beyond 100 picosecond limit imposed current electronics. spatial phase-sweep uses linear array light sources spatial separation 3 mm them, thereby resulting time shift 10 picoseconds, translates 100 gfps transient imaging theory. demonstrate prototype transient imaging results using spatial phase-sweep.",3
"structured signal recovery quadratic measurements: breaking sample complexity barriers via nonconvex optimization. paper concerns problem recovering unknown structured signal $x \in r^n$ $m$ quadratic measurements form $y_r=|<a_r,x>|^2$ $r=1,2,...,m$. focus under-determined setting number measurements significantly smaller dimension signal ($m<<n$). formulate recovery problem nonconvex optimization problem prior structural information signal enforced constrains optimization variables. prove projected gradient descent, initialized neighborhood desired signal, converges unknown signal linear rate. results hold constraint set (convex nonconvex) providing convergence guarantees global optimum even objective function constraint set nonconvex. furthermore, results hold number measurements constant factor away minimal number measurements required uniquely identify unknown signal. results provide first provably tractable algorithm data-poor regime, breaking local sample complexity barriers emerged recent literature. companion paper demonstrate favorable properties optimization problem may enable similar results continue hold globally (over entire ambient space). collectively two papers utilize develop powerful tools uniform convergence empirical processes may broader implications rigorous understanding constrained nonconvex optimization heuristics. mathematical results paper also pave way new generation data-driven phase-less imaging systems utilize prior information significantly reduce acquisition time enhance image reconstruction, enabling nano-scale imaging unprecedented speeds resolutions.",3
"deep learning ising model near criticality. well established neural networks deep architectures perform better shallow networks many tasks machine learning. statistical physics, recent interest representing physical data generative modelling, focus shallow neural networks. natural question ask whether deep neural networks hold advantage shallow networks representing data. investigate question using unsupervised, generative graphical models learn probability distribution two-dimensional ising system. deep boltzmann machines, deep belief networks, deep restricted boltzmann networks trained thermal spin configurations system, compared shallow architecture restricted boltzmann machine. benchmark models, focussing accuracy generating energetic observables near phase transition, quantities difficult approximate. interestingly, training generative networks, observe accuracy essentially depends number neurons first hidden layer network, model details network depth model type. evidence shallow networks efficient deep networks representing physical probability distributions associated ising systems near criticality.",2
"robustness analysis hottopixx, linear programming model factoring nonnegative matrices. although nonnegative matrix factorization (nmf) np-hard general, shown recently tractable assumption input nonnegative data matrix close separable (separability requires columns input matrix belongs cone spanned small subset columns). since then, several algorithms designed handle subclass nmf problems. particular, bittorf, recht, r\'e tropp (`factoring nonnegative matrices linear programs', nips 2012) proposed linear programming model, referred hottopixx. paper, provide new general robustness analysis method. particular, design provably robust variant using post-processing strategy allows us deal duplicates near duplicates dataset.",18
"gemini: natural language system spoken-language understanding. gemini natural language understanding system developed spoken language applications. paper describes architecture gemini, paying particular attention resolving tension robustness overgeneration. gemini features broad-coverage unification-based grammar english, fully interleaved syntactic semantic processing all-paths, bottom-up parser, utterance-level parser find interpretations sentences might analyzable complete sentences. gemini also includes novel components recognizing correcting grammatical disfluencies, parse preferences. paper presents component-by-component view gemini, providing detailed relevant measurements size, efficiency, performance.",1
"complete dictionary recovery sphere i: overview geometric picture. consider problem recovering complete (i.e., square invertible) matrix $\mathbf a_0$, $\mathbf \in \mathbb{r}^{n \times p}$ $\mathbf = \mathbf a_0 \mathbf x_0$, provided $\mathbf x_0$ sufficiently sparse. recovery problem central theoretical understanding dictionary learning, seeks sparse representation collection input signals finds numerous applications modern signal processing machine learning. give first efficient algorithm provably recovers $\mathbf a_0$ $\mathbf x_0$ $o(n)$ nonzeros per column, suitable probability model $\mathbf x_0$. contrast, prior results based efficient algorithms either guarantee recovery $\mathbf x_0$ $o(\sqrt{n})$ zeros per column, require multiple rounds sdp relaxation work $\mathbf x_0$ $o(n^{1-\delta})$ nonzeros per column (for constant $\delta \in (0, 1)$). } algorithmic pipeline centers around solving certain nonconvex optimization problem spherical constraint. paper, provide geometric characterization objective landscape. particular, show problem highly structured: high probability, (1) ""spurious"" local minimizers; (2) around saddle points objective negative directional curvature. distinctive structure makes problem amenable efficient optimization algorithms. companion paper (arxiv:1511.04777), design second-order trust-region algorithm sphere provably converges local minimizer arbitrary initializations, despite presence saddle points.",3
"gini-regularized optimal transport application spatio-temporal forecasting. rapidly growing product lines services require finer-granularity forecast considers geographic locales. however open question remains, assess quality spatio-temporal forecast? manuscript introduce metric evaluate spatio-temporal forecasts. metric based opti- mal transport (ot) problem. metric propose constrained ot objec- tive function using gini impurity function regularizer. demonstrate computer experiments qualitative quantitative charac- teristics gini regularized ot problem. moreover, show gini regularized ot problem converges classical ot problem, gini regularized problem considered function {\lambda}, regularization parame-ter. convergence classical ot solution faster state-of-the-art entropic-regularized ot[cuturi, 2013] results numerically stable algorithm.",18
"classical verification quantum depth. present two protocols classical verification quantum depth. protocols allow purely classical verifier distinguish devices different quantum circuit depths even presence classical computation. show device quantum circuit depth rejected verifier even prover applies additional polynomial-time classical computation cheat. hand, verifier accepts device quantum circuit depth d' d'>d. first protocol, introduce additional untrusted quantum machine shares entanglements target machine. applying robust self-test, first protocol certifies depth target machine information theoretic security nearly optimal separation. protocol relies oracle separation problem quantum depth chia, chung lai [stoc 2020] transformation oracle separation problem two-player non-local game. second protocol certifies quantum depth single device based quantum hardness learning errors. protocol relies noisy trapdoor claw-free function family idea pointer chasing force prover keep quantum coherence preceding message exchanges completed. knowledge, give first constructions distinguishing hybrid quantum-classical computers different circuit depths unrelativized models.",17
"collective argumentation. extension abstract argumentation framework, called collective argumentation, introduced attack relation defined directly among sets arguments. extension turns suitable, particular, representing semantics disjunctive logic programs. two special kinds collective argumentation considered opponents share arguments.",3
"learning deep convolutional embeddings face representation using joint sample- set-based supervision. work, investigate several methods strategies learn deep embeddings face recognition, using joint sample- set-based optimization. explain framework expands traditional learning set-based supervision together strategies used maintain set characteristics. we, then, briefly review related set-based loss functions, subsequently propose novel max-margin loss maximizes maximum possible inter-class margin assistance support vector machines (svms). implicitly pushes samples towards correct side margin vector perpendicular hyperplane strength exponentially growing towards negative side hyperplane. show introduced loss outperform previous sample-based set-based ones terms verification faces two commonly used benchmarks.",3
"expoiting syntactic structure language modeling. paper presents language model develops syntactic structure uses extract meaningful information word history, thus enabling use long distance dependencies. model assigns probability every joint sequence words--binary-parse-structure headword annotation operates left-to-right manner --- therefore usable automatic speech recognition. model, probabilistic parameterization, set experiments meant evaluate predictive power presented; improvement standard trigram modeling achieved.",3
"lazily adapted constant kinky inference nonparametric regression model-reference adaptive control. techniques known nonlinear set membership prediction, lipschitz interpolation kinky inference approaches machine learning utilise presupposed lipschitz properties compute inferences unobserved function values. provided bound true best lipschitz constant target function known priori offer convergence guarantees well bounds around predictions. considering general setting builds hoelder continuity relative pseudo-metrics, propose online method estimating hoelder constant online function value observations possibly corrupted bounded observational errors. utilising compute adaptive parameters within kinky inference rule gives rise nonparametric machine learning method, establish strong universal approximation guarantees. is, show prediction rule learn continuous function limit increasingly dense data within worst-case error bound depends level observational uncertainty. apply method context nonparametric model-reference adaptive control (mrac). across range simulated aircraft roll-dynamics performance metrics approach outperforms recently proposed alternatives based gaussian processes rbf-neural networks. discrete-time systems, provide stability guarantees learning-based controllers batch online learning setting.",10
"throughput optimal scheduling policy quantum switch. study quantum switch creates shared end-to-end entangled quantum states multiple sets users connected it. user connected switch via optical link across bipartite bell-state entangled states generated time-slot certain probabilities, switch merges entanglements links create end-to-end entanglements users. one qubit entanglement link stored switch qubit entanglement stored user corresponding link. assuming qubits entanglements links decipher one time-slot, characterize capacity region, defined set arrival rates requests end-to-end entanglements exists scheduling policy stabilizes switch. propose max-weight scheduling policy show stabilizes switch arrival rates lie capacity region. also provide numerical results support analysis.",17
"gaussian process conditional copulas applications financial time series. estimation dependencies multiple variables central problem analysis financial time series. common approach express dependencies terms copula function. typically copula function assumed constant may inaccurate covariates could large influence dependence structure data. account this, bayesian framework estimation conditional copulas proposed. framework parameters copula non-linearly related arbitrary conditioning variables. evaluate ability method predict time-varying dependencies several equities currencies observe consistent performance gains compared static copula models time-varying copula methods.",18
"progress animation ema-controlled tongue model acoustic-visual speech synthesis. present technique animation 3d kinematic tongue model, one component talking head acoustic-visual (av) speech synthesizer. skeletal animation approach adapted make use deformable rig controlled tongue motion capture data obtained electromagnetic articulography (ema), tongue surface extracted volumetric magnetic resonance imaging (mri) data. initial results shown future work outlined.",3
"happiness paradox: friends happier. individuals social networks experience so-called friendship paradox: less popular friends average. effect may explain recent findings widespread social network media use leads reduced happiness. however relation popularity happiness poorly understood. friendship paradox necessarily imply happiness paradox individuals less happy friends. report first direct observation significant happiness paradox large-scale online social network $39,110$ twitter users. results reveal popular individuals indeed happier majority individuals experience significant happiness paradox. magnitude latter effect shaped complex interactions individual popularity, happiness, fact users cluster assortatively level happiness. results indicate topology online social networks distribution happiness populations cause widespread psycho-social effects affect well-being billions individuals.",3
"three-dimensional extinction maps: inverting inter-calibrated extinction catalogues. 3d maps extinction density galaxy built inversion catalogues distance-extinction pairs individual target stars. spatial resolution maps achieved increases spatial density targets, subsequently combination catalogues. however, requires careful inter-calibration. aim develop methods inter-calibration two different catalogues. used reference spectrophotometric catalogue. principal component analysis performed g,gb,gr,j,h,k multi-colour space second catalog. subspace constituted two first components split cells estimated deviations reference. deviations computed using targets reference located short spatial distance secondary target. corrections filtering deduced cell multi-colour space. applied technique two different datasets: one hand, spectrophotometric catalogue, and, hand, catalogue extinctions based photometry gaia edr3 2mass. calibration, dispersion extinction among neighbouring points second catalogue reduced, regardless whether reference targets present locally. weak structures apparent. extinction high galactic latitude targets tightly correlated dust emission, property acquired first catalogue. hierarchical inversion technique applied two merged inter-calibrated catalogues produce 3d extinction density maps corresponding different volumes maximum spatial resolution. maximum resolution 10pc 3000x3000x800pc3 volume around sun, maximum size maps 10x10x0.8 kpc3 resolution 50pc. maps downloaded used means on-line tools.",0
"particle swarm optimization: efficient method tracing periodic orbits 3d galactic potentials. propose particle swarm optimization (pso) alternative method locating periodic orbits three--dimensional (3d) model barred galaxies. develop appropriate scheme transforms problem finding periodic orbits problem detecting global minimizers function, defined poincar\'{e} surface section (pss) hamiltonian system. combining pso method deflection techniques, succeeded tracing systematically several periodic orbits system. method succeeded tracing initial conditions periodic orbits cases newton iterative techniques difficulties. particular, found families 2d 3d periodic orbits associated inner 8:1 12:1 resonances, radial 4:1 corotation resonances 3d ferrers bar model. main advantages proposed algorithm simplicity, ability work using function values solely, well ability locate many periodic orbits per run given jacobian constant.",0
"film: visual reasoning general conditioning layer. introduce general-purpose conditioning method neural networks called film: feature-wise linear modulation. film layers influence neural network computation via simple, feature-wise affine transformation based conditioning information. show film layers highly effective visual reasoning - answering image-related questions require multi-step, high-level process - task proven difficult standard deep learning methods explicitly model reasoning. specifically, show visual reasoning tasks film layers 1) halve state-of-the-art error clevr benchmark, 2) modulate features coherent manner, 3) robust ablations architectural modifications, 4) generalize well challenging, new data examples even zero-shot.",3
"statistics streaming sparse regression. present sparse analogue stochastic gradient descent guaranteed perform well similar conditions lasso. linear regression setup irrepresentable noise features, algorithm recovers support set optimal parameter vector high probability, achieves statistically quasi-optimal rate convergence op(k log(d)/t), k sparsity solution, number features, number training examples. meanwhile, algorithm require computational resources stochastic gradient descent. experiments, find method substantially out-performs existing streaming algorithms real simulated data.",10
"goldilocks principle: reading children's books explicit memory representations. introduce new test well language models capture meaning children's books. unlike standard language modelling benchmarks, distinguishes task predicting syntactic function words predicting lower-frequency words, carry greater semantic content. compare range state-of-the-art models, different way encoding previously read. show models store explicit representations long-term contexts outperform state-of-the-art neural language models predicting semantic content words, although advantage observed syntactic function words. interestingly, find amount text encoded single memory representation highly influential performance: sweet-spot, big small, single words full sentences allows meaningful information text effectively retained recalled. further, attention window-based memories trained effectively self-supervision. assess generality principle applying cnn qa benchmark, involves identifying named entities paraphrased summaries news articles, achieve state-of-the-art performance.",3
"balanced k-means min-cut clustering. clustering effective technique data mining generate groups matter interest. among various clustering approaches, family k-means algorithms min-cut algorithms gain popularity due simplicity efficacy. classical k-means algorithm partitions number data points several subsets iteratively updating clustering centers associated data points. contrast, weighted undirected graph constructed min-cut algorithms partition vertices graph two sets. however, existing clustering algorithms tend cluster minority data points subset, shall avoided target dataset balanced. achieve accurate clustering balanced dataset, propose leverage exclusive lasso k-means min-cut regulate balance degree clustering results. optimizing objective functions build atop exclusive lasso, make clustering result much balanced possible. extensive experiments several large-scale datasets validate advantage proposed algorithms compared state-of-the-art clustering algorithms.",3
"optical magnetism wavefront control arrays strontium atoms. analyzing parameters electronic transitions, show bosonic sr atoms planar optical lattices engineered exhibit optical magnetism higher-order electromagnetic multipoles harnessed wavefront control incident light. resonant $\lambda\simeq 2.6\mu$m light $^3d_1\rightarrow {^3}p_0$ transition mediates cooperative interactions atoms atoms trapped deeply subwavelength optical lattice. atoms exhibit collective excitation eigenmodes, e.g., strong cooperative magnetic response optical frequencies, despite individual atoms negligible coupling magnetic component light. provide detailed scheme utilize excitations cooperative modes consisting arrays electromagnetic multipoles form atomic huygens' surface, complete $2\pi$ phase control transmitted light almost reflection, allowing nearly arbitrary wavefront shaping. numerical examples, achieved controlling atomic level shifts sr off-resonant ${^3p}_j\rightarrow {^3d}_1$ transitions, results simultaneous excitation arrays electric dipoles electric quadrupoles magnetic dipoles. demonstrate wavefront engineering sr array realizing steering incident beam generation baby-skyrmion texture transmitted light via topologically nontrivial transition gaussian beam poincar\'{e} beam, contains possible polarizations single cross-section.",17
"structural identifiability compartmental models infectious disease transmission influenced data type. model identifiability confirmed, inferences infectious disease transmission models may reliable, might lead misleading recommendations. structural identifiability analysis characterizes whether possible obtain unique solutions unknown model parameters, given model structure. work, studied structural identifiability typical deterministic compartmental models infectious disease transmission, focusing influence data type considered model output identifiability unknown model parameters, including initial conditions. defined 26 model versions, unique combination underlying compartmental structure data type(s) considered model output(s). four compartmental model structures three common data types disease surveillance (incidence, prevalence detected vector counts) studied. structural identifiability parameters varied depending type model output. general, models multiple data types outputs structurally identifiable parameters, models single data type output. study highlights importance careful consideration data types integral part inference process compartmental infectious disease transmission models.",15
"conditional plausibility measures bayesian networks. general notion algebraic conditional plausibility measures defined. probability measures, ranking functions, possibility measures, (under appropriate definitions) sets probability measures viewed defining algebraic conditional plausibility measures. shown technology bayesian networks applied algebraic conditional plausibility measures.",3
"coherent bi-directional virtual detector 1-d schrödinger equation. virtual detector commonly utilized technique measure properties wavefunction simulation. one type virtual detector measures probability density current set position time, permitting instantaneous measurement momentum boundary. may used boundary condition quantum classical simulation. however, tool measuring spectra, possesses several problems stemming incoherent nature. another form virtual detector measures wavefunction's complex value set position real space time fourier analyzes produce energy spectrum. spectra produces exact provided wavefunction propagated detector one direction. otherwise produce spectrum includes interference forward backward propagating wavepackets. propose virtual detector maintains benefits coherent virtual detector also able resolve direction propagation mitigate nonphysical interference use second measurement point. show that, continuum limit, bi-directional virtual detector reproduce equivalent wavefunction assuming globally constant potential. therefore equivalent exact spectrum.",17
"rgtsvm: support vector machines gpu r. rgtsvm provides fast flexible support vector machine (svm) implementation r language. distinguishing feature rgtsvm support vector classification support vector regression tasks implemented graphical processing unit (gpu), allowing libraries scale millions examples >100-fold improvement performance existing implementations. nevertheless, rgtsvm retains feature parity interface compatible popular e1071 svm package r. altogether, rgtsvm enables large svm models created experienced novice practitioners.",18
"machine translation using semantic web technologies: survey. large number machine translation approaches developed recently aim migrating content easily across languages. however, literature suggests many obstacles must dealt achieve better automatic translations. central issue machine translation systems must handle ambiguity. promising way overcoming problem using semantic web technologies. article presents results systematic review approaches rely semantic web technologies within machine translation approaches translating texts. overall, survey suggests semantic web technologies enhance quality machine translation outputs various problems, combination still infancy.",3
"optimal transport generative modeling: vegan cookbook. study unsupervised generative modeling terms optimal transport (ot) problem true (but unknown) data distribution $p_x$ latent variable model distribution $p_g$. show ot problem equivalently written terms probabilistic encoders, constrained match posterior prior distributions latent space. relaxed, constrained optimization problem leads penalized optimal transport (pot) objective, efficiently minimized using stochastic gradient descent sampling $p_x$ $p_g$. show pot 2-wasserstein distance coincides objective heuristically employed adversarial auto-encoders (aae) (makhzani et al., 2016), provides first theoretical justification aaes known authors. also compare pot popular techniques like variational auto-encoders (vae) (kingma welling, 2014). theoretical results include (a) better understanding commonly observed blurriness images generated vaes, (b) establishing duality wasserstein gan (arjovsky bottou, 2017) pot 1-wasserstein distance.",18
"joint modeling registration cell populations cohorts high-dimensional flow cytometric data. systems biomedicine, experimenter encounters different potential sources variation data individual samples, multiple experimental conditions, multi-variable network-level responses. multiparametric cytometry, often used analyzing patient samples, issues critical. computational methods identify cell populations individual samples, without ability automatically match across samples, difficult compare characterize populations typical experiments, responding various stimulations distinctive particular patients time-points, especially many samples. joint clustering matching (jcm) multi-level framework simultaneous modeling registration populations across cohort. jcm models every population robust multivariate probability distribution. simultaneously, jcm fits random-effects model construct overall batch template -- used registering populations across samples, classifying new samples. tackling systems-level variation, jcm supports practical biomedical applications involving large cohorts.",18
"point estimators gamma beta distributions. let $x_1,\ldots,x_n$ random sample gamma distribution density $f(x)=\lambda^{\alpha}x^{\alpha-1}e^{-\lambda x}/\gamma(\alpha)$, $x>0$, $\alpha>0$ (the shape parameter) $\lambda>0$ (the reciprocal scale parameter) unknown. main result shows uniformly minimum variance unbiased estimator (umvue) shape parameter, $\alpha$, exists $n\geq 4$; moreover, finite variance $n\geq 6$. precisely, form umvue given parametric functions $\alpha$, $\lambda$, $1/\alpha$ $1/\lambda$. furthermore, highly efficient estimating procedure two-parameter beta distribution also given. based stein-type covariance identity beta distribution, followed application theory $u$-statistics delta-method. msc: primary 62f10; 62f12; secondary 62e15. key words phrases: unbiased estimation; gamma distribution; beta distribution; ye-chen-type closed-form estimators; asymptotic efficiency; $u$-statistics; stein-type covariance identity; delta-method.",10
"intel realsense stereoscopic depth cameras. present comprehensive overview stereoscopic intel realsense rgbd imaging systems. discuss systems' mode-of-operation, functional behavior include models expected performance, shortcomings, limitations. provide information systems' optical characteristics, correlation algorithms, properties affect different applications, including 3d reconstruction gesture recognition. discussion covers intel realsense r200 intel realsense d400 (formally rs400).",3
"active search sparse signals region sensing. autonomous systems used search sparse signals large space; e.g., aerial robots deployed localize threats, detect gas leaks, respond distress calls. intuitively, search algorithms may increase efficiency collecting aggregate measurements summarizing large contiguous regions. however, existing search methods either ignore possibility region observations (e.g., bayesian optimization multi-armed bandits) make strong assumptions sensing mechanism allow measurement arbitrarily encode signals entire environment (e.g., compressive sensing). propose algorithm actively collects data search sparse signals using noisy measurements average values rectangular regions (including single points), based greedy maximization information gain. analyze algorithm 1d show requires $\tilde{o}(\frac{n}{\mu^2}+k^2)$ measurements recover $k$ signal locations small bayes error, $\mu$ $n$ signal strength size search space, respectively. also show active designs fundamentally efficient passive designs region sensing, contrasting results arias-castro, candes, davenport (2013). demonstrate empirical performance algorithm search problem using satellite image data high dimensions.",18
"efficient sparse-winograd convolutional neural networks. convolutional neural networks (cnns) computationally intensive, limits application mobile devices. energy dominated number multiplies needed perform convolutions. winograd's minimal filtering algorithm (lavin, 2015) network pruning (han et al., 2015) reduce operation count, two methods cannot directly combined $-$ applying winograd transform fills sparsity weights activations. propose two modifications winograd-based cnns enable methods exploit sparsity. first, move relu operation winograd domain increase sparsity transformed activations. second, prune weights winograd domain exploit static weight sparsity. models cifar-10, cifar-100 imagenet datasets, method reduces number multiplications $10.4\times$, $6.8\times$ $10.8\times$ respectively loss accuracy less $0.1\%$, outperforming previous baselines $2.0\times$-$3.0\times$. also show moving relu winograd domain allows aggressive pruning.",3
"dialog context language modeling recurrent neural networks. work, propose contextual language models incorporate dialog level discourse information language modeling. previous works contextual language model treat preceding utterances sequence inputs, without considering dialog interactions. design recurrent neural network (rnn) based contextual language models specially track interactions speakers dialog. experiment results switchboard dialog act corpus show proposed model outperforms conventional single turn based rnn language model 3.3% perplexity. proposed models also demonstrate advantageous performance competitive contextual language models.",3
"sparse coding multitask transfer learning. investigate use sparse coding dictionary learning context multitask transfer learning. central assumption learning method tasks parameters well approximated sparse linear combinations atoms dictionary high infinite dimensional space. assumption, together large quantity available data multitask transfer learning settings, allows principled choice dictionary. provide bounds generalization error approach, settings. numerical experiments one synthetic two real datasets show advantage method single task learning, previous method based orthogonal dense representation tasks related method learning task grouping.",3
"learning invariant feature spaces transfer skills reinforcement learning. people learn wide range tasks experience, also learn observing creatures. accelerate acquisition new skills even observed agent differs substantially learning agent terms morphology. paper, examine reinforcement learning algorithms transfer knowledge morphologically different agents (e.g., different robots). introduce problem formulation two agents tasked learning multiple skills sharing information. method uses skills learned agents train invariant feature spaces used transfer skills one agent another. process learning invariant feature spaces viewed kind ""analogy making"", implicit learning partial correspondences two distinct domains. evaluate transfer learning algorithm two simulated robotic manipulation skills, illustrate transfer knowledge simulated robotic arms different numbers links, well simulated arms different actuation mechanisms, one robot torque-driven tendon-driven.",3
"complexity issues randomization strategies frank-wolfe algorithms machine learning. frank-wolfe algorithms convex minimization recently gained considerable attention optimization machine learning communities, properties make suitable choice variety applications. however, iteration requires optimize linear model, clever implementation crucial make algorithms viable large-scale datasets. purpose, approximation strategies based random sampling proposed several researchers. work, perform experimental study effectiveness techniques, analyze possible alternatives provide guidelines based results.",18
"nonlinear unsupervised clustering hyperspectral images applications anomaly detection active learning. problem unsupervised learning segmentation hyperspectral images significant challenge remote sensing. high dimensionality hyperspectral data, presence substantial noise, overlap classes contribute difficulty automatically clustering segmenting hyperspectral images. article, propose unsupervised learning technique combines geometric estimation class modes diffusion-inspired labeling incorporates spatial spectral information. mode estimation incorporates geometry hyperspectral data using diffusion distance promote learning unique mode class. class modes used label points joint spatial-spectral nonlinear diffusion process. proposed method, called spatial-spectral diffusion learning (dlss), shown perform competitively benchmark state-of-the-art hyperspectral clustering methods variety synthetic real datasets. proposed methods shown enjoy low computational complexity fast empirical runtime. two variations proposed method also discussed. first variation combines proposed method mode estimation partial least squares regression (plsr) efficiently segment chemical plumes hyperspectral images anomaly detection. second variation incorporates active learning allow user request labels small number pixels, dramatically improve overall clustering results. extensive experimental analysis demonstrate efficacy proposed methods, robustness choices parameters.",3
"context-dependent fine-grained entity type tagging. entity type tagging task assigning category labels mention entity document. standard systems focus small set types, recent work (ling weld, 2012) suggests using large fine-grained label set lead dramatic improvements downstream tasks. absence labeled training data, existing fine-grained tagging systems obtain examples automatically, using resolved entities types extracted knowledge base. however, since appropriate type often depends context (e.g. washington could tagged either city government), procedure result spurious labels, leading poorer generalization. propose task context-dependent fine type tagging, set acceptable labels mention restricted deducible local context (e.g. sentence document). introduce new resources task: 12,017 mentions annotated context-dependent fine types, provide baseline experimental results data.",3
"sentence similarity learning lexical decomposition composition. conventional sentence similarity methods focus similar parts two input sentences, simply ignore dissimilar parts, usually give us clues semantic meanings sentences. work, propose model take account similarities dissimilarities decomposing composing lexical semantics sentences. model represents word vector, calculates semantic matching vector word based words sentence. then, word vector decomposed similar component dissimilar component based semantic matching vector. this, two-channel cnn model employed capture features composing similar dissimilar components. finally, similarity score estimated composed feature vectors. experimental results show model gets state-of-the-art performance answer sentence selection task, achieves comparable result paraphrase identification task.",3
"leveraging patient similarity time series data healthcare predictive models. patient time series classification faces challenges high degrees dimensionality missingness. light patient similarity theory, study explores effective temporal feature engineering reduction, missing value imputation, change point detection methods afford similarity-based classification models desirable accuracy enhancement. select piecewise aggregation approximation method extract fine-grain temporal features propose minimalist method impute missing values temporal features. dimensionality reduction, adopt gradient descent search method feature weight assignment. propose new patient status directional change definitions based medical knowledge clinical guidelines value ranges different patient status levels, develop method detect change points indicating positive negative patient status changes. evaluate effectiveness proposed methods context early intensive care unit mortality prediction. evaluation results show k-nearest neighbor algorithm incorporates methods select propose significantly outperform relevant benchmarks early icu mortality prediction. study makes contributions time series classification early icu mortality prediction via identifying enhancing temporal feature engineering reduction methods similarity-based time series classification.",3
"scoring searching bayesian networks causal associative priors. significant theoretical advantage search-and-score methods learning bayesian networks accept informative prior beliefs possible network, thus complementing data. paper, method presented assigning priors based beliefs presence absence certain paths true network. beliefs correspond knowledge possible causal associative relations pairs variables. type knowledge naturally arises prior experimental observational data, among others. addition, novel search-operator proposed take advantage prior knowledge. experiments show that, using path beliefs improves learning skeleton, well edge directions network.",3
"interacting conceptual spaces : grammatical composition concepts. categorical compositional approach meaning successfully applied natural language processing, outperforming models mainstream empirical language processing tasks. show approach generalized conceptual space models cognition. order this, first introduce category convex relations new setting categorical compositional semantics, emphasizing convex structure important conceptual space applications. show construct conceptual spaces various types nouns, adjectives verbs. finally show means examples concepts systematically combined establish meanings composite phrases meanings constituent parts. provides mathematical underpinnings new compositional approach cognition.",3
"improving accuracy power transfer learning using meta-analytic database. typical cohorts brain imaging studies large enough systematic testing information contained images. build testable working hypotheses, investigators thus rely analysis previous work, sometimes formalized so-called meta-analysis. brain imaging, approach underlies specification regions interest (rois) usually selected basis coordinates previously detected effects. paper, propose use database images, rather coordinates, frame problem transfer learning: learning discriminant model reference task apply different related new task. facilitate statistical analysis small cohorts, use sparse discriminant model selects predictive voxels reference task thus provides principled procedure define rois. benefits approach twofold. first uses reference database prediction, i.e. provide potential biomarkers clinical setting. second increases statistical power new task. demonstrate set 18 pairs functional mri experimental conditions approach gives good prediction. addition, specific transfer situation involving different scanners different locations, show voxel selection based transfer learning leads higher detection power small cohorts.",18
"stochastic neural networks monotonic activation functions. propose laplace approximation creates stochastic unit smooth monotonic activation function, using gaussian noise. paper investigates application stochastic approximation training family restricted boltzmann machines (rbm) closely linked bregman divergences. family, call exponential family rbm (exp-rbm), subset exponential family harmoniums expresses family members choice smooth monotonic non-linearity neuron. using contrastive divergence along gaussian approximation, show exp-rbm learn useful representations using novel stochastic units.",18
"recurrent neural network grammars. introduce recurrent neural network grammars, probabilistic models sentences explicit phrase structure. explain efficient inference procedures allow application parsing language modeling. experiments show provide better parsing english single previously published supervised generative model better language modeling state-of-the-art sequential rnns english chinese.",3
"aggregation method sparse logistic regression. $l_1$ regularized logistic regression become workhorse data mining bioinformatics: widely used many classification problems, particularly ones many features. however, $l_1$ regularization typically selects many features so-called false positives unavoidable. paper, demonstrate analyze aggregation method sparse logistic regression high dimensions. approach linearly combines estimators suitable set logistic models different underlying sparsity patterns balance predictive ability model interpretability. numerical performance proposed aggregation method investigated using simulation studies. also analyze published genome-wide case-control dataset evaluate usefulness aggregation method multilocus association mapping.",18
"large deviation principles monge--ampère equation (following berman, hultgren). mostly exposition, aimed accessible geometers, analysts, probabilists, fundamental recent theorem r. berman recent developments j. hultgren, asserts second boundary value problem real monge--amp\`ere equation admits probabilistic interpretation, terms many particle limit permanental point processes satisfying large deviation principle rate function given explicitly using optimal transport. alternative proof step berman--hultgren theorem presented allowing deal ""tempratures"" simultaneously instead first reducing zero-temperature case.",10
"probabilistic data analysis probabilistic programming. probabilistic techniques central data analysis, different approaches difficult apply, combine, compare. paper introduces composable generative population models (cgpms), computational abstraction extends directed graphical models used describe compose broad class probabilistic data analysis techniques. examples include hierarchical bayesian models, multivariate kernel methods, discriminative machine learning, clustering algorithms, dimensionality reduction, arbitrary probabilistic programs. also demonstrate integration cgpms bayesdb, probabilistic programming platform express data analysis tasks using modeling language structured query language. practical value illustrated two ways. first, cgpms used analysis identifies satellite data records probably violate kepler's third law, composing causal probabilistic programs non-parametric bayes 50 lines probabilistic code. second, several representative data analysis tasks, report lines code accuracy measurements various cgpms, plus comparisons standard baseline solutions python matlab libraries.",3
"analysis kernel mean matching covariate shift. real supervised learning scenarios, uncommon training test sample follow different probability distributions, thus rendering necessity correct sampling bias. focusing particular covariate shift problem, derive high probability confidence bounds kernel mean matching (kmm) estimator, whose convergence rate turns depend regularity measure regression function also capacity measure kernel. comparing kmm natural plug-in estimator, establish superiority former hence provide concrete evidence/understanding effectiveness kmm covariate shift.",3
"on-the-fly learning perpetual learning machine. despite promise brain-inspired machine learning, deep neural networks (dnn) frustratingly failed bridge deceptively large gap learning memory. here, introduce perpetual learning machine; new type dnn capable brain-like dynamic 'on fly' learning exists self-supervised state perpetual stochastic gradient descent. thus, provide means unify learning memory within machine learning framework. also explore elegant duality abstraction synthesis: yin yang deep learning.",3
"non-associative learning representation nervous system nematode caenorhabditis elegans. caenorhabditis elegans (c. elegans) illustrated remarkable behavioral plasticities including complex non-associative associative learning representations. understanding principles mechanisms presumably leads constructive inspirations design efficient learning algorithms. present study, postulate novel approach modeling single neurons synapses study mechanisms underlying learning c. elegans nervous system. regard, construct precise mathematical model sensory neurons include multi-scale details genes, ion channels ion pumps, together dynamic model synapses comprised neurotransmitters receptors kinetics. recapitulate mechanosensory habituation mechanism, non-associative learning process, elements neural network tune parameters result repeated input stimuli. accordingly, quantitatively demonstrate roots plasticity neuronal synaptic-level representations. findings potentially give rise development new bio-inspired learning algorithms.",15
"nonparametric testing heterogeneous correlation. presence weak overall correlation, may useful investigate correlation significantly substantially pronounced subpopulation. two different testing procedures compared. based rankings values two variables data set large number n observations. first maintains level gaussian copulas; second adapts general alternatives sense number parameters used test grows n. analysis wine quality illustrates methods detect heterogeneity association chemical properties wine, attributable mix different cultivars.",18
"classifiers reject option early time-series classification. early classification time-series data dynamic environment challenging problem great importance signal processing. paper proposes classifier architecture reject option capable online decision making without need wait entire time series signal present. main idea classify odor/gas signal acceptable accuracy early possible. instead using posterior probability classifier, proposed method uses ""agreement"" ensemble decide whether accept reject candidate label. introduced algorithm applied bio-chemistry problem odor classification build novel electronic-nose called forefront-nose. experimental results wind tunnel test-bed facility confirms robustness forefront-nose compared standard classifiers earliness recognition perspectives.",3
"weakly supervised object boundaries. state-of-the-art learning based boundary detection methods require extensive training data. since labelling object boundaries one expensive types annotations, need relax requirement carefully annotate images make training affordable extend amount training data. paper propose technique generate weakly supervised annotations show bounding box annotations alone suffice reach high-quality object boundaries without using object-specific boundary annotations. proposed weak supervision techniques achieve top performance object boundary detection task, outperforming large margin current fully supervised state-of-the-art methods.",3
"workings genetic algorithms: genoclique fixing hypothesis. recently reported simple genetic algorithm (sga) capable performing remarkable form sublinear computation straightforward connection general problem interacting attributes data-mining. paper explain sga leverage computational proficiency perform efficient adaptation broad class fitness functions. based relative ease practical fitness function might belong broad class, submit new hypothesis workings genetic algorithms. explain hypothesis superior building block hypothesis, and, way empirical validation, present results experiment use simple mechanism called clamping dramatically improved performance sga uniform crossover large, randomly generated instances max 3-sat problem.",3
"multiclass feature learning hyperspectral image classification: sparse hierarchical solutions. paper, tackle question discovering effective set spatial filters solve hyperspectral classification problems. instead fixing priori filters parameters using expert knowledge, let model find within random draws (possibly infinite) space possible filters. define active set feature learner includes model features improve classifier. end, consider fast linear classifier, multiclass logistic classification, show good representation (the filters discovered), simple classifier reach least state art performances. apply proposed active set learner four hyperspectral image classification problems, including agricultural urban classification different resolutions, well multimodal data. also propose hierarchical setting, allows generate complex banks features better describe nonlinearities present data.",18
"adaptive inference general graphical models. many algorithms applications involve repeatedly solving variations inference problem; example may want introduce new evidence model perform updates conditional dependencies. goal adaptive inference take advantage preserved model perform inference rapidly scratch. paper, describe techniques adaptive inference general graphs support marginal computation updates conditional probabilities dependencies logarithmic time. give experimental results implementation algorithm, demonstrate potential performance benefit study protein structure.",3
"open source pattern recognition toolbox matlab. pattern recognition machine learning becoming integral parts algorithms wide range applications. different algorithms approaches machine learning include different tradeoffs performance computation, algorithm development often necessary explore variety different approaches given task. toolbox unified framework across multiple pattern recognition techniques enables algorithm developers ability rapidly evaluate different choices prior deployment. matlab widely used environment algorithm development prototyping, although several matlab toolboxes pattern recognition currently available either incomplete, expensive, restrictively licensed. work describe matlab toolbox pattern recognition machine learning known prt (pattern recognition toolbox), licensed permissive mit license. prt includes many popular techniques data preprocessing, supervised learning, clustering, regression feature selection, well methodology combining components using simple, uniform syntax. resulting algorithms evaluated using cross-validation variety scoring metrics ensure robust performance algorithm deployed. paper presents overview prt well example usage fisher's iris dataset.",18
"interaction information causal inference: case directed triangle. interaction information one multivariate generalizations mutual information, expresses amount information shared among set variables, beyond information, shared proper subset variables. unlike (conditional) mutual information, always non-negative, interaction information negative. utilize property find direction causal influences among variables triangle topology mild assumptions.",3
"convergence maximum variance unfolding. maximum variance unfolding one main methods (nonlinear) dimensionality reduction. study large sample limit, providing specific rates convergence standard assumptions. find consistent underlying submanifold isometric convex subset, provide simple examples fails consistent.",18
"comparison theorem pro-étale fundamental group. [bhatt-scholze, pro-\'etale topology schemes], b. bhatt p. scholze introduced notion pro-\'etale fundamental group connected locally topologically noetherian scheme $x$. topological group classifies geometric covers $x$. paper, prove $x$ scheme locally finite type $\mathbb{c}$, pro-\'etale fundamental group $x$ ""algebraizable completion"" topological fundamental group associated analytic space $x^{\rm an}$ $x$.",10
"tricolor dags machine translation. machine translation (mt) recently formulated terms constraint-based knowledge representation unification theories, becoming evident possible design practical mt system without adequate method handling mismatches semantic representations source target languages. paper, introduce idea ``information-based'' mt, considerably flexible interlingual mt conventional transfer-based mt.",1
"learning multi-grid generative convnets minimal contrastive divergence. paper proposes minimal contrastive divergence method learning energy-based generative convnet models images multiple grids (or scales) simultaneously. grid, learn energy-based probabilistic model energy function defined bottom-up convolutional neural network (convnet cnn). learning model requires generating synthesized examples model. within iteration learning algorithm, observed training image, generate synthesized images multiple grids initializing finite-step mcmc sampling minimal 1 x 1 version training image. synthesized image subsequent grid obtained finite-step mcmc initialized synthesized image generated previous coarser grid. obtaining synthesized examples, parameters models multiple grids updated separately simultaneously based differences synthesized observed examples. call learning method multi-grid minimal contrastive divergence. show method learn realistic energy-based generative convnet models, outperforms original contrastive divergence (cd) persistent cd.",18
"zagreb index, signless laplacian eigenvalues signless laplacian energy graph. let $g$ simple graph order $n$ size $m$. quantity $m_1(g)=\displaystyle\sum_{i=1}^{n}d^2_{v_i}$ called first zagreb index $g$, $d_{v_i}$ degree vertex $v_i$, $i=1,2,\dots,n$. signless laplacian matrix graph $g$ $q(g)=d(g)+a(g)$, $a(g)$ $d(g)$ denote, respectively, adjacency diagonal matrix vertex degrees $g$. let $q_1\geq q_2\geq \dots\geq q_n\geq 0$ signless laplacian eigenvalues $g$. largest signless laplacian eigenvalue $q_1$ called signless laplacian spectral radius $q$-index $g$ denoted $q(g)$. let $s^+_k(g)=\displaystyle\sum_{i=1}^{k}q_i$ $l_k(g)=\displaystyle\sum_{i=0}^{k-1}q_{n-i}$, $1\leq k\leq n$, respectively denote sum $k$ largest smallest signless laplacian eigenvalues $g$. signless laplacian energy $g$ defined $qe(g)=\displaystyle\sum_{i=1}^{n}|q_i-\overline{d}|$, $\overline{d}=\frac{2m}{n}$ average vertex degree $g$. article, obtain upper bounds first zagreb index $m_1(g)$ show bound best possible. using bounds, obtain several upper bounds graph invariant $s^+_k(g)$ characterize extremal cases. consequence, find upper bounds $q$-index lower bounds graph invariant $l_k(g)$ terms various graph parameters determine extremal cases. application, obtain upper bounds signless laplacian energy graph characterize extremal cases.",10
"shift: zero flop, zero parameter alternative spatial convolutions. neural networks rely convolutions aggregate spatial information. however, spatial convolutions expensive terms model size computation, grow quadratically respect kernel size. paper, present parameter-free, flop-free ""shift"" operation alternative spatial convolutions. fuse shifts point-wise convolutions construct end-to-end trainable shift-based modules, hyperparameter characterizing tradeoff accuracy efficiency. demonstrate operation's efficacy, replace resnet's 3x3 convolutions shift-based modules improved cifar10 cifar100 accuracy using 60% fewer parameters; additionally demonstrate operation's resilience parameter reduction imagenet, outperforming resnet family members. finally show shift operation's applicability across domains, achieving strong performance fewer parameters classification, face verification style transfer.",3
"objective classes micro-facial expression recognition. micro-expressions brief spontaneous facial expressions appear face person conceals emotion, making different normal facial expressions subtlety duration. currently, emotion classes within casme ii dataset based action units self-reports, creating conflicts machine learning training. show classifying expressions using action units, instead predicted emotion, removes potential bias human reporting. proposed classes tested using lbp-top, hoof hog 3d feature descriptors. experiments evaluated two benchmark facs coded datasets: casme ii samm. best result achieves 86.35\% accuracy classifying proposed 5 classes casme ii using hog 3d, outperforming result state-of-the-art 5-class emotional-based classification casme ii. results indicate classification based action units provides objective method improve micro-expression recognition.",3
"asymptotic expansions traces thermoelastic operators. obtain asymptotic expansions traces thermoelastic operators dirichlet neumann boundary conditions riemannian manifold, give effective method calculate coefficients asymptotic expansions. coefficients provide precise geometric information. particular, explicitly calculate first two coefficients concerning volumes manifold boundary. application, combining results isoperimetric inequality show $n$-dimensional geodesic ball uniquely determined isometry thermoelastic spectrum among bounded thermoelastic bodies boundary.",10
"stochastic gradient descent algorithms strongly convex functions o(1/t) convergence rates. weighting scheme proportional t, traditional stochastic gradient descent (sgd) algorithm achieves high probability convergence rate o({\kappa}/t) strongly convex functions, instead o({\kappa} ln(t)/t). also prove accelerated sgd algorithm also achieves rate o({\kappa}/t).",3
"automatic instrument recognition polyphonic music using convolutional neural networks. traditional methods tackle many music information retrieval tasks typically follow two-step architecture: feature engineering followed simple learning algorithm. ""shallow"" architectures, feature engineering learning typically disjoint unrelated. additionally, feature engineering difficult, typically depends extensive domain expertise. paper, present application convolutional neural networks task automatic musical instrument identification. model, feature extraction learning algorithms trained together end-to-end fashion. show convolutional neural network trained raw audio achieve performance surpassing traditional methods rely hand-crafted features.",3
"deep label distribution learning label ambiguity. convolutional neural networks (convnets) achieved excellent recognition performance various visual recognition tasks. large labeled training set one important factors success. however, difficult collect sufficient training images precise labels domains apparent age estimation, head pose estimation, multi-label classification semantic segmentation. fortunately, ambiguous information among labels, makes tasks different traditional classification. based observation, convert label image discrete label distribution, learn label distribution minimizing kullback-leibler divergence predicted ground-truth label distributions using deep convnets. proposed dldl (deep label distribution learning) method effectively utilizes label ambiguity feature learning classifier learning, help prevent network over-fitting even training set small. experimental results show proposed approach produces significantly better results state-of-the-art methods age estimation head pose estimation. time, also improves recognition performance multi-label classification semantic segmentation tasks.",3
"ron: reverse connection objectness prior networks object detection. present ron, efficient effective framework generic object detection. motivation smartly associate best region-based (e.g., faster r-cnn) region-free (e.g., ssd) methodologies. fully convolutional architecture, ron mainly focuses two fundamental problems: (a) multi-scale object localization (b) negative sample mining. address (a), design reverse connection, enables network detect objects multi-levels cnns. deal (b), propose objectness prior significantly reduce searching space objects. optimize reverse connection, objectness prior object detector jointly multi-task loss function, thus ron directly predict final detection results locations various feature maps. extensive experiments challenging pascal voc 2007, pascal voc 2012 ms coco benchmarks demonstrate competitive performance ron. specifically, vgg-16 low resolution 384x384 input size, network gets 81.3% map pascal voc 2007, 80.7% map pascal voc 2012 datasets. superiority increases datasets become larger difficult, demonstrated results ms coco dataset. 1.5g gpu memory test phase, speed network 15 fps, 3x faster faster r-cnn counterpart.",3
"accelerated variance reduced block coordinate descent. algorithms fast convergence, small number data access, low per-iteration complexity particularly favorable big data era, due demand obtaining \emph{highly accurate solutions} problems \emph{a large number samples} \emph{ultra-high} dimensional space. existing algorithms lack least one qualities, thus inefficient handling big data challenge. paper, propose method enjoying merits accelerated convergence rate $o(\frac{1}{k^2})$. empirical studies large scale datasets one million features conducted show effectiveness methods practice.",18
"analog sparse approximation applications compressed sensing. recent research shown performance signal processing tasks often significantly improved using signal models based sparse representations, signal approximated using small number elements fixed dictionary. unfortunately, inference model involves solving non-smooth optimization problems computationally expensive. significant efforts focused developing digital algorithms specifically problem, algorithms inappropriate many applications time power requirements necessary solve large optimization problems. based recent work computational neuroscience, explore potential advantages continuous time dynamical systems solving sparse approximation problems implemented analog vlsi. specifically, simulated task recovering synthetic mri data acquired via compressive sensing techniques, show systems potentially perform recovery time scales 10-20{\mu}s, supporting datarates 50-100 khz (orders magnitude faster digital algorithms). furthermore, show analytically wide range sparse approximation problems solved basic architecture, including approximate $\ell^p$ norms, modified $\ell^1$ norms, re-weighted $\ell^1$ $\ell^2$, block $\ell^1$ norm classic tikhonov regularization.",10
"unified framework order-of-magnitude confidence relations. aim work provide unified framework ordinal representations uncertainty lying crosswords possibility probability theories. confidence relations events commonly found monotonic reasoning, inconsistency management, qualitative decision theory. start either probability theory, making qualitative, possibility theory, making expressive. show two trends converge class genuine probability theories. provide characterization results useful tools preserve qualitative nature possibility rankings, enjoying power expressivity additive representations.",3
"inferring graphs cascades: sparse recovery framework. network inference problem, one seeks recover edges unknown graph observations cascades propagating graph. paper, approach problem sparse recovery perspective. introduce general model cascades, including voter model independent cascade model, provide first algorithm recovers graph's edges high probability $o(s\log m)$ measurements $s$ maximum degree graph $m$ number nodes. furthermore, show algorithm also recovers edge weights (the parameters diffusion process) robust context approximate sparsity. finally prove almost matching lower bound $\omega(s\log\frac{m}{s})$ validate approach empirically synthetic graphs.",3
"communication-efficient distributed sparse linear discriminant analysis. propose communication-efficient distributed estimation method sparse linear discriminant analysis (lda) high dimensional regime. method distributes data size $n$ $m$ machines, estimates local sparse lda estimator machine using data subset size $n/m$. distributed estimation, method aggregates debiased local estimators $m$ machines, sparsifies aggregated estimator. show aggregated estimator attains statistical rate centralized estimation method, long number machines $m$ chosen appropriately. moreover, prove method attain model selection consistency milder condition centralized method. experiments synthetic real datasets corroborate theory.",18
"computing egomotion local loop closures egocentric videos. finding camera pose important step many egocentric video applications. widely reported that, state art slam algorithms fail egocentric videos. paper, propose robust method camera pose estimation, designed specifically egocentric videos. egocentric video, camera views scene point multiple times wearer's head sweeps back forth. use specific motion profile perform short loop closures aligned wearer's footsteps. egocentric videos, depth estimation usually noisy. important departure, use 2d computations rotation averaging rely upon depth estimates. two modification results much stable algorithm evident experiments various egocentric video datasets different egocentric applications. proposed algorithm resolves long standing problem egocentric vision unlocks new usage scenarios future applications.",3
"$m_w$ helps select $z^\prime$ models $b \rightarrow \ell \ell$ anomalies. shown ref. \cite{allanach:2021kzj}, third family hypercharge ($y_3$) model changes standard model prediction $m_w$ whilst simultaneously explaining anomalies $b\to s\ell\ell$ transitions via heavy $z^\prime$ gauge boson spawned spontaneously broken gauged $u(1)_{y_3}$ symmetry. 2022 cdf ii measurement $m_w$, far standard model prediction, somewhat disfavours $y_3$ model. here, generalise gauge charge assignments anomaly-free combination $s y_3 + (b_3-l_3)$ show incorporating 2022 cdf ii measurement $m_w$ selects viable domain integers $s$ $t$. example, $s=1, t=-3$ yields $p-$value .08 two-parameter global fit 277 electroweak flavour changing $b$ data, much improving sm $p-$value $1\times 10^{-6}$.",8
"low complexity algorithm $o(\sqrt{t})$ regret finite constraint violations online convex optimization long term constraints. paper considers online convex optimization complicated constraint set, typically consists multiple functional constraints set constraint. conventional projection based online projection algorithm (zinkevich, 2003) difficult implement due potentially high computation complexity projection operation. paper, relax functional constraints allowing violated round still requiring satisfied long term. type relaxed online convex optimization (with long term constraints) first considered mahdavi et al. (2012). prior work proposes algorithm achieve $o(\sqrt{t})$ regret $o(t^{3/4})$ constraint violations general problems another algorithm achieve $o(t^{2/3})$ bound regret constraint violations constraint set described finite number linear constraints. recent extension jenatton et al. (2016) achieve $o(t^{\max\{\beta,1-\beta\}})$ regret $o(t^{1-\beta/2})$ constraint violations $\beta\in (0,1)$. current paper proposes new simple algorithm yields improved performance comparison prior works. new algorithm achieves $o(\sqrt{t})$ regret bound finite constraint violations.",10
"medical image segmentation localization using deformable templates. paper presents deformable templates tool segmentation localization biological structures medical images. structures represented prototype template, combined parametric warp mapping used deform original shape. localization procedure achieved using multi-stage, multi-resolution algorithm de-signed reduce computational complexity time. algorithm initially identifies regions image likely contain desired objects examines regions progressively increasing resolutions. final stage algorithm involves warping prototype template match localized objects. algorithm presented along results four example applications using mri, x-ray ultrasound images.",3
"chr-based implementation known arc-consistency. classical clp(fd) systems, domains variables completely known beginning constraint propagation process. however, systems interacting external environment, acquiring whole domains variables beginning constraint propagation may cause waste computation time, even obsolescence acquired data time use. cases, interactive constraint satisfaction problem (icsp) model proposed extension csp model, make possible start constraint propagation even domains fully known, performing acquisition domain elements necessary, without need restarting propagation every acquisition. paper, show solver two sorted clp language, defined previous work, express icsps, implemented constraint handling rules (chr) language, declarative language particularly suitable high level implementation constraint solvers.",3
"entropy-growth-based model emotionally charged online dialogues. analyze emotionally annotated massive data irc (internet relay chat) model dialogues participants assuming driving force discussion entropy growth emotional probability distribution. process claimed correlated emergence power-law distribution discussion lengths observed dialogues. perform numerical simulations based noticed phenomenon obtaining good agreement real data. finally, propose method artificially prolong duration discussion relies entropy emotional probability distribution.",14
"learning koopman invariant subspaces dynamic mode decomposition. spectral decomposition koopman operator attracting attention tool analysis nonlinear dynamical systems. dynamic mode decomposition popular numerical algorithm koopman spectral analysis; however, often need prepare nonlinear observables manually according underlying dynamics, always possible since may priori knowledge them. paper, propose fully data-driven method koopman spectral analysis based principle learning koopman invariant subspaces observed data. end, propose minimization residual sum squares linear least-squares regression estimate set functions transforms data form linear regression fits well. introduce implementation neural networks evaluate performance empirically using nonlinear dynamical systems applications.",3
"dual ascent framework lagrangean decomposition combinatorial problems. propose general dual ascent framework lagrangean decomposition combinatorial problems. although methods type shown efficiency number problems, far general algorithm applicable multiple problem types. work, propose general algorithm. depends several parameters, used optimize performance particular setting. demonstrate efficacy method graph matching multicut problems, outperforms state-of-the-art solvers including based subgradient optimization off-the-shelf linear programming solvers.",3
"unsupervised aspect term extraction b-lstm & crf using automatically labelled datasets. aspect term extraction (ate) identifies opinionated aspect terms texts one tasks semeval aspect based sentiment analysis (absa) contest. small amount available datasets supervised ate costly human annotation aspect term labelling give rise need unsupervised ate. paper, introduce architecture achieves top-ranking performance supervised ate. moreover, used efficiently feature extractor classifier unsupervised ate. second contribution method automatically construct datasets ate. train classifier automatically labelled datasets evaluate human annotated semeval absa test sets. compared strong rule-based baseline, obtain dramatically higher f-score attain precision values 80%. unsupervised method beats supervised absa baseline semeval, preserving high precision scores.",3
"generalizable features unsupervised learning. humans learn predictive model world use model reason future events consequences actions. contrast machine predictors, exhibit impressive ability generalize unseen scenarios reason intelligently settings. one important aspect ability physical intuition(lake et al., 2016). work, explore potential unsupervised learning find features promote better generalization settings outside supervised training distribution. task predicting stability towers square blocks. demonstrate unsupervised model, trained predict future frames video sequence stable unstable block configurations, yield features support extrapolating stability prediction blocks configurations outside training set distribution",18
"interference management over-the-air federated learning multi-cell wireless networks. federated learning (fl) resource-constrained wireless networks recently attracted much attention. however, existing studies consider one fl task single-cell wireless networks ignore impact downlink/uplink inter-cell interference learning performance. paper, investigate fl multi-cell wireless network, cell performs different fl task over-the-air computation (aircomp) adopted enable fast uplink gradient aggregation. conduct convergence analysis aircomp-assisted fl systems, taking account inter-cell interference downlink uplink model/gradient transmissions, reveals distorted model/gradient exchanges induce gap hinder convergence fl. characterize pareto boundary error-induced gap region quantify learning performance trade-off among different fl tasks, based formulate optimization problem minimize sum error-induced gaps cells. tackle coupling downlink uplink transmissions well coupling among multiple cells, propose cooperative multi-cell fl optimization framework achieve efficient interference management downlink uplink transmission design. results demonstrate proposed algorithm achieves much better average learning performance multiple cells non-cooperative baseline schemes.",5
"abstractive meeting summarization usingdependency graph fusion. automatic summarization techniques meeting conversations developed far primarily extractive, resulting poor summaries. improve this, propose approach generate abstractive summaries fusing important content several utterances. meeting generally comprised several discussion topic segments. topic segment within meeting conversation, aim generate one sentence summary important utterances using integer linear programming-based sentence fusion approach. experimental results show method generate informative summaries baselines.",3
"accelerating kernel classifiers borders mapping. support vector machines (svm) kernel techniques represent family powerful statistical classification methods high accuracy broad applicability. use significant portion training data, however, slow, especially large problems. piecewise linear classifiers similarly versatile, yet additional advantages simplicity, ease interpretation and, number component linear classifiers large, speed. show simple, piecewise linear classifier trained kernel-based classifier order improve classification speed. method works finding root difference conditional probabilities pairs opposite classes build representation decision boundary. tested 17 different datasets, succeeded improving classification speed svm 9 factors high 88 times more. method best suited problems continuum features data smooth probability functions. component linear classifiers built individually existing classifier, rather simultaneous optimization procedure, classifier also fast train.",18
"scene labeling knowledge-based rules employing constrained integer linear programing. scene labeling task segment image meaningful regions categorize classes objects comprised image. commonly used methods typically find local features segment label using classifiers. afterward, labeling smoothed order make sure neighboring regions receive similar labels. however, ignore expressive non-local dependencies among regions due expensive training inference. paper, propose use high-level knowledge regarding rules inference incorporate dependencies among regions image improve scores classification. towards aim, extract rules data transform constraints integer programming optimize structured problem assigning labels super-pixels (consequently pixels) image. addition, propose use soft-constraints scenarios, allowing violating constraint imposing penalty, make model flexible. assessed approach three datasets obtained promising results.",3
"mostly exploration-free algorithms contextual bandits. contextual bandit literature traditionally focused algorithms address exploration-exploitation tradeoff. particular, greedy algorithms exploit current estimates without exploration may sub-optimal general. however, exploration-free greedy algorithms desirable practical settings exploration may costly unethical (e.g., clinical trials). surprisingly, find simple greedy algorithm rate-optimal sufficient randomness observed contexts. prove always case two-armed bandit general class context distributions satisfy condition term covariate diversity. furthermore, even absent condition, show greedy algorithm rate-optimal nonzero probability. thus, standard bandit algorithms may unnecessarily explore. motivated results, introduce greedy-first, new algorithm uses observed contexts rewards determine whether follow greedy algorithm explore. prove algorithm rate-optimal without additional assumptions context distribution number arms. extensive simulations demonstrate greedy-first successfully reduces experimentation outperforms existing (exploration-based) contextual bandit algorithms thompson sampling ucb.",18
"hierarchical dirichlet process model multiple levels clustering human eeg seizure modeling. driven multi-level structure human intracranial electroencephalogram (ieeg) recordings epileptic seizures, introduce new variant hierarchical dirichlet process---the multi-level clustering hierarchical dirichlet process (mlc-hdp)---that simultaneously clusters datasets multiple levels. seizure dataset contains brain activity recorded typically hundred individual channels seizure patient. mlc-hdp model clusters channels-types, seizure-types, patient-types simultaneously. describe model implementation detail. also present results simulation study comparing mlc-hdp similar model, nested dirichlet process finally demonstrate mlc-hdp's use modeling seizures across multiple patients. find mlc-hdp's clustering comparable independent human physician clusterings. knowledge, mlc-hdp model first epilepsy literature capable clustering seizures within patients.",18
"interactive grounded language acquisition generalization 2d world. build virtual agent learning language 2d maze-like world. agent sees images surrounding environment, listens virtual teacher, takes actions receive rewards. interactively learns teacher's language scratch based two language use cases: sentence-directed navigation question answering. learns simultaneously visual representations world, language, action control. disentangling language grounding computational routines sharing concept detection function language grounding prediction, agent reliably interpolates extrapolates interpret sentences contain new word combinations new words missing training sentences. new words transferred answers language prediction. language ability trained evaluated population 1.6 million distinct sentences consisting 119 object words, 8 color words, 9 spatial-relation words, 50 grammatical words. proposed model significantly outperforms five comparison methods interpreting zero-shot sentences. addition, demonstrate human-interpretable intermediate outputs model appendix.",3
"practical fundamental limits neutral atom entanglement using rydberg dressing. revisit implementation two-qubit entangling gate, m{\o}lmer-s{\o}rensen gate, using adiabatic rydberg dressing paradigm. study implementation rapid adiabatic passage using two-photon transition, require use ultra-violet laser, implemented using amplitude modulation one field laser frequencies fixed. find entangling gate fidelities, comparable one-photon excitation, achievable two-photon excitation. moreover, address adiabatic dressing protocol used implement entangling gates outside regime perfect rydberg blockade. show using adiabatic dressing achieve scaling gate fidelity set fundamental limits entanglement generated rydberg interactions simultaneously retaining limited population doubly excited rydberg state. allows fast high fidelity gates atoms separated beyond blockade radius.",17
"graph based manifold regularized deep neural networks automatic speech recognition. deep neural networks (dnns) successfully applied wide variety acoustic modeling tasks recent years. include applications dnns either discriminative feature extraction hybrid acoustic modeling scenario. despite rapid progress area, number challenges remain training dnns. paper presents effective way training dnns using manifold learning based regularization framework. framework, parameters network optimized preserve underlying manifold based relationships speech feature vectors minimizing measure loss network outputs targets. achieved incorporating manifold based locality constraints objective criterion dnns. empirical evidence provided demonstrate training network manifold constraints preserves structural compactness hidden layers network. manifold regularization applied train bottleneck dnns feature extraction hidden markov model (hmm) based speech recognition. experiments work conducted aurora-2 spoken digits aurora-4 read news large vocabulary continuous speech recognition tasks. performance measured terms word error rate (wer) tasks. shown manifold regularized dnns result 37% reduction wer relative standard dnns.",18
computer-aided diagnosis lung carcinoma using deep learning - pilot study. aim: early detection correct diagnosis lung cancer important steps improving patient outcome. study aims assess deep learning models perform best lung cancer diagnosis. methods: non-small cell lung carcinoma small cell lung carcinoma biopsy specimens consecutively obtained stained. specimen slides diagnosed two experienced pathologists (over 20 years). several deep learning models trained discriminate cancer non-cancer biopsies. result: deep learning models give reasonable auc 0.8810 0.9119. conclusion: deep learning analysis could help speed detection process whole-slide image (wsi) keep comparable detection rate human observer.,3
"solution degree constrained spanning tree using hybrid ga. real life, always urge reach goal minimum effort i.e., minimum constrained path. path may shortest route practical life, either physical electronic medium. scenario represents ambiance graph find spanning tree custom design criteria. here, chosen minimum degree spanning tree, generated real time minimum turnaround time. problem np-complete nature [1, 2]. solution approach, general, approximate. used heuristic approach, namely hybrid genetic algorithm (ga), motivated criteria encoded data structures graph. compare experimental result existing approximate algorithm result encouraging interested use future applications.",3
"significance analysis high-dimensional, low-sample size partially labeled data. classification clustering important topics statistical learning. natural question herein whether predefined classes really different one another, whether clusters really there. specifically, may interested knowing whether two classes defined class labels (when provided), two clusters tagged clustering algorithm (where class labels provided), underlying distribution. although challenging questions high-dimensional, low-sample size data, recent development both. however, costly manually place labels observations, often small portion class labels available. article, propose significance analysis approach type data, namely partially labeled data. method makes use whole data tries test class difference labels observed. compared testing method ignores label information, method provides greater power, meanwhile, maintaining size, illustrated comprehensive simulation study. theoretical properties proposed method studied emphasis high-dimensional, low-sample size setting. simulated examples help understand information extracted labeled data effective. real data example illustrates usefulness proposed method.",18
"improving statistical multimedia information retrieval model using ontology. typical ir system delivers stores information affected problem matching user query available content web. use ontology represents extracted terms form network graph consisting nodes, edges, index terms etc. mentioned ir approaches provide relevance thus satisfying users query. paper also emphasis analyzing multimedia documents performs calculation extracted terms using different statistical formulas. proposed model developed reduces semantic gap satisfies user needs efficiently.",3
"using inspiration synaptic plasticity rules optimize traffic flow distributed engineered networks. controlling flow routing data fundamental problem many distributed networks, including transportation systems, integrated circuits, internet. brain, synaptic plasticity rules discovered regulate network activity response environmental inputs, enable circuits stable yet flexible. here, develop new neuro-inspired model network flow control depends modifying edge weights activity-dependent manner. show two fundamental plasticity rules (long-term potentiation long-term depression) cast distributed gradient descent algorithm regulating traffic flow engineered networks. characterize, via simulation analytically, different forms edge-weight update rules affect network routing efficiency robustness. find close correspondence certain classes synaptic weight update rules derived experimentally brain rules commonly used engineering, suggesting common principles both.",3
"analysis t-sne algorithm data visualization. first line attack exploratory data analysis data visualization, i.e., generating 2-dimensional representation data makes clusters similar points visually identifiable. standard johnson-lindenstrauss dimensionality reduction produce data visualizations. t-sne heuristic van der maaten hinton, based non-convex optimization, become de facto standard visualization wide range applications. work gives formal framework problem data visualization - finding 2-dimensional embedding clusterable data correctly separates individual clusters make visually identifiable. give rigorous analysis performance t-sne natural, deterministic condition ""ground-truth"" clusters (similar conditions assumed earlier analyses clustering) underlying data. first provable guarantees t-sne constructing good data visualizations. show deterministic condition satisfied considerably general probabilistic generative models clusterable data mixtures well-separated log-concave distributions. finally, give theoretical evidence t-sne provably succeeds partially recovering cluster structure even deterministic condition met.",3
"review action recognition detection methods. computer vision, action recognition refers act classifying action present given video action detection involves locating actions interest space and/or time. videos, contain photometric information (e.g. rgb, intensity values) lattice structure, contain information assist identifying action imaged. process action recognition detection often begins extracting useful features encoding ensure features specific serve task action recognition detection. encoded features processed classifier identify action class spatial and/or temporal locations. report, thorough review various action recognition detection algorithms computer vision provided analyzing two-step process typical action recognition detection algorithm: (i) extraction encoding features, (ii) classifying features action classes. efforts ensure computer vision-based algorithms reach capabilities humans identifying actions irrespective various nuisance variables may present within field view, state-of-the-art methods reviewed remaining problems addressed final chapter.",3
"pseudo mask augmented object detection. work, present novel effective framework facilitate object detection instance-level segmentation information supervised bounding box annotation. starting joint object detection instance segmentation network, propose recursively estimate pseudo ground-truth object masks instance-level object segmentation network training, enhance detection network top-down segmentation feedbacks. pseudo ground truth mask network parameters optimized alternatively mutually benefit other. obtain promising pseudo masks iteration, embed graphical inference incorporates low-level image appearance consistency bounding box annotations refine segmentation masks predicted segmentation network. approach progressively improves object detection performance incorporating detailed pixel-wise information learned weakly-supervised segmentation network. extensive evaluation detection task pascal voc 2007 2012 [12] verifies proposed approach effective.",3
"learning abstract classes using deep learning. humans generally good learning abstract concepts objects scenes (e.g.\ spatial orientation, relative sizes, etc.). last years convolutional neural networks achieved almost human performance recognizing concrete classes (i.e.\ specific object categories). paper tests performance current cnn (googlenet) task differentiating abstract classes trivially differentiable humans. trained tested cnn two abstract classes horizontal vertical orientation determined well network able transfer learned classes other, previously unseen objects.",3
"mixed membership models time series. article discuss consequences mixed membership perspective time series analysis. abstract form, mixed membership model aims associate individual entity set attributes based collection observed data. although much literature mixed membership models considers setting exchangeable collections data associated member set entities, equally natural consider problems entire time series viewed entity goal characterize time series terms set underlying dynamic attributes ""dynamic regimes"". indeed, perspective already present classical hidden markov model, dynamic regimes referred ""states"", collection states realized sample path underlying process viewed mixed membership characterization observed time series. goal review richer modeling possibilities time series provided recent developments mixed membership framework.",18
"optimal neuron selection: nk echo state networks reinforcement learning. paper introduces nk echo state network. problem learning nk echo state network reduced problem optimizing special form spin glass problem known nk landscape. weight adjustment used; learning accomplished spinning (turning on) spinning (turning off) neurons order find combination neurons work together achieve desired computation. special types nk landscapes, exact global solution obtained polynomial time using dynamic programming. nk echo state network applied reinforcement learning problem requiring recurrent network: balancing two poles cart given velocity information. empirical results shows nk echo state network learns rapidly yields good generalization.",3
"classifying clustering schemes. many clustering schemes defined optimizing objective function defined partitions underlying set finite metric space. paper, construct framework studying happens instead impose various structural conditions clustering schemes, general heading functoriality. functoriality refers idea one able compare results clustering algorithms one varies data set, example adding points applying functions it. show within framework, one prove theorems analogous one j. kleinberg, example one obtains existence uniqueness theorem instead non-existence result. obtain full classification clustering schemes satisfying condition refer excisiveness. classification changed varying notion maps finite metric spaces. conditions occur naturally one considers clustering statistical version geometric notion connected components. varying degree functoriality one requires schemes possible construct richer families clustering schemes exhibit sensitivity density.",18
"x-view: graph-based semantic multi-view localization. global registration multi-view robot data challenging task. appearance-based global localization approaches often fail drastic view-point changes, representations limited view-point invariance. work based idea human-made environments contain rich semantics used disambiguate global localization. here, present x-view, multi-view semantic global localization system. x-view leverages semantic graph descriptor matching global localization, enabling localization drastically different view-points. approach general terms semantic input data, present evaluate implementation visual data. demonstrate system experiments publicly available synthia dataset, realistic urban dataset recorded simulator, real-world streetview data. findings show x-view able globally localize aerial-to-ground, ground-to-ground robot data drastically different view-points. approach achieves accuracy 85 % global localizations multi-view case, benchmarked baseline appearance-based methods reach 75 %.",3
"driving mode shock-driven turbulence. turbulence interstellar medium (ism) crucial process star formation. shocks produced supernova explosions, jets, radiation massive stars, galactic spiral-arm dynamics amongst common drivers turbulence ism. however, fully understood shocks drive turbulence, particular whether shock driving solenoidal(rotational, divergence-free) compressive (potential, curl-free) mode driving turbulence. mode turbulence driving profound consequences star formation, compressive driving producing three times larger density dispersion, order magnitude higher star formation rate solenoidal driving. here, use hydrodynamical simulations shock inducing turbulent motions structured, multi-phase medium. done context laser-induced shock, propagating foam material, preparation experiment performed national ignition facility (nif). specifically, analyse density velocity distributions shocked turbulent medium, measure turbulence driving parameter $b=(\sigma^{2 \gamma}_{\rho /\langle \rho \rangle}-1)^{1/2} (1-\sigma_{\rho \langle \rho \rangle}^{-2})^{-1/2}\mathcal{m}^{-1}\gamma^{-1/2}$ density dispersion $\sigma_{\rho / \langle \rho \rangle}$, turbulent mach number $\mathcal{m}$, polytropic exponent $\gamma$. purely solenoidal purely compressive driving correspond $b \sim 1/3$ $b \sim 1$, respectively. using simulations shock driven multi-phase medium structures different sizes $\gamma < 1$, find $b \sim 1$ cases, showing shock-driven turbulence consistent strongly compressive driving.",0
"general approach belief change answer set programming. address problem belief change (nonmonotonic) logic programming answer set semantics. unlike previous approaches belief change logic programming, formal techniques analogous distance-based belief revision propositional logic. developing results, build upon model theory logic programs furnished se models. since se models provide formal, monotonic characterisation logic programs, adapt techniques area belief revision belief change logic programs. introduce methods revising merging logic programs, respectively. former, study subset-based revision well cardinality-based revision, show satisfy majority agm postulates revision. merging, consider operators following arbitration merging ic merging, respectively. also present encodings computing revision well merging logic programs within logic programming framework, giving rise direct implementation approach terms off-the-shelf answer set solvers. encodings reflect turn fact change operators increase complexity base formalism.",3
"dematel-based completion method incomplete pairwise comparison matrix ahp. pairwise comparison matrix crucial component ahp, presents prefer- ence relations among alternatives. however, many cases, pairwise comparison matrix difficult complete, obstructs subsequent operations clas- sical ahp. paper, based dematel ability derive total relation matrix direct relation matrix, new completion method incomplete pairwise comparison matrix proposed. proposed method provides new per- spective estimate missing values explicit physical meaning. besides, proposed method low computational cost. promising method wide application multi-criteria decision-making.",10
"persistent clustering theorem j. kleinberg. construct framework studying clustering algorithms, includes two key ideas: persistence functoriality. first encodes idea output clustering scheme carry multiresolution structure, second idea one able compare results clustering algorithms one varies data set, example adding points applying functions it. show within framework, one prove theorem analogous one j. kleinberg, one obtains existence uniqueness theorem instead non-existence result. explore properties unique scheme, stability convergence established.",18
"sparse-then-dense alignment based 3d map reconstruction method endoscopic capsule robots. since development capsule endoscopcy technology, substantial progress made converting passive capsule endoscopes robotic active capsule endoscopes controlled doctor. however, robotic capsule endoscopy still challenges. particular, use devices generate precise globally consistent three-dimensional (3d) map entire inner organ remains unsolved problem. global 3d maps inner organs would help doctors detect location size diseased areas accurately, precisely, intuitively, thus permitting accurate intuitive diagnoses. proposed 3d reconstruction system built modular fashion including preprocessing, frame stitching, shading-based 3d reconstruction modules. propose efficient scheme automatically select key frames huge quantity raw endoscopic images. together bundle fusion approach aligns selected key frames jointly globally consistent way, significant improvement mosaic 3d map accuracy reached. best knowledge, framework first complete pipeline endoscopic capsule robot based 3d map reconstruction containing necessary steps reliable accurate endoscopic 3d map. qualitative evaluations, real pig stomach employed. moreover, first time literature, detailed comprehensive quantitative analysis proposed pipeline modules performed using non-rigid esophagus gastro duodenoscopy simulator, four different endoscopic cameras, magnetically activated soft capsule robot (masce), sub-millimeter precise optical motion tracker fine-scale 3d optical scanner.",3
"foundations modern language resource archives. number serious reasons convince increasing amount researchers store relevant material centers call ""language resource archives"". combine duty taking care long-term preservation well task give access material different user groups. access meant sense active interaction data made possible support integration new data, new versions commentaries sort. modern language resource archives adhere number basic principles fulfill requirements involved federations create joint language resource domains making even simple researchers access data. paper makes attempt formulate essential pillars language resource archives adhere to.",3
"deceptiveness neutrality - nd family fitness landscapes. considerable number mutations effects fitness values, fitness landscape said neutral. order study interplay neutrality, exists many real-world applications, performances metaheuristics, useful design landscapes make possible tune precisely neutral degree distribution. even though many neutral landscape models already designed, none general enough create landscapes specific neutral degree distributions. propose three steps design landscapes: first using algorithm construct landscape whose distribution roughly fits target one, use simulated annealing heuristic bring closer two distributions finally affect fitness values neutral network. using new family fitness landscapes able highlight interplay deceptiveness neutrality.",3
"automatic topography high-dimensional data sets non-parametric density peak clustering. data analysis high-dimensional spaces aims obtaining synthetic description data set, revealing main structure salient features. introduce approach charting data spaces, providing topography probability distribution data harvested. topography includes information number height probability peaks, depth ""valleys"" separating them, relative location peaks hierarchical organization. topography reconstructed using unsupervised variant density peak clustering exploiting non-parametric density estimator, automatically measures density manifold containing data. importantly, density estimator provides estimate error. key feature, allows distinguishing genuine probability peaks density fluctuations due finite sampling.",18
"quantum pattern retrieval qubit networks hebb interactions. qubit networks long-range interactions inspired hebb rule used quantum associative memories. starting uniform superposition, unitary evolution generated interactions drives network quantum phase transition critical computation time, ferromagnetic order guarantees measurement retrieves stored memory. maximum memory capacity p qubit networks reached memory density p/n=1.",17
"fast ilp-based heuristic robust design body wireless sensor networks. consider problem optimally designing body wireless sensor network, taking account uncertainty data generation biosensors. since related min-max robustness integer linear programming (ilp) problem difficult solve even state-of-the-art commercial optimization solvers, propose original heuristic solution. heuristic combines deterministic probabilistic variable fixing strategies, guided information coming strengthened linear relaxations ilp robust model, includes large neighborhood search reparation improvement generated solutions, formulated ilp problem solved exactly. computational tests realistic instances show heuristic finds solutions much higher quality state-of-the-art solver effective benchmark heuristic.",10
"restricted strong convexity implies weak submodularity. connect high-dimensional subset selection submodular maximization. results extend work das kempe (2011) setting linear regression arbitrary objective functions. greedy feature selection, connection allows us obtain strong multiplicative performance bounds several methods without statistical modeling assumptions. also derive recovery guarantees form standard assumptions. work shows greedy algorithms perform within constant factor best possible subset-selection solution broad class general objective functions. methods allow direct control number obtained features opposed regularization parameters implicitly control sparsity. proof technique uses concept weak submodularity initially defined das kempe. draw connection convex analysis submodular set function theory may independent interest statistical learning applications combinatorial structure.",18
"noise tolerance risk minimization. paper explore noise tolerant learning classifiers. formulate problem follows. assume ${\bf unobservable}$ training set noise-free. actual training set given learning algorithm obtained ideal data set corrupting class label example. probability class label example corrupted function feature vector example. would account kinds noisy data one encounters practice. say learning method noise tolerant classifiers learnt ideal noise-free data noisy data, classification accuracy noise-free data. paper analyze noise tolerance properties risk minimization (under different loss functions), generic method learning classifiers. show risk minimization 0-1 loss function impressive noise tolerance properties squared error loss tolerant uniform noise; risk minimization loss functions noise tolerant. conclude paper discussion implications theoretical results.",3
"probabilistic interpretations mycin's certainty factors. paper examines quantities used mycin reason uncertainty, called certainty factors. shown original definition certainty factors inconsistent functions used mycin combine quantities. inconsistency used argue redefinition certainty factors terms intuitively appealing desiderata associated combining functions. shown redefinition accommodates unlimited number probabilistic interpretations. interpretations shown monotonic transformations likelihood ratio p(eih)/p(el h). construction interpretations provides insight assumptions implicit certainty factor model. particular, shown uncertainty propagated inference network accordance desiderata, evidence must conditionally independent given hypothesis negation inference network must tree structure. emphasized assumptions implicit model rarely true practical applications. methods relaxing assumptions suggested.",3
"probabilities sentences expressive logic. automated reasoning uncertain knowledge many applications. one difficulty developing systems lack completely satisfactory integration logic probability. address problem directly. expressive languages like higher-order logic ideally suited representing reasoning structured knowledge. uncertain knowledge modeled using graded probabilities rather binary truth-values. main technical problem studied paper following: given set sentences, probability true, probability ascribed (query) sentences? natural wish-list, among others, probability distribution (i) consistent knowledge base, (ii) allows consistent inference procedure particular (iii) reduces deductive logic limit probabilities 0 1, (iv) allows (bayesian) inductive reasoning (v) learning limit particular (vi) allows confirmation universally quantified hypotheses/sentences. translate wish-list technical requirements prior probability show probabilities satisfying criteria exist. also give explicit constructions several general characterizations probabilities satisfy criteria various (counter) examples. also derive necessary sufficient conditions extending beliefs finitely many sentences suitable probabilities sentences, particular least dogmatic least biased ones. conclude brief outlook developed theory might used approximated autonomous reasoning agents. theory step towards globally consistent empirically satisfactory unification probability logic.",3
"hierarchical dirichlet process tracking complex topical structure evolution application autism research literature. paper describe novel framework discovery topical content data corpus, tracking complex structural changes across temporal dimension. contrast previous work model impose prior rate documents added corpus adopt markovian assumption overly restricts type changes model capture. key technical contribution framework based (i) discretization time epochs, (ii) epoch-wise topic discovery using hierarchical dirichlet process-based model, (iii) temporal similarity graph allows modelling complex topic changes: emergence disappearance, evolution, splitting merging. power proposed framework demonstrated medical literature corpus concerned autism spectrum disorder (asd) - increasingly important research subject significant social healthcare importance. addition collected asd literature corpus make freely available, contributions also include two free online tools built aids asd researchers. used semantically meaningful navigation searching, well knowledge discovery large rapidly growing corpus literature.",3
"holistic, instance-level human parsing. object parsing -- task decomposing object semantic parts -- traditionally formulated category-level segmentation problem. consequently, multiple objects image, current methods cannot count number objects scene, determine part belongs object. address problem segmenting parts objects instance-level, pixel image assigned part label, well identity object belongs to. moreover, show approach benefits us obtaining segmentations coarser granularities well. proposed network trained end-to-end given detections, begins category-level segmentation module. thereafter, differentiable conditional random field, defined variable number instances every input image, reasons identity part associating human detection. contrast approaches, method handle varying number people image holistic network produces state-of-the-art results instance-level part human segmentation, together competitive results category-level part segmentation, achieved single forward-pass neural network.",3
"robust data geometric structure aligned close yet discriminative domain adaptation. domain adaptation (da) transfer learning aims leverage labeled data related source domain achieve informed knowledge transfer help classification unlabeled data target domain. paper, propose novel da method, namely robust data geometric structure aligned, close yet discriminative domain adaptation (rsa-cdda), brings closer, latent joint subspace, source target data distributions, aligns inherent hidden source target data geometric structures performing discriminative da repulsing interclass source target data. proposed method performs domain adaptation source target solving unified model, incorporates data distribution constraints, particular via nonparametric distance, i.e., maximum mean discrepancy (mmd), well constraints inherent hidden data geometric structure segmentation alignment source target, low rank sparse representation. rsa-cdda achieves search joint subspace solving proposed unified model iterative optimization, alternating rayleigh quotient algorithm inexact augmented lagrange multiplier algorithm. extensive experiments carried standard da benchmarks, i.e., 16 cross-domain image classification tasks, verify effectiveness proposed method, consistently outperforms state-of-the-art methods.",3
"anonymous conference key agreement linear quantum networks. sharing multi-partite quantum entanglement parties allows diverse secure communication tasks performed. among them, conference key agreement (cka), extension key distribution multiple parties, received much attention recently. interestingly, cka also performed way protects identities participating parties, therefore providing anonymity. work, propose anonymous cka protocol three parties implemented highly practical network setting. specifically, line quantum repeater nodes used build linear cluster state among nodes, used anonymously establish secret key three them. nodes need share maximally entangled pairs neighbours, therefore avoiding necessity central server sharing entangled states. repeater setup makes protocol excellent candidate implementation future quantum networks. explicitly prove protocol protects identities participants one another perform analysis key rate finite regime, contributing quest identifying feasible quantum communication tasks network architectures beyond point-to-point.",17
"survey expressivity deep neural networks. survey results neural network expressivity described ""on expressive power deep neural networks"". paper motivates develops three natural measures expressiveness, display exponential dependence depth network. fact, measures related fourth quantity, trajectory length. quantity grows exponentially depth network, responsible depth sensitivity observed. results translate consequences networks training. suggest parameters earlier network greater influence expressive power -- particular, given layer, influence expressivity determined remaining depth network layer. verified experiments mnist cifar-10. also explore effect training input-output map, find trades stability expressivity.",18
"complex network classification convolutional neural network. classifying large scale networks several categories distinguishing according fine structures great importance several applications real life. however, studies complex networks focus properties single network seldom classification, clustering, comparison different networks, network treated whole. due non-euclidean properties data, conventional methods hardly applied networks directly. paper, propose novel framework complex network classifier (cnc) integrating network embedding convolutional neural network tackle problem network classification. training classifiers synthetic complex network data real international trade network data, show cnc classify networks high accuracy robustness, also extract features networks automatically.",3
"complexity modified instances. paper study complexity solving problem solution similar instance known. problem relevant whenever instances may change time time, known solutions may remain valid change. consider two scenarios: first one, known solution problem change; second case, assume additional information, found search solution, also known. first setting, techniques theory np-completeness suffice show complexity results. second case, negative results proved using techniques compilability, often related size considered changes.",3
"tv-gan: generative adversarial network based thermal visible face recognition. work tackles face recognition task images captured using thermal camera sensors operate non-light environment. greatly increase scope benefits current security surveillance systems, performing task using thermal images challenging problem compared face recognition task visible light domain (vld). partly due much smaller amount number thermal imagery data collected compared vld data. unfortunately, direct application existing strong face recognition models trained using vld data thermal imagery data produce satisfactory performance. due existence domain gap thermal vld images. end, propose thermal-to-visible generative adversarial network (tv-gan) able transform thermal face images corresponding vld images whilst maintaining identity information sufficient enough existing vld face recognition models perform recognition. examples presented figure 1. unlike previous methods, proposed tv-gan uses explicit closed-set face recognition loss regularize discriminator network training. information conveyed generator network forms gradient loss. experiment, show using additional explicit regularization discriminator network, tv-gan able preserve identity information translating thermal image person seen tv-gan.",3
"general method robust bayesian modeling. robust bayesian models appealing alternatives standard models, providing protection data contains outliers departures model assumptions. historically, robust models mostly developed case-by-case basis; examples include robust linear regression, robust mixture models, bursty topic models. paper develop general approach robust bayesian modeling. show turn existing bayesian model robust model, develop generic strategy computing it. use method study robust variants several models, including linear regression, poisson regression, logistic regression, probabilistic topic models. discuss connections methods existing approaches, especially empirical bayes james-stein estimation.",18
"deep, dense, low-rank gaussian conditional random fields. work introduce fully-connected graph structure deep gaussian conditional random field (g-crf) model. express pairwise interactions pixels inner-products low-dimensional embeddings, delivered new subnetwork deep architecture. efficiently minimize resulting energy solving resulting low-rank linear system conjugate gradients, derive analytic expression gradient embeddings allows us train end-to-end backpropagation. demonstrate merit approach achieving state art results three challenging computer vision benchmarks, namely semantic segmentation, human parts segmentation, saliency estimation. implementation fully gpu based, built top caffe library, made publicly available.",3
"decomposition quandle rings dihedral quandles. let $k = \mathbb r$ $\mathbb c$ $\mathcal r_n$ dihedral quandle order $n$: article, give decomposition quandle ring $k[\mathcal r_n]$ indecomposable right $k[\mathcal r_n]$-modules even $n \in \mathbb n$. follows decomposition $k[\mathcal r_n]$ given [eft19, prop. 4.18(2)] valid case $n$ divisible $4.$",10
"constrained classification ranking via quantiles. machine learning applications, classification accuracy primary metric interest. binary classifiers face class imbalance often evaluated $f_\beta$ score, area precision-recall curve, precision k, more. maximization many metrics expressed constrained optimization problem, constraint function classifier's predictions. paper propose novel framework learning constraints expressed predicted positive rate (or negative rate) subset training data. explicitly model threshold classifier must operate satisfy constraint, yielding surrogate loss function avoids complexity constrained optimization. method model-agnostic marginally expensive minimization unconstrained loss. experiments variety benchmarks show competitive performance relative existing baselines.",3
"metalda: topic model efficiently incorporates meta information. besides text content, documents associated words usually come rich sets meta informa- tion, categories documents semantic/syntactic features words, like encoded word embeddings. incorporating meta information directly generative process topic models improve modelling accuracy topic quality, especially case word-occurrence information training data insufficient. paper, present topic model, called metalda, able leverage either document word meta information, jointly. two data argumentation techniques, derive efficient gibbs sampling algorithm, benefits fully local conjugacy model. moreover, algorithm favoured sparsity meta information. extensive experiments several real world datasets demonstrate model achieves comparable improved performance terms perplexity topic quality, particularly handling sparse texts. addition, compared models using meta information, model runs significantly faster.",3
"interfacing constraint-based grammars generation algorithms. constraint-based grammars can, principle, serve major linguistic knowledge source parsing generation. surface generation starts input semantics representations may vary across grammars. many declarative grammars, concept derivation implicitly built parsing. may thus interpretable generation algorithm. show linguistically plausible semantic analyses cause severe problems semantic-head-driven approaches generation (shdg). use sereal, variant shdg disco grammar german source examples. propose new, general approach explicitly accounts interface grammar generation algorithm adding control-oriented layer linguistic knowledge base reorganizes semantics way suitable generation.",3
"novel transfer learning method based common space mapping weighted domain matching. paper, propose novel learning framework problem domain transfer learning. map data two domains one single common space, learn classifier common space. adapt common classifier two domains adding two adaptive functions respectively. common space, target domain data points weighted matched target domain term distributions. weighting terms source domain data points target domain classification responses also regularized local reconstruction coefficients. novel transfer learning framework evaluated benchmark cross-domain data sets, outperforms existing state-of-the-art transfer learning methods.",3
"alleviating media bias intelligent agent blogging. consumers mass media must comprehensive, balanced plural selection news get unbiased perspective; achieving goal challenging, laborious time consuming. news stories development time, (in)consistency, different level coverage across media outlets challenges conscientious reader overcome order alleviate bias. paper present intelligent agent framework currently facilitating analysis main sources on-line news el salvador. show prior tools text analysis web 2.0 technologies combined minimal manual intervention help individuals rational decision process, holding media outlets accountable work.",3
"improving max-sum decimation solve loopy distributed constraint optimization problems. context solving large distributed constraint optimization problems (dcop), belief-propagation approximate inference algorithms candidates choice. however, general, factor graph loopy (i.e. cyclic), solution methods suffer bad performance, due non-convergence many exchanged messages. improve performances max-sum inference algorithm solving loopy constraint optimization problems, propose take inspiration belief-propagation-guided dec-imation used solve sparse random graphs (k-satisfiability). propose novel decimaxsum method, parameterized terms policies decide trigger decimation, variables decimate, values assign decimated variables. based empirical evaluation classical bp benchmark (the ising model), combinations policies exhibit better performance state-of-the-art competitors.",3
"total capture: 3d deformation model tracking faces, hands, bodies. present unified deformation model markerless capture multiple scales human movement, including facial expressions, body motion, hand gestures. initial model generated locally stitching together models individual parts human body, refer ""frankenstein"" model. model enables full expression part movements, including face hands single seamless model. using large-scale capture people wearing everyday clothes, optimize frankenstein model create ""adam"". adam calibrated model shares skeleton hierarchy initial model express hair clothing geometry, making directly usable fitting people normally appear everyday life. finally, demonstrate use models total motion tracking, simultaneously capturing large-scale body movements subtle face hand motion social group people.",3
"automatic classification bright retinal lesions via deep network features. diabetic retinopathy timely diagonalized color eye fundus images experienced ophthalmologists, order recognize potential retinal features identify early-blindness cases. paper, proposed extract deep features last fully-connected layer of, four different, pre-trained convolutional neural networks. features feeded non-linear classifier discriminate three-class diabetic cases, i.e., normal, exudates, drusen. averaged across 1113 color retinal images collected six publicly available annotated datasets, deep features approach perform better classical bag-of-words approach. proposed approaches average accuracy 91.23% 92.00% 13% improvement traditional state art methods.",3
"tips, guidelines tools managing multi-label datasets: mldr.datasets r package cometa data repository. new proposals field multi-label learning algorithms growing number steadily last years. experimentation associated always goes phases: selection datasets, partitioning, training, analysis results and, finally, comparison existing methods. last step often hampered since involves using exactly datasets, partitioned way using validation strategy. paper present set tools whose objective facilitate management multi-label datasets, aiming standardize experimentation procedure. two main tools r package, mldr.datasets, web repository datasets, cometa. together, tools simplify collection datasets, partitioning, documentation export multiple formats, among functions. tips, recommendations guidelines good experimental analysis multi-label methods also presented.",3
"adversarial phenomenon eyes bayesian deep learning. deep learning models vulnerable adversarial examples, i.e.\ images obtained via deliberate imperceptible perturbations, model misclassifies high confidence. however, class confidence incomplete picture uncertainty. therefore use principled bayesian methods capture model uncertainty prediction observing adversarial misclassification. provide extensive study different bayesian neural networks attacked white-box black-box setups. behaviour networks noise, attacks clean test data compared. observe bayesian neural networks uncertain predictions adversarial perturbations, behaviour similar one observed random gaussian perturbations. thus, conclude bayesian neural networks considered detecting adversarial examples.",18
"robust kernel density estimation scaling projection hilbert space. robust parameter estimation well studied parametric density estimation, little investigation robust density estimation nonparametric setting. present robust version popular kernel density estimator (kde). estimators, robust version kde useful since sample contamination common issue datasets. ""robustness"" means nonparametric density estimate straightforward topic explore paper. construct robust kde scale traditional kde project nearest weighted kde $l^2$ norm. yields scaled projected kde (spkde). squared $l^2$ norm penalizes point-wise errors superlinearly causes weighted kde allocate weight high density regions. demonstrate robustness spkde numerical experiments consistency result shows asymptotically spkde recovers uncontaminated density sufficient conditions contamination.",18
"complex networks analysis language complexity. methods statistical physics, involving complex networks, increasingly used quantitative analysis linguistic phenomena. paper, represented pieces text different levels simplification co-occurrence networks found topological regularity correlated negatively textual complexity. furthermore, less complex texts distance concepts, represented nodes, tended decrease. complex networks metrics treated multivariate pattern recognition techniques, allowed us distinguish original texts simplified versions. original text, two simplified versions generated manually increasing number simplification operations. expected, distinction easier strongly simplified versions, relevant metrics node strength, shortest paths diversity. also, discrimination complex texts improved higher hierarchical network metrics, thus pointing usefulness considering wider contexts around concepts. though accuracy rate distinction high methods using deep linguistic knowledge, complex network approach still useful rapid screening texts whenever assessing complexity essential guarantee accessibility readers limited reading ability",14
"nonlinear stability asymptotic behavior periodic wave trains reaction-diffusion systems $c_{\mathrm{ub}}^2$-perturbations. present nonlinear stability theory periodic wave trains reaction-diffusion systems, relies pure $l^\infty$-estimates only. analysis shows localization periodicity requirements perturbations, present current literature, completely lifted. inspired previous works considering localized perturbations, decompose semigroup generated linearization wave train introduce spatio-temporal phase modulation capture critical dynamics, governed viscous burgers' equation. aim close nonlinear stability argument iterative estimates corresponding duhamel formulation, where, hampered lack localization, must rely diffusive smoothing render decay semigroup. yet, decay strong enough control terms duhamel formulation. address difficulty applying cole-hopf transform eliminate critical burgers'-type nonlinearities. ultimately, establish nonlinear stability diffusively spectrally stable wave trains $c_{\mathrm{ub}}^2$-perturbations. moreover, show perturbed solution converges modulated wave train, whose phase wavenumber approximated solutions associated viscous hamilton-jacobi burgers' equation, respectively.",10
"deep convolutional networks pancreas segmentation ct imaging. automatic organ segmentation important prerequisite many computer-aided diagnosis systems. high anatomical variability organs abdomen, pancreas, prevents many segmentation methods achieving high accuracies compared segmentation organs like liver, heart kidneys. recently, availability large annotated training sets accessibility affordable parallel computing resources via gpus made feasible ""deep learning"" methods convolutional networks (convnets) succeed image classification tasks. methods advantage used classification features trained directly imaging data. present fully-automated bottom-up method pancreas segmentation computed tomography (ct) images abdomen. method based hierarchical coarse-to-fine classification local image regions (superpixels). superpixels extracted abdominal region using simple linear iterative clustering (slic). initial probability response map generated, using patch-level confidences two-level cascade random forest classifiers, superpixel regions probabilities larger 0.5 retained. retained superpixels serve highly sensitive initial input pancreas surroundings convnet samples bounding box around superpixel different scales (and random non-rigid deformations training time) order assign distinct probability superpixel region pancreas not. evaluate method ct images 82 patients (60 training, 2 validation, 20 testing). using convnets achieve average dice scores 68%+-10% (range, 43-80%) testing. shows promise accurate pancreas segmentation, using deep learning approach compares favorably state-of-the-art methods.",3
automation feature engineering iot analytics. paper presents approach automation interpretable feature selection internet things analytics (iota) using machine learning (ml) techniques. authors conducted survey different people involved different iota based application development tasks. survey reveals feature selection time consuming niche skill demanding part entire workflow. paper shows feature selection successfully automated without sacrificing decision making accuracy thereby reducing project completion time cost hiring expensive resources. several pattern recognition principles state art (soa) ml techniques followed design overall approach proposed automation. three data sets considered establish proof-of-concept. experimental results show proposed automation able reduce time feature selection $2$ days instead $4-6$ months would required absence automation. reduction time achieved without sacrifice accuracy decision making process. proposed method also compared multi layer perceptron (mlp) model state art works iota uses mlp based deep learning. moreover feature selection method compared soa feature reduction technique namely principal component analysis (pca) variants. results obtained show proposed method effective.,18
"crl ntcir2. developed systems two types ntcir2. one enhenced version system developed ntcir1 irex. submitted retrieval results jj cc tasks. variety parameters tried system. used characteristics newspapers locational information cc tasks. system got good results tasks. system portable system avoids free parameters much possible. system submitted retrieval results jj, je, ee, ej, cc tasks. system automatically determined number top documents weight original query used automatic-feedback retrieval. also determined relevant terms quite robustly. ej je tasks, used document expansion augment initial queries. achieved good results, except cc tasks.",3
"log-densenet: sparsify densenet. skip connections increasingly utilized deep neural networks improve accuracy cost-efficiency. particular, recent densenet efficient computation parameters, achieves state-of-the-art predictions directly connecting feature layer previous ones. however, densenet's extreme connectivity pattern may hinder scalability high depths, applications like fully convolutional networks, full densenet connections prohibitively expensive. work first experimentally shows one key advantage skip connections short distances among feature layers backpropagation. specifically, using fixed number skip connections, connection patterns shorter backpropagation distance among layers accurate predictions. following insight, propose connection template, log-densenet, which, comparison densenet, slightly increases backpropagation distances among layers 1 ($1 + \log_2 l$), uses $l\log_2 l$ total connections instead $o(l^2)$. hence, log-densenets easier densenets implement scale. demonstrate effectiveness design principle showing better performance densenets tabula rasa semantic segmentation, competitive results visual recognition.",3
"variants rmsprop adagrad logarithmic regret bounds. adaptive gradient methods become recently popular, particular shown useful training deep neural networks. paper analyzed rmsprop, originally proposed training deep neural networks, context online convex optimization show $\sqrt{t}$-type regret bounds. moreover, propose two variants sc-adagrad sc-rmsprop show logarithmic regret bounds strongly convex functions. finally, demonstrate experiments new variants outperform adaptive gradient techniques stochastic gradient descent optimization strongly convex functions well training deep neural networks.",3
"initiation fiber fuse damage high-power ytterbium-doped fiber lasers. fiber fuse effect occur spontaneously propagate along optical fibers cause wide-spread damage; threatens applications involving optical fibers. paper presents two results. first, establishes initiation fiber fuse (iff) silica fibers caused defect-induced absorption. critical temperatures critical optical powers iff simulated first time using 3d solid-state heat transfer model heat source generated defect-induced absorption. method, formation energies defects uniquely determined, offers critical information chemical reasons fiber fuse. second, paper offers method evaluate operating temperatures fiber lasers. general analytical solutions operating temperatures along gain fibers deduced. results 976-nm laser-diode-(ld)-pumped 1018-nm tandem-pumped ytterbium-doped fiber (ydf) amplifiers using 10/130-{\mu}m ydfs calculated. potential limits caused fiber fuse discussed.",14
"paths equally spaced points circle. consider n points evenly spaced circle, path n-1 chords uses point once. m=floor(n/2) possible chord lengths, path defines multiset n-1 elements drawn {1,2,..,m}. first problem consider characterize multisets realized path. buratti conjectured multisets realized n prime, generalized conjecture n proposed horak rosa. previously conjecture proved n \leq 19 n=23; extend n\leq 37 (oeis sequence a352568). second problem determine number distinct (euclidean) path lengths realized. conjecture; extend current knowledge n\leq 16 n\leq 37 (oeis sequence a030077). n prime, twice prime, power 2, prove two paths length multiset chord lengths.",10
"modelling dependency completion sentence comprehension bayesian hierarchical mixture process: case study involving chinese relative clauses. present case-study demonstrating usefulness bayesian hierarchical mixture modelling investigating cognitive processes. sentence comprehension, widely assumed distance linguistic co-dependents affects latency dependency resolution: longer distance, longer retrieval time (the distance-based account). alternative theory, direct-access, assumes retrieval times mixture two distributions: one distribution represents successful retrievals (these independent dependency distance) represents initial failure retrieve correct dependent, followed reanalysis leads successful retrieval. implement models bayesian hierarchical models show direct-access model explains chinese relative clause reading time data better distance account.",18
"explainable graph theory-based identification meter-transformer mapping. distributed energy resources better environment may cause transformer overload distribution grids, calling recovering meter-transformer mapping provide situational awareness, i.e., transformer loading. challenge lies recovering meter-transformer (m.t.) mapping two common scenarios, e.g., large distances meter parent transformer high similarity meter's consumption pattern non-parent transformer's meters. past methods either assume variety data transmission grid ignore two common scenarios mentioned above. therefore, propose utilize observation via spectral embedding using property inter-transformer meter consumptions noise data limited k smallest eigenvalues voltage-based laplacian matrix smaller next smallest eigenvalue ideal laplacian matrix. also provide guarantee based understanding. furthermore, partially relax assumption utilizing location information aid voltage information areas geographically far away similar voltages. numerical simulations ieee test systems real feeders partner utility show proposed method correctly identifies m.t. mapping.",10
"increased phase coherence length porous topological insulator. surface area bi2te3 thin films increased introducing nanoscale porosity. temperature dependent resistivity magnetotransport measurements conducted as-grown porous samples (23 70 nm). longitudinal resistivity porous samples became metallic, indicating increased surface area resulted transport surface-like. weak antilocalization (wal) present samples, remarkably phase coherence length doubled porous samples. increase likely due large fermi velocity dirac surface states. results show introduction nanoporosity destroy topological surface states rather enhances them, making nanostructured materials promising low energy electronics, spintronics thermoelectrics.",2
mesa: maximum entropy simulated annealing. probabilistic reasoning systems combine different probabilistic rules probabilistic facts arrive desired probability values consequences. paper describe mesa-algorithm (maximum entropy simulated annealing) derives joint distribution variables propositions. takes account reliability probability values resolve conflicts contradictory statements. joint distribution represented terms marginal distributions therefore allows process large inference networks determine desired probability values high precision. procedure derives maximum entropy distribution subject given constraints. applied inference networks arbitrary topology may extended number directions.,3
"discriminative representation convolutional features indoor scene recognition. indoor scene recognition multi-faceted challenging problem due diverse intra-class variations confusing inter-class similarities. paper presents novel approach exploits rich mid-level convolutional features categorize indoor scenes. traditionally used convolutional features preserve global spatial structure, desirable property general object recognition. however, argue structuredness much helpful large variations scene layouts, e.g., indoor scenes. propose transform structured convolutional activations another highly discriminative feature space. representation transformed space incorporates discriminative aspects target dataset, also encodes features terms general object categories present indoor scenes. end, introduce new large-scale dataset 1300 object categories commonly present indoor scenes. proposed approach achieves significant performance boost previous state art approaches five major scene classification datasets.",3
"characterization homology $d$-manifolds $g_2\leq 3$. $g$-vector simplicial complex contains lot informations combinatorial topological structure complex. far several classification results structure normal pseudomanifolds homology manifolds given respect value $g_2$. known $g_2=0$, normal pseudomanifolds dimension least three stacked spheres. case $g_2=1$ $2$, prime homology manifolds polytopal spheres obtained sort retriangulations join operation previous one. article given combinatorial characterization homology $d$-manifolds $g_2=3$, $d\geq 3$ obtained operations like join, retriangulations connected sum. further, given structural result prime normal $d$-pseudomanifolds $g_2=3$. results together [9] classifies (combinatorially) normal $3$-pseudomanifolds $g_2=3$.",10
"rigidity 3d spherical caps via $μ$-bubbles. using gromov's $\mu$-bubble technique, show $3$-dimensional spherical caps rigid perturbations reduce metric, scalar curvature, mean curvature along boundary. several generalizations result discussed.",10
"convolutional neural network gravitational-wave early alert: going frequency. present latest development machine-learning pipeline pre-merger alerts gravitational waves coming binary neutron stars. work starts convolutional neural networks introduced previous paper (physrevd.103.102003) searched three classes early inspirals simulated gaussian noise colored design-sensitivity power-spectral density ligo. new network able search type binary neutron stars, take account detectors available, see events even earlier previous one. study performance method three different types noise: gaussian o3 noise, real o3 noise, predicted o4 noise. show network performs almost well non-gaussian noise gaussian noise: method robust w.r.t. glitches artifacts present real noise. although would able trigger bnss detected o3 signal-to-noise ratio weak, expect network find around 3 bnss o4 time merger 3 88 advance.",6
"hebbian/anti-hebbian neural network linear subspace learning: derivation multidimensional scaling streaming data. neural network models early sensory processing typically reduce dimensionality streaming input data. networks learn principal subspace, sense principal component analysis (pca), adjusting synaptic weights according activity-dependent learning rules. derived principled cost function rules nonlocal hence biologically implausible. time, biologically plausible local rules postulated rather derived principled cost function. here, bridge gap, derive biologically plausible network subspace learning streaming data minimizing principled cost function. departure previous work, cost quantified representation, reconstruction, error, adopt multidimensional scaling (mds) cost function streaming data. resulting algorithm relies biologically plausible hebbian anti-hebbian local learning rules. stochastic setting, synaptic weights converge stationary state projects input data onto principal subspace. data generated nonstationary distribution, network track principal subspace. thus, result makes step towards algorithmic theory neural computation.",15
"minimal cylinders three-dimensional heisenberg group. study minimal cylinders three-dimensional heisenberg group ${\rm nil}_3$ using generalized weierstrass type representation, so-called loop group method. characterize non-vertical minimal cylinders terms pairs two closed plane curves signed area. moreover, byproduct construction, spacelike cmc cylinders also obtained.",10
"towards asp-based architecture autonomous uavs dynamic environments (extended abstract). traditional ai reasoning techniques used successfully many domains, including logistics, scheduling game playing. paper part project aimed investigating techniques extended coordinate teams unmanned aerial vehicles (uavs) dynamic environments. specifically challenging real-world environments uavs network-enabled devices must communicate coordinate -- communication actions neither reliable free. network-centric environments common military, public safety commercial applications, yet research (even multi-agent planning) usually takes communications among distributed agents given. address challenge developing agent architecture reasoning algorithms based answer set programming (asp). although asp used successfully number applications, best knowledge first practical application complete asp-based agent architecture. also first practical application asp involving combination centralized reasoning, decentralized reasoning, execution monitoring, reasoning network communications.",3
"tunable gmm kernels. recently proposed ""generalized min-max"" (gmm) kernel efficiently linearized, direct applications large-scale statistical learning fast near neighbor search. linearized gmm kernel extensively compared linearized radial basis function (rbf) kernel. large number classification tasks, tuning-free gmm kernel performs (surprisingly) well compared best-tuned rbf kernel. nevertheless, one would naturally expect gmm kernel ought improved introduce tuning parameters. paper, study three simple constructions tunable gmm kernels: (i) exponentiated-gmm (or egmm) kernel, (ii) powered-gmm (or pgmm) kernel, (iii) exponentiated-powered-gmm (epgmm) kernel. pgmm kernel still efficiently linearized modifying original hashing procedure gmm kernel. 60 publicly available classification datasets, verify proposed tunable gmm kernels typically improve original gmm kernel. datasets, improvements astonishingly significant. example, 11 popular datasets used testing deep learning algorithms tree methods, experiments show proposed tunable gmm kernels strong competitors trees deep nets. previous studies developed tree methods including ""abc-robust-logitboost"" demonstrated excellent performance 11 datasets (and datasets), establishing second-order tree-split formula new derivatives multi-class logistic loss. compared tree methods like ""abc-robust-logitboost"" (which slow need substantial model sizes), tunable gmm kernels produce largely comparable results.",18
"generalized compression dictionary distance universal similarity measure. present new similarity measure based information theoretic measures superior normalized compression distance clustering problems inherits useful properties conditional kolmogorov complexity. show normalized compression dictionary size normalized compression dictionary entropy computationally efficient, need perform compression eliminated. also scale linearly exponential vector size growth content independent. show normalized compression dictionary distance compressor independent, limited lossless compressors, gives space optimizations implementation speed improvement real-time big data applications. introduced measure applicable machine learning tasks parameter-free unsupervised clustering, supervised learning classification regression, feature selection, applicable big data problems order magnitude speed increase.",18
"geometry discrete exponential families application exponential random graph models. explosion interest statistical models analyzing network data, considerable interest class exponential random graph (erg) models, especially connection difficulties computing maximum likelihood estimates. issues associated difficulties relate broader structure discrete exponential families. paper re-examines issues two parts. first consider closure $k$-dimensional exponential families distribution discrete base measure polyhedral convex support $\mathrm{p}$. show normal fan $\mathrm{p}$ geometric object plays fundamental role deriving statistical geometric properties corresponding extended exponential families. discuss relevance maximum likelihood estimation, theoretical computational standpoint. second, apply results analysis erg models. particular, means detailed example, provide characterization properties erg models, and, particular, certain behaviors erg models known degeneracy.",18
"competitive multi-agent inverse reinforcement learning sub-optimal demonstrations. paper considers problem inverse reinforcement learning zero-sum stochastic games expert demonstrations known optimal. compared previous works decouple agents game assuming optimality expert strategies, introduce new objective function directly pits experts nash equilibrium strategies, design algorithm solve reward function context inverse reinforcement learning deep neural networks model approximations. setting model algorithm decouple agent. order find nash equilibrium large-scale games, also propose adversarial training algorithm zero-sum stochastic games, show theoretical appeal non-existence local optima objective function. numerical experiments, demonstrate nash equilibrium inverse reinforcement learning algorithms address games amenable previous approaches using tabular representations. moreover, sub-optimal expert demonstrations algorithms recover reward functions strategies good quality.",18
"bidirectional deep-readout echo state networks. propose deep architecture classification multivariate time series. means recurrent untrained reservoir generate vectorial representation embeds temporal relationships data. improve memorization capability, implement bidirectional reservoir, whose last state captures also past dependencies input. apply dimensionality reduction final reservoir states obtain compressed fixed size representations time series. subsequently fed deep feedforward network trained perform final classification. test architecture benchmark datasets real-world use-case blood samples classification. results show method performs better standard echo state network and, time, achieves results comparable fully-trained recurrent network, faster training.",3
"consistent model data? information-theoretic model check. choice model class fundamental statistical learning system identification, matter whether class derived physical principles generic black-box. develop method evaluate specified model class assessing capability reproducing data similar observed data record. model check based information-theoretic properties models viewed data generators applicable e.g. sequential data nonlinear dynamical models. method understood specific two-sided posterior predictive test. apply information-theoretic model check synthetic real data compare classical whiteness test.",18
"differentially private bayesian learning distributed data. many applications machine learning, example health care, would benefit methods guarantee privacy data subjects. differential privacy (dp) become established standard protecting learning results. standard dp algorithms require single trusted party access entire data, clear weakness. consider dp bayesian learning distributed setting, party holds single sample samples data. propose learning strategy based secure multi-party sum function aggregating summaries data holders gaussian mechanism dp. method builds asymptotically optimal practically efficient dp bayesian inference rapidly diminishing extra cost.",18
"unsupervised segmentation action segments egocentric videos using gaze. unsupervised segmentation action segments egocentric videos desirable feature tasks activity recognition content-based video retrieval. reducing search space finite set action segments facilitates faster less noisy matching. however, exist substantial gap machine understanding natural temporal cuts continuous human activity. work reports novel gaze-based approach segmenting action segments videos captured using egocentric camera. gaze used locate region-of-interest inside frame. tracking two simple motion-based parameters inside successive regions-of-interest, discover finite set temporal cuts. present several results using combinations (of two parameters) dataset, i.e., brisgaze-actions. dataset contains egocentric videos depicting several daily-living activities. quality temporal cuts improved implementing two entropy measures.",3
"decentralized algorithms outperform centralized algorithms? case study decentralized parallel stochastic gradient descent. distributed machine learning systems nowadays, including tensorflow cntk, built centralized fashion. one bottleneck centralized algorithms lies high communication cost central node. motivated this, ask, decentralized algorithms faster centralized counterpart? although decentralized psgd (d-psgd) algorithms studied control community, existing analysis theory show advantage centralized psgd (c-psgd) algorithms, simply assuming application scenario decentralized network available. paper, study d-psgd algorithm provide first theoretical analysis indicates regime decentralized algorithms might outperform centralized algorithms distributed stochastic gradient descent. d-psgd comparable total computational complexities c-psgd requires much less communication cost busiest node. conduct empirical study validate theoretical analysis across multiple frameworks (cntk torch), different network configurations, computation platforms 112 gpus. network configurations low bandwidth high latency, d-psgd one order magnitude faster well-optimized centralized counterparts.",10
"sar image despeckling using convolutional. synthetic aperture radar (sar) images often contaminated multiplicative noise known speckle. speckle makes processing interpretation sar images difficult. propose deep learning-based approach called, image despeckling convolutional neural network (id-cnn), automatically removing speckle input noisy images. particular, id-cnn uses set convolutional layers along batch normalization rectified linear unit (relu) activation function component-wise division residual layer estimate speckle trained end-to-end fashion using combination euclidean loss total variation (tv) loss. extensive experiments synthetic real sar images show proposed method achieves significant improvements state-of-the-art speckle reduction methods.",3
"word-alignment-based segment-level machine translation evaluation using word embeddings. one important problems machine translation (mt) evaluation evaluate similarity translation hypotheses different surface forms reference, especially segment level. propose use word embeddings perform word alignment segment-level mt evaluation. performed experiments three types alignment methods using word embeddings. evaluated proposed methods various translation datasets. experimental results show proposed methods outperform previous word embeddings-based methods.",3
"low-rank matrix recovery row-and-column affine measurements. propose study row-and-column affine measurement scheme low-rank matrix recovery. measurement linear combination elements one row one column matrix $x$. setting arises naturally applications different domains. however, current algorithms developed standard matrix recovery problems perform well case, hence need developing new algorithms theory problem. propose simple algorithm problem based singular value decomposition ($svd$) least-squares ($ls$), term \alg. prove (a simplified version of) algorithm recover $x$ exactly minimum possible number measurements noiseless case. general noisy case, prove performance guarantees reconstruction accuracy frobenius norm. simulations, row-and-column design \alg algorithm show improved speed, comparable cases better accuracy compared standard measurements designs algorithms. theoretical experimental results suggest proposed row-and-column affine measurements scheme, together recovery algorithm, may provide powerful framework affine matrix reconstruction.",3
"time-contrastive networks: self-supervised learning video. propose self-supervised approach learning representations robotic behaviors entirely unlabeled videos recorded multiple viewpoints, study representation used two robotic imitation settings: imitating object interactions videos humans, imitating human poses. imitation human behavior requires viewpoint-invariant representation captures relationships end-effectors (hands robot grippers) environment, object attributes, body pose. train representations using metric learning loss, multiple simultaneous viewpoints observation attracted embedding space, repelled temporal neighbors often visually similar functionally different. words, model simultaneously learns recognize common different-looking images, different similar-looking images. signal causes model discover attributes change across viewpoint, change across time, ignoring nuisance variables occlusions, motion blur, lighting background. demonstrate representation used robot directly mimic human poses without explicit correspondence, used reward function within reinforcement learning algorithm. representations learned unlabeled collection task-related videos, robot behaviors pouring learned watching single 3rd-person demonstration human. reward functions obtained following human demonstrations learned representation enable efficient reinforcement learning practical real-world robotic systems. video results, open-source code dataset available https://sermanet.github.io/imitate",3
numerical weather prediction stochastic modeling: objective criterion choice global radiation forecasting. numerous methods exist developed global radiation forecasting. two popular types numerical weather predictions (nwp) predictions using stochastic approaches. propose compute parameter noted constructed part mutual information quantity measures mutual dependence two variables. calculated objective establish relevant method nwp stochastic models concerning current problem.,18
"context aware nonnegative matrix factorization clustering. article propose method refine clustering results obtained nonnegative matrix factorization (nmf) technique, imposing consistency constraints final labeling data. research community focused effort initialization optimization part method, without paying attention final cluster assignments. propose game theoretic framework object clustered represented player, choose cluster membership. information obtained nmf used initialize strategy space players weighted graph used model interactions among players. interactions allow players choose cluster coherent clusters chosen similar players, property guaranteed nmf, since produces soft clustering data. results common benchmarks show model able improve performances many nmf formulations.",3
"mixture model-based functional discriminant analysis curve classification. statistical approaches functional data analysis concern paradigm individuals functions curves rather finite dimensional vectors. paper, particularly focus modeling classification functional data temporal curves presenting regime changes time. specifically, propose new mixture model-based discriminant analysis approach functional data using specific hidden process regression model. approach particularly adapted handle problem complex-shaped classes curves, class composed several sub-classes, deal regime changes within homogeneous sub-class. model explicitly integrates heterogeneity class curves via mixture model formulation, regime changes within sub-class hidden logistic process. approach allows therefore fitting flexible curve-models class complex-shaped curves presenting regime changes unsupervised learning scheme, automatically summarize finite number homogeneous clusters, decomposed several regimes. model parameters learned maximizing observed-data log-likelihood class using dedicated expectation-maximization (em) algorithm. comparisons simulated data real data alternative approaches, including functional linear discriminant analysis functional mixture discriminant analysis polynomial regression mixtures spline regression mixtures, show proposed approach provides better results regarding discrimination results significantly improves curves approximation.",18
"any-gram kernels sentence classification: sentiment analysis case study. any-gram kernels flexible efficient way employ bag-of-n-gram features learning textual data. also compatible use word embeddings word similarities accounted for. original any-gram kernels implemented top tree kernels, propose new approach independent tree kernels efficient. also propose effective way make use word embeddings original any-gram formulation. applied task sentiment classification, new formulation achieves significantly better performance.",3
"general theory training learning machine. though deep learning pushing machine learning new stage, basic theories machine learning still limited. principle learning, role prior knowledge, role neuron bias, basis choosing neural transfer function cost function, etc., still far clear. paper, present general theoretical framework machine learning. classify prior knowledge common problem-dependent parts, consider aim learning maximally incorporate them. principle suggested maximizing former design risk minimization principle, neural transfer function, cost function, well pretreatment samples, endowed role maximizing latter. role neuron bias explained different angle. develop monte carlo algorithm establish input-output responses, control input-output sensitivity learning machine controlling individual neurons. applications function approaching smoothing, pattern recognition classification, provided illustrate train general learning machines based theory algorithm. method may addition induce new applications, transductive inference.",18
"efficient learning family nonconvex regularizers redistributing nonconvexity. use convex regularizers allows easy optimization, though often produce biased estimation inferior prediction performance. recently, nonconvex regularizers attracted lot attention outperformed convex ones. however, resultant optimization problem much harder. paper, large class nonconvex regularizers, propose move nonconvexity regularizer loss. nonconvex regularizer transformed familiar convex regularizer, resultant loss function still guaranteed smooth. learning convexified regularizer performed existing efficient algorithms originally designed convex regularizers (such proximal algorithm, frank-wolfe algorithm, alternating direction method multipliers stochastic gradient descent). extensions made convexified regularizer closed-form proximal step, loss function nonconvex, nonsmooth. extensive experiments variety machine learning application scenarios show optimizing transformed problem much faster running state-of-the-art original problem.",10
"surfacing contextual hate speech words within social media. social media platforms recently seen increase occurrence hate speech discourse led calls improved detection methods. rely annotated data, keywords, classification technique. approach provides good coverage, fall short dealing new terms produced online extremist communities act original sources words alternate hate speech meanings. code words (which created adopted words) designed evade automatic detection often benign meanings regular discourse. example, ""skypes"", ""googles"", ""yahoos"" instances words alternate meaning used hate speech. overlap introduces additional challenges relying keywords collection data specific hate speech, downstream classification. work, develop community detection approach finding extremist hate speech communities collecting data members. also develop word embedding model learns alternate hate speech meaning words demonstrate candidacy code words several annotation experiments, designed determine possible recognize word used hate speech without knowing alternate meaning. report inter-annotator agreement rate k=0.871, k=0.676 data drawn extremist community keyword approach respectively, supporting claim hate speech detection contextual task depend fixed list keywords. goal advance domain providing high quality hate speech dataset addition learned code words fed existing classification approaches, thus improving accuracy automated detection.",3
"beyondplanck cosmoglobe: open science, reproducibility, data longevity. beyondplanck cosmoglobe collaborations implemented first integrated bayesian end-to-end analysis pipeline cmb experiments. primary long-term motivation work develop common analysis platform supports efficient global joint analysis complementary radio, microwave, sub-millimeter experiments. strict prerequisite succeed broad participation cmb community, two foundational aspects program therefore reproducibility open science. paper, discuss efforts toward aim. also discuss measures toward facilitating easy code data distribution, community-based code documentation, user-friendly compilation procedures, etc. work represents first publicly released end-to-end cmb analysis pipeline includes raw data, source code, parameter files, documentation. argue complete pipeline release requirement major future publicly-funded cmb experiments, noting full public release significantly increases data longevity ensuring data quality improved whenever better processing techniques, complementary datasets, computing power become available, thereby also taxpayers' value money; providing raw data final products sufficient guarantee full reproducibility future.",0
"statistical modeling continuous speech recognition (csr)(invited talk). automatic continuous speech recognition (csr) sufficiently mature variety real world applications possible including large vocabulary transcription interactive spoken dialogues. paper reviews evolution statistical modelling techniques underlie current-day systems, specifically hidden markov models (hmms) n-grams. starting description speech signal parameterisation, various modelling assumptions consequences discussed. describes various techniques effects assumptions mitigated. despite progress made, limitations current modelling techniques still evident. paper therefore concludes brief review fundamental modelling work progress.",3
"neural method explicit mapping quasi-curvature locally linear embedding image retrieval. paper proposed new explicit nonlinear dimensionality reduction using neural networks image retrieval tasks. first proposed quasi-curvature locally linear embedding (qlle) training set. qlle guarantees linear criterion neighborhood sample. then, neural method (nm) proposed out-of-sample problem. combining qlle nm, provide explicit nonlinear dimensionality reduction approach efficient image retrieval. experimental results three benchmark datasets illustrate method get better performance state-of-the-art out-of-sample methods.",3
"instrumenting smt solver solve hybrid network reachability problems. pddl+ planning semantics rooted hybrid automata (ha) recent work shown modeled network has. addressing complexity nonlinear pddl+ planning requires space time efficient reasoning. unfortunately, existing solvers either address nonlinear dynamics natively support networks automata. present new algorithm, called hnsolve, guides variable selection dreal satisfiability modulo theories (smt) solver reasoning network encodings nonlinear pddl+ planning has. hnsolve tightly integrates dreal solving discrete abstraction ha network. hnsolve finds composite runs ha network ignore continuous variables, respect mode jumps synchronization labels. hnsolve admissibly detects dead-ends discrete abstraction, posts conflict clauses prune smt solver's search. evaluate benefits hnsolve algorithm pddl+ benchmark problems demonstrate performance respect prior work.",3
"mpead: multi-population ea diagrams. multi-population evolutionary algorithms are, nature, highly complex difficult describe. even two populations working concert (or opposition) present myriad potential configurations often difficult relate using text alone. little effort made, however, depict kinds systems, relying solely simple structural connections (related using ad hoc diagrams) populations often leaving crucial details. paper, propose notation accompanying formalism consistently powerfully depicting structures relationships within intuitive consistent way. using notation, examine simple co-evolutionary systems discover new configurations simple process ""drawing whiteboard"". finally, demonstrate even complex, highly-interconnected systems large numbers populations understood ease using advanced features formalism",3
"recurrent neural networks driver activity anticipation via sensory-fusion architecture. anticipating future actions human widely studied problem robotics requires spatio-temporal reasoning. work propose deep learning approach anticipation sensory-rich robotics applications. introduce sensory-fusion architecture jointly learns anticipate fuse information multiple sensory streams. architecture consists recurrent neural networks (rnns) use long short-term memory (lstm) units capture long temporal dependencies. train architecture sequence-to-sequence prediction manner, explicitly learns predict future given partial temporal context. introduce novel loss layer anticipation prevents over-fitting encourages early anticipation. use architecture anticipate driving maneuvers several seconds happen natural driving data set 1180 miles. context maneuver anticipation comes multiple sensors installed vehicle. approach shows significant improvement state-of-the-art maneuver anticipation increasing precision 77.4% 90.5% recall 71.2% 87.4%.",3
"new constraint tensor-to-scalar ratio $planck$ bicep/keck array data using profile likelihood. derive new upper bound tensor-to-scalar ratio parameter $r$ using frequentist profile likelihood method. vary relevant cosmological parameters $\lambda$cdm model, well nuisance parameters. unlike bayesian analysis using markov chain monte carlo (mcmc), analysis independent choice priors. using $planck$ public release 4, bicep/keck array 2018, $planck$ cmb lensing, bao data, find upper limit $r<0.037$ 95% c.l., similar bayesian mcmc result $r<0.038$ flat prior $r$ conditioned $planck$ lowleb covariance matrix.",0
"learning deep features source color laser printer identification based cascaded learning. color laser printers fast printing speed high resolution, forgeries using color laser printers cause significant harm society. source printer identification technique employed countermeasure forgeries. paper presents color laser printer identification method based cascaded learning deep neural networks. refiner network trained adversarial training refine synthetic dataset halftone color decomposition. halftone color decomposing convnet trained refined dataset, trained knowledge transferred printer identifying convnet enhance accuracy. robustness rotation scaling considered training process, considered existing methods. experiments performed eight color laser printers, performance compared several existing methods. experimental results clearly show proposed method outperforms existing source color laser printer identification methods.",3
"cascaded scene flow prediction using semantic segmentation. given two consecutive frames pair stereo cameras, 3d scene flow methods simultaneously estimate 3d geometry motion observed scene. many existing approaches use superpixels regularization, may predict inconsistent shapes motions inside rigidly moving objects. instead assume scenes consist foreground objects rigidly moving front static background, use semantic cues produce pixel-accurate scene flow estimates. cascaded classification framework accurately models 3d scenes iteratively refining semantic segmentation masks, stereo correspondences, 3d rigid motion estimates, optical flow fields. evaluate method challenging kitti autonomous driving benchmark, show accounting motion segmented vehicles leads state-of-the-art performance.",3
enumeration sentences compactness. presented julia meta-program discovers compact theories data exist. writes candidate theories julia validates: tossing bad theories keeping good theories. compactness measured metric: number space-time derivatives. underlying algorithm applicable wide variety combinatorics problems compactness serves cut search space.,3
"robust ensemble clustering using probability trajectories. although many successful ensemble clustering approaches developed recent years, still two limitations existing approaches. first, mostly overlook issue uncertain links, may mislead overall consensus process. second, generally lack ability incorporate global information refine local links. address two limitations, paper, propose novel ensemble clustering approach based sparse graph representation probability trajectory analysis. particular, present elite neighbor selection strategy identify uncertain links locally adaptive thresholds build sparse graph small number probably reliable links. argue small number probably reliable links lead significantly better consensus results using graph links regardless reliability. random walk process driven new transition probability matrix utilized explore global information graph. derive novel dense similarity measure sparse graph analyzing probability trajectories random walkers, based two consensus functions proposed. experimental results multiple real-world datasets demonstrate effectiveness efficiency approach.",18
"wavelet-based semantic features hyperspectral signature discrimination. hyperspectral signature classification quantitative analysis approach hyperspectral imagery performs detection classification constituent materials pixel level scene. classification procedure operated directly hyperspectral data performed using features extracted corresponding hyperspectral signatures containing information like signature's energy shape. paper, describe technique applies non-homogeneous hidden markov chain (nhmc) models hyperspectral signature classification. basic idea use statistical models (such nhmc) characterize wavelet coefficients capture spectrum semantics (i.e., structural information) multiple levels. experimental results show approach based nhmc models outperform existing approaches relevant classification tasks.",3
"image retrieval based binary signature ang s-kgraph. paper, introduce optimum approach querying similar images large digital-image databases. work based rbir (region-based image retrieval) method uses multiple regions key retrieval images. method significantly improves accuracy queries. however, also increases cost computing. reduce expensive computational cost, implement binary signature encoder maps image identification binary. order fasten lookup, binary signatures images classified help s-kgraph. finally, work evaluated corel's images.",3
"ensemble multi-task gaussian process regression multiple latent processes. multi-task/multi-output learning seeks exploit correlation among tasks enhance performance learning solving task independently. paper, investigate problem context gaussian processes (gps) propose new model learns mixture latent processes decomposing covariance matrix sum structured hidden components controlled latent gp input features ""weight"" tasks. sum structure, propose parallelizable parameter learning algorithm predetermined initialization ""weights"". also notice ensemble parameter learning approach using mini-batches training data reduces computation complexity learning also improves regression performance. evaluate model two datasets, smaller swiss jura dataset another relatively larger atms dataset noaa. substantial improvements observed compared established alternatives.",18
"precision boson-jet azimuthal decorrelation hadron colliders. azimuthal angular decorrelation vector boson jet sensitive qcd radiation, used probe quark-gluon plasma heavy-ion collisions. using recoil-free jet definition, sensitivity contamination soft radiation measurement reduced, complication non-global logarithms eliminated theoretical calculation. specifically consider $p_t^n$ recombination scheme, well $n\to \infty$ limit, known winner-take-all scheme. jet definitions also significantly simplify calculation track-based measurement, preferred due superior angular resolution. present detailed discussion factorization soft-collinear effective theory, revealing transverse momentum $\vec q_t$ complicated azimuthal angle. show potential glauber contributions spoil factorization formalism, least including order $\alpha_s^3$. resummation carried using renormalization group, necessary ingredients collected calculated. conclude detailed phenomenological study, finding enhanced matching correction high jet $p_t$ due electroweak collinear enhancement boson emission di-jets. also compare pythia event generator, finding observable robust hadronization underlying event.",8
"private posterior distributions variational approximations. privacy preserving mechanisms differential privacy inject additional randomness form noise data, beyond sampling mechanism. ignoring additional noise lead inaccurate invalid inferences. paper, incorporate privacy mechanism explicitly likelihood function treating original data missing, end goal estimating posterior distributions model parameters. leads principled way performing valid statistical inference using private data, however, corresponding likelihoods intractable. paper, derive fast accurate variational approximations tackle intractable likelihoods arise due privacy. focus estimating posterior distributions parameters naive bayes log-linear model, sufficient statistics model shared using differentially private interface. using simulation study, show posterior approximations outperform naive method ignoring noise addition mechanism.",18
"constructive algorithm feedforward neural networks medical diagnostic reasoning. research search alternatives resolution complex medical diagnosis human knowledge apprehended general fashion. successful application examples show human diagnostic capabilities significantly worse neural diagnostic system. research describes constructive neural network algorithm backpropagation; offer approach incremental construction nearminimal neural network architectures pattern classification. algorithm starts minimal number hidden units single hidden layer; additional units added hidden layer one time improve accuracy network get optimal size neural network. algorithm tested several benchmarking classification problems including cancer1, heart, diabetes good generalization ability.",3
"swsi: low-cost commercial-quality whole slide imaging system android ios smartphones. paper, scalable whole slide imaging (swsi), novel high-throughput, cost-effective robust whole slide imaging system android ios platforms introduced analyzed. swsi, mainstream smartphone connected optical eyepiece manually controlled microscope automatically controlled capture sequences mega-pixel fields views synthesized giga-pixel virtual slides. remote servers carry majority computation asynchronously support clients running satisfying frame rates without sacrificing image quality robustness. typical 15x15mm sample digitized 30 seconds 4x 3 minutes 10x object magnification, costing $1. virtual slide quality considered comparable existing high-end scanners thus satisfying clinical usage surveyed pathologies. scan procedure features supporting magnification 100x, recoding z-stacks, specimen-type-neutral giving real-time feedback, deemed work-flow-friendly reliable.",14
"deep nonparametric estimation discrete conditional distributions via smoothed dyadic partitioning. present approach deep estimation discrete conditional probability distributions. models several applications, including generative modeling audio, image, video data. approach combines two main techniques: dyadic partitioning graph-based smoothing discrete space. recursively decomposing dimension series binary splits smoothing resulting distribution using graph-based trend filtering, impose strict structure model achieve much higher sample efficiency. demonstrate advantages model series benchmarks synthetic real-world datasets, cases reducing error nearly half comparison popular methods literature. models implemented tensorflow publicly available https://github.com/tansey/sdp .",18
"shift-net: image inpainting via deep feature rearrangement. deep convolutional networks (cnns) exhibited potential image inpainting producing plausible results.however, existing methods, e.g., context encoder, missing parts predicted propagating surrounding convolutional features fully connected layer, intends produce semantically plausible blurry result. paper, introduce special shift-connection layer u-net architecture, namely shift-net, filling missing regions shape sharp structures fine-detailed textures. end, encoder feature known region shifted serve estimation missing parts. guidance loss introduced decoder feature minimize distance decoder feature fully connected layer ground truth encoder feature missing parts. constraint, decoder feature missing region used guide shift encoder feature known region. end-to-end learning algorithm developed train shift-net. experiments paris streetview places datasets demonstrate efficiency effectiveness shift-net producing sharper, fine-detailed, visually plausible results.",3
"learning compose neural networks question answering. describe question answering model applies images structured knowledge bases. model uses natural language strings automatically assemble neural networks collection composable modules. parameters modules learned jointly network-assembly parameters via reinforcement learning, (world, question, answer) triples supervision. approach, term dynamic neural model network, achieves state-of-the-art results benchmark datasets visual structured domains.",3
"linear time computation moments sum-product networks. bayesian online algorithms sum-product networks (spns) need update posterior distribution seeing one single additional instance. so, must compute moments model parameters distribution. best existing method computing moments scales quadratically size spn, although scales linearly trees. unfortunate scaling makes bayesian online algorithms prohibitively expensive, except small tree-structured spns. propose optimal linear-time algorithm works even spn general directed acyclic graph (dag), significantly broadens applicability bayesian online algorithms spns. three key ingredients design analysis algorithm: 1). edge graph, construct linear time reduction moment computation problem joint inference problem spns. 2). using property spn computes multilinear polynomial, give efficient procedure polynomial evaluation differentiation without expanding network may contain exponentially many monomials. 3). propose dynamic programming method reduce computation moments edges graph quadratic linear. demonstrate usefulness linear time algorithm applying develop linear time assume density filter (adf) spns.",3
"deep learning markov random field semantic segmentation. semantic segmentation tasks well modeled markov random field (mrf). paper addresses semantic segmentation incorporating high-order relations mixture label contexts mrf. unlike previous works optimized mrfs using iterative algorithm, solve mrf proposing convolutional neural network (cnn), namely deep parsing network (dpn), enables deterministic end-to-end computation single forward pass. specifically, dpn extends contemporary cnn model unary terms additional layers devised approximate mean field (mf) algorithm pairwise terms. several appealing properties. first, different recent works required many iterations mf back-propagation, dpn able achieve high performance approximating one iteration mf. second, dpn represents various types pairwise terms, making many existing models special cases. furthermore, pairwise terms dpn provide unified framework encode rich contextual information high-dimensional data, images videos. third, dpn makes mf easier parallelized speeded up, thus enabling efficient inference. dpn thoroughly evaluated standard semantic image/video segmentation benchmarks, single dpn model yields state-of-the-art segmentation accuracies pascal voc 2012, cityscapes dataset camvid dataset.",3
"fitness uniform deletion: simple way preserve diversity. commonly experienced problem population based optimisation methods gradual decline population diversity tends occur time. slow system's progress even halt completely population converges local optimum cannot escape. paper present fitness uniform deletion scheme (fuds), simple somewhat unconventional approach problem. fuds deletion operation modified delete individuals ""common"" sense exist many individuals similar fitness population. makes impossible population collapse collection highly related individuals similar fitness. experimental results range optimisation problems confirm this, particular deceptive optimisation problems performance significantly robust variation selection intensity.",3
"intelligent emergency message broadcasting vanet using pso. new type mobile ad hoc network called vehicular ad hoc networks (vanet) created fertile environment research. research, protocol particle swarm optimization contention based broadcast (pcbb) proposed, fast andeffective dissemination emergency messages within geographical area distribute emergency message achieve safety system, research help vanet system achieve safety goals intelligent efficient way.",3
"unveiling relationship complex networks metrics word senses. automatic disambiguation word senses (i.e., identification meanings used given context word multiple meanings) essential applications machine translation information retrieval, represents key step developing so-called semantic web. humans disambiguate words straightforward fashion, apply computers. paper address problem word sense disambiguation (wsd) treating texts complex networks, show word senses distinguished upon characterizing local structure around ambiguous words. goal obtain best possible disambiguation system, nevertheless found half cases approach outperforms traditional shallow methods. show hierarchical connectivity clustering words usually relevant features wsd. results reported shine light relationship semantic structural parameters complex networks. also indicate combined traditional techniques complex network approach may useful enhance discrimination senses large texts",14
"learning combinations sigmoids gradient estimation. develop new approach learn parameters regression models hidden variables. nutshell, estimate gradient regression function set random points, cluster estimated gradients. centers clusters used estimates parameters hidden units. justify approach studying toy model, whereby regression function linear combination sigmoids. prove indeed estimated gradients concentrate around parameter vectors hidden units, provide non-asymptotic bounds number required samples. best knowledge, comparable guarantees proven linear combinations sigmoids.",18
"optimal distributed algorithm operator extrapolation stochastic aggregative games. work studies nash equilibrium seeking class stochastic aggregative games, player expectation-valued objective function depending local strategy aggregate players' strategies. propose distributed algorithm operator extrapolation, player maintains estimate aggregate exchanging information neighbors time-varying network, updates decision mirror descent method. operator extrapolation search direction applied two step historical gradient samples utilized accelerate convergence. strongly monotone assumption pseudo-gradient mapping, prove proposed algorithm achieve optimal convergence rate $\mathcal{o}(1/k)$ nash equilibrium seeking stochastic games. finally, algorithm performance demonstrated via numerical simulations.",10
"approximate bayesian inference gauge theory. published paper [sengupta, 2016], proposed brain (and self-organized biological artificial systems) characterized via mathematical apparatus gauge theory. picture emerges approach suggests biological system (from neuron organism) cast resolving uncertainty external milieu, either changing internal states relationship environment. using formal arguments, shown gauge theory neuronal dynamics -- based approximate bayesian inference -- potential shed new light phenomena thus far eluded formal description, attention link action perception. here, describe technical apparatus enables variational inference manifolds. particularly, novel contribution paper algorithm utlizes schild's ladder parallel transport sufficient statistics (means, covariances, etc.) statistical manifold.",15
"reward maximization uncertainty: leveraging side-observations networks. study stochastic multi-armed bandit (mab) problem presence side-observations across actions occur result underlying network structure. model, bipartite graph captures relationship actions common set unknowns choosing action reveals observations unknowns connected to. models common scenario online social networks users respond friends' activity, thus providing side information other's preferences. contributions follows: 1) derive asymptotic lower bound (with respect time) function bi-partite network structure regret uniformly good policy achieves maximum long-term average reward. 2) propose two policies - randomized policy; policy based well-known upper confidence bound (ucb) policies - explore action rate function network position. show, mild assumptions, policies achieve asymptotic lower bound regret multiplicative factor, independent network structure. finally, use numerical examples real-world social network routing example network demonstrate benefits obtained policies existing policies.",3
"semi-supervised clustering methods. cluster analysis methods seek partition data set homogeneous subgroups. useful wide variety applications, including document processing modern genetics. conventional clustering methods unsupervised, meaning outcome variable anything known relationship observations data set. many situations, however, information clusters available addition values features. example, cluster labels observations may known, certain observations may known belong cluster. cases, one may wish identify clusters associated particular outcome variable. review describes several clustering algorithms (known ""semi-supervised clustering"" methods) applied situations. majority methods modifications popular k-means clustering method, several described detail. brief description semi-supervised clustering algorithms also provided.",18
"traits personable. information personalization fertile ground application ai techniques. article relate personalization ability capture partial information information-seeking interaction. specific focus personalizing interactions web sites. using ideas partial evaluation explanation-based generalization, present modeling methodology reasoning personalization. approach helps identify seven tiers `personable traits' web sites.",3
"supercomputers strong coupling gravity curvature torsion. many theories gravity spoiled strongly coupled modes: high computational cost hamiltonian analysis obstruct identification modes. computer algebra implementation hamiltonian constraint algorithm curvature torsion theories presented. non-riemannian poincar\'e gauge theories suffer notoriously strong coupling. implementation forms package (the `hamiltonian gauge gravity surveyor' - higgs) xact tensor manipulation suite mathematica. poisson brackets evaluated parallel, meaning hamiltonian analysis done silicon, scale. accordingly higgs designed survey whole lagrangian space high-performance computing resources (clusters supercomputers). demonstrate this, space `outlawed' poincar\'e gauge theories surveyed, massive parity-even/odd vector parity-odd tensor torsion particle accompanies usual graviton. survey spans possible configurations teleparallel-style multiplier fields might used kill-off strongly coupled modes, results analysed subsequent work. brackets known primary secondary constraints theories made available future study. demonstrations also given using higgs - desktop computer - run dirac-bergmann algorithm specific theories, einstein-cartan theory minimal extensions.",6
"cultural diffusion trends facebook photographs. online social media social vehicle people share various moments lives friends, playing sports, cooking dinner taking selfie fun, via visual means, is, photographs. study takes closer look popular visual concepts illustrating various cultural lifestyles aggregated, de-identified photographs. perform analysis macroscopic microscopic levels, gain novel insights global local visual trends well dynamics interpersonal cultural exchange diffusion among facebook friends. processed images automatically classifying visual content convolutional neural network (cnn). various statistical tests, find socially tied individuals likely post images showing similar cultural lifestyles. identify main cause observed social correlation, use shuffle test preference-based matched estimation (pme) test distinguish effects influence homophily. results indicate visual content user's photographs temporally, although necessarily causally, correlated photographs friends, may suggest effect influence. paper demonstrates facebook photographs exhibit diverse cultural lifestyles preferences social interaction mediated visual channel social media effective mechanism cultural diffusion.",3
"low-rank modeling applications image analysis. low-rank modeling generally refers class methods solve problems representing variables interest low-rank matrices. achieved great success various fields including computer vision, data mining, signal processing bioinformatics. recently, much progress made theories, algorithms applications low-rank modeling, exact low-rank matrix recovery via convex programming matrix completion applied collaborative filtering. advances brought attentions topic. paper, review recent advance low-rank modeling, state-of-the-art algorithms, related applications image analysis. first give overview concept low-rank modeling challenging problems area. then, summarize models algorithms low-rank matrix recovery illustrate advantages limitations numerical experiments. next, introduce applications low-rank modeling context image analysis. finally, conclude paper discussions.",3
"influence-receptivity model topic based information cascades. consider problem estimating latent structure social network based observational data information diffusion processes, {\it cascades}. given cascade, observe time node/agent infected source infection. existing literature focused estimating network diffusion matrix without underlying assumptions structure network. propose novel model inferring network diffusion matrix based intuition information datum likely propagate among two nodes interested similar topics, common information. particular, model endows node influence vector (how authoritative topic) receptivity vector (how susceptible topic). show node-topic structure estimated observed cascades. estimated model used build recommendation system based receptivity vectors, well marketing based influence vectors.",18
"context-aware learning using transferable features classification breast cancer histology images. convolutional neural networks (cnns) recently used variety histology image analysis. however, availability large dataset major prerequisite training cnn limits use computational pathology community. previous studies, cnns demonstrated potential terms feature generalizability transferability accompanied better performance. considering traits cnn, propose simple yet effective method leverages strengths cnn combined advantages including contextual information, particularly designed small dataset. method consists two main steps: first uses activation features cnn trained patch-based classification trains separate classifier using features overlapping patches perform image-based classification using contextual information. proposed framework outperformed state-of-the-art method breast cancer classification.",3
acoustic modeling using shallow cnn-htsvm architecture. high-accuracy speech recognition especially challenging large datasets available. possible bridge gap careful knowledge-driven parsing combined biologically inspired cnn learning guarantees vapnik chervonenkis (vc) theory. work presents shallow-cnn-htsvm (hierarchical tree support vector machine classifier) architecture uses predefined knowledge-based set rules statistical machine learning techniques. show gross errors present even state-of-the-art systems avoided accurate acoustic model built hierarchical fashion. cnn-htsvm acoustic model outperforms traditional gmm-hmm models htsvm structure outperforms mlp multi-class classifier. importantly isolate performance acoustic model provide results frame phoneme level considering true robustness model. show even small amount data accurate robust recognition rates obtained.,3
"gpflow: gaussian process library using tensorflow. gpflow gaussian process library uses tensorflow core computations python front end. distinguishing features gpflow uses variational inference primary approximation method, provides concise code use automatic differentiation, engineered particular emphasis software testing able exploit gpu hardware.",18
"inventory control involving unknown demand discrete nonperishable items - analysis newsvendor-based policy. inventory control unknown demand distribution considered, emphasis placed case involving discrete nonperishable items. focus adaptive policy every period uses, much possible, optimal newsvendor ordering quantity empirical distribution learned period. policy assessed using regret criterion, measures price paid ambiguity demand distribution $t$ periods. guarantees latter's separation critical newsvendor parameter $\beta=b/(h+b)$, constant upper bound regret found. without prior information demand distribution, show regret grow faster rate $t^{1/2+\epsilon}$ $\epsilon>0$. view known lower bound, almost best one could hope for. simulation studies involving along policies also conducted.",18
"predictive analytics approach reducing avoidable hospital readmission. hospital readmission become critical metric quality cost healthcare. medicare anticipates nearly $17 billion paid 20% patients readmitted within 30 days discharge. although several interventions transition care management discharge reengineering practiced recent years, effectiveness sustainability depends well identify target patients high risk rehospitalization. based literature, current risk prediction models fail reach acceptable accuracy level; none considers patient's history readmission impacts patient attribute changes time; often discriminate planned unnecessary readmissions. tackling drawbacks, develop new readmission metric based administrative data identify potentially avoidable readmissions types readmission. propose tree based classification method estimate predicted probability readmission directly incorporate patient's history readmission risk factors changes time. proposed methods validated 2011-12 veterans health administration data inpatients hospitalized heart failure, acute myocardial infarction, pneumonia, chronic obstructive pulmonary disease state michigan. results shows improved discrimination power compared literature (c-statistics>80%) good calibration.",18
"towards stability optimality stochastic gradient descent. iterative procedures parameter estimation based stochastic gradient descent allow estimation scale massive data sets. however, theory practice, suffer numerical instability. moreover, statistically inefficient estimators true parameter value. address two issues, propose new iterative procedure termed averaged implicit sgd (ai-sgd). statistical efficiency, ai-sgd employs averaging iterates, achieves optimal cram\'{e}r-rao bound strong convexity, i.e., optimal unbiased estimator true parameter value. numerical stability, ai-sgd employs implicit update iteration, related proximal operators optimization. practice, ai-sgd achieves competitive performance state-of-the-art procedures. furthermore, stable averaging procedures employ proximal updates, simple implement requires fewer tunable hyperparameters procedures employ proximal updates.",18
"latent constrained correlation filter. correlation filters special classifiers designed shift-invariant object recognition, robust pattern distortions. recent literature shows combining set sub-filters trained based single small group images obtains best performance. idea equivalent estimating variable distribution based data sampling (bagging), interpreted finding solutions (variable distribution approximation) directly sampled data space. however, methodology fails account variations existed data. paper, introduce intermediate step -- solution sampling -- data sampling step form subspace, optimal solution estimated. specifically, propose new method, named latent constrained correlation filters (lccf), mapping correlation filters given latent subspace, develop new learning framework latent subspace embeds distribution-related constraints original problem. solve optimization problem, introduce subspace based alternating direction method multipliers (sadmm), proven converge saddle point. approach successfully applied three different tasks, including eye localization, car detection object tracking. extensive experiments demonstrate lccf outperforms state-of-the-art methods. source code publicly available. https://github.com/bczhangbczhang/.",3
"cyborg astrobiologist: scouting red beds uncommon features geological significance. `cyborg astrobiologist' (ca) undergone second geological field trial, red sandstone site northern guadalajara, spain, near riba de santiuste. cyborg astrobiologist wearable computer video camera system demonstrated capability find uncommon interest points geological imagery real-time field. first (of three) geological structures studied outcrop nearly homogeneous sandstone, exhibits oxidized-iron impurities red absence iron impurities white. white areas ``red beds'' turned white iron removed chemical reduction, perhaps biological agent. computer vision system found one instance several (iron-free) white spots uncommon therefore interesting, well several small dark nodules. second geological structure contained white, textured mineral deposits surface sandstone, found ca interesting. third geological structure 50 cm thick paleosol layer, fossilized root structures plants, found ca interesting. quasi-blind comparison cyborg astrobiologist's interest points images interest points determined afterwards human geologist shows cyborg astrobiologist concurred human geologist 68% time (true positive rate), 32% false positive rate 32% false negative rate. (abstract abridged).",3
"contour polygonal approximation using shortest path networks. contour polygonal approximation simplified representation contour line segments, main characteristics contour remain small number line segments. paper presents novel method polygonal approximation based complex networks theory. convert point contour vertex, model regular network. transform network small-world complex network applying transformations edges. analyzing network properties, especially geodesic path, compute polygonal approximation. paper presents main characteristics method, well functionality. evaluate proposed method using benchmark contours, compare results polygonal approximation methods.",14
"stochastic model human visual attention dynamic bayesian network. recent studies field human vision science suggest human responses stimuli visual display non-deterministic. people may attend different locations visual input time. based knowledge, propose new stochastic model visual attention introducing dynamic bayesian network predict likelihood humans typically focus video scene. proposed model composed dynamic bayesian network 4 layers. model provides framework simulates combines visual saliency response cognitive state person estimate probable attended regions. sample-based inference markov chain monte-carlo based particle filter stream processing multi-core processors enable us estimate human visual attention near real time. experimental results demonstrated model performs significantly better predicting human visual attention compared previous deterministic models.",3
"survey statistical network models. networks ubiquitous science become focal point discussion everyday life. formal statistical models analysis network data emerged major topic interest diverse areas study, involve form graphical representation. probability models graphs date back 1959. along empirical studies social psychology sociology 1960s, early works generated active network community substantial literature 1970s. effort moved statistical literature late 1970s 1980s, past decade seen burgeoning network literature statistical physics computer science. growth world wide web emergence online networking communities facebook, myspace, linkedin, host specialized professional network communities intensified interest study networks network data. goal review provide reader entry point burgeoning literature. begin overview historical development statistical network modeling introduce number examples studied network literature. subsequent discussion focuses number prominent static dynamic network models interconnections. emphasize formal model descriptions, pay special attention interpretation parameters estimation. end description open problems challenges machine learning statistics.",18
"objective prior unifies objective bayes information-based inference. three principle paradigms statistical inference: (i) bayesian, (ii) information-based (iii) frequentist inference. describe objective prior (the weighting $w$-prior) unifies objective bayes information-based inference. $w$-prior chosen make marginal probability unbiased estimator predictive performance model. definition several natural interpretations. perspective information content prior, $w$-prior uniformly maximally uninformative. $w$-prior also understood result uniform density distinguishable models parameter space. finally demonstrate $w$-prior equivalent akaike information criterion (aic) regular models asymptotic limit. $w$-prior appears generically applicable statistical inference free {\it ad hoc} regularization. mechanism suppressing complexity analogous aic: model complexity reduces model predictivity. expect new objective-bayes approach inference widely-applicable machine-learning problems including singular models.",18
"multivariate hawkes processes large-scale inference. paper, present framework fitting multivariate hawkes processes large-scale problems number events observed history $n$ number event types $d$ (i.e. dimensions). proposed low-rank hawkes process (lrhp) framework introduces low-rank approximation kernel matrix allows perform nonparametric learning $d^2$ triggering kernels using $o(ndr^2)$ operations, $r$ rank approximation ($r \ll d,n$). comes major improvement existing state-of-the-art inference algorithms $o(nd^2)$. furthermore, low-rank approximation allows lrhp learn representative patterns interaction event types, may valuable analysis complex processes real world datasets. efficiency scalability approach illustrated numerical experiments simulated well real datasets.",18
"neural dynamic programming musical self similarity. present neural sequence model designed specifically symbolic music. model based learned edit distance mechanism generalises classic recursion computer science, leading neural dynamic program. repeated motifs detected learning transformations them. represent arising computational dependencies using novel data structure, edit tree; perspective suggests natural approximations afford scaling otherwise cubic time algorithm. demonstrate model real synthetic data; cases out-performs strong stacked long short-term memory benchmark.",3
"note pac bayesian theorem. prove general exponential moment inequalities averages [0,1]-valued iid random variables use tighten pac bayesian theorem. logarithmic dependence sample count enumerator pac bayesian bound halved.",3
"adaptive-rate sparse signal reconstruction application compressive background subtraction. propose analyze online algorithm reconstructing sequence signals limited number linear measurements. signals assumed sparse, unknown support, evolve time according generic nonlinear dynamical model. algorithm, based recent theoretical results $\ell_1$-$\ell_1$ minimization, recursive computes number measurements taken time on-the-fly. example, apply algorithm compressive video background subtraction, problem stated follows: given set measurements sequence images static background, simultaneously reconstruct image separating foreground background. performance method illustrated sequences real images: observe allows dramatic reduction number measurements respect state-of-the-art compressive background subtraction schemes.",10
"deep learning spatio-temporal modeling: dynamic traffic flows high frequency trading. deep learning applies hierarchical layers hidden variables construct nonlinear high dimensional predictors. goal develop train deep learning architectures spatio-temporal modeling. training deep architecture achieved stochastic gradient descent (sgd) drop-out (do) parameter regularization goal minimizing out-of-sample predictive mean squared error. illustrate methodology, predict sharp discontinuities traffic flow data, secondly, develop classification rule predict short-term futures market prices function order book depth. finally, conclude directions future research.",18
"measuring sample quality kernels. approximate markov chain monte carlo (mcmc) offers promise rapid sampling cost biased inference. since standard mcmc diagnostics fail detect biases, researchers developed computable stein discrepancy measures provably determine convergence sample target distribution. approach recently combined theory reproducing kernels define closed-form kernel stein discrepancy (ksd) computable summing kernel evaluations across pairs sample points. develop theory weak convergence ksds based stein's method, demonstrate commonly used ksds fail detect non-convergence even gaussian targets, show kernels slowly decaying tails provably determine convergence large class target distributions. resulting convergence-determining ksds suitable comparing biased, exact, deterministic sample sequences simpler compute parallelize alternative stein discrepancies. use tools compare biased samplers, select sampler hyperparameters, improve upon existing ksd approaches one-sample hypothesis testing sample quality improvement.",18
"information field based global bayesian inference jet transport coefficient. bayesian statistical inference powerful tool model-data comparisons extractions physical parameters often unknown functions system variables. existing bayesian analyses often rely explicit parametrizations unknown function. introduce long-range correlations impose fictitious constraints physical parameters regions variable space probed experimental data. develop information field (if) approach modeling prior distribution unknown function free long-range correlations. apply approach first global bayesian inference jet transport coefficient $\hat q$ function temperature ($t$) existing experimental data single-inclusive hadron, di-hadron $\gamma$-hadron spectra heavy-ion collisions rhic lhc energies. extracted $\hat q/t^3$ exhibits strong $t$-dependence result progressive constraining power data central collisions higher colliding energies incrementally included. method guarantees extracted $t$-dependence biased specific functional form.",8
"time series prediction : predicting stock price. time series forecasting widely used multitude domains. paper, present four models predict stock price using spx index input time series data. martingale ordinary linear models require strongest assumption stationarity use baseline models. generalized linear model requires lesser assumptions unable outperform martingale. empirical testing, rnn model performs best comparing two models, update input lstm instantaneously, also beat martingale. addition, introduce online batch algorithm discrepancy measure inform readers newest research time series predicting method, require stationarity non mixing assumptions time series data. finally, apply forecasting practice, introduce basic trading strategies create win win zero sum situations.",18
"generalized canonical correlation analysis classification. multiple multivariate data sets, derive conditions generalized canonical correlation analysis (gcca) improves classification performance projected datasets, compared standard canonical correlation analysis (cca) using two data sets. illustrate theoretical results simulations real data experiment.",18
"adversarial posenet: structure-aware convolutional network human pose estimation. human pose estimation monocular images, joint occlusions overlapping upon human bodies often result deviated pose predictions. circumstances, biologically implausible pose predictions may produced. contrast, human vision able predict poses exploiting geometric constraints joint inter-connectivity. address problem incorporating priors structure human bodies, propose novel structure-aware convolutional network implicitly take priors account training deep network. explicit learning constraints typically challenging. instead, design discriminators distinguish real poses fake ones (such biologically implausible ones). pose generator (g) generates results discriminator fails distinguish real ones, network successfully learns priors.",3
"automatic generation synthetic colonoscopy videos domain randomization. increasing number colonoscopic guidance assistance systems rely machine learning algorithms require large amount high-quality training data. order ensure high performance, latter resemble substantial portion possible configurations. particularly addresses varying anatomy, mucosa appearance image sensor characteristics likely deteriorated motion blur inadequate illumination. limited amount readily available training data hampers account possible configurations results reduced generalization capabilities machine learning models. propose exemplary solution synthesizing colonoscopy videos substantial appearance anatomical variations enables learn discriminative domain-randomized representations interior colon mimicking real-world settings.",5
"hilbert scheme computer vision. multiview geometry study two-dimensional images three-dimensional scenes, foundational subject computer vision. determine universal groebner basis multiview ideal n generic cameras. cameras move, multiview varieties vary family dimension 11n-15. family distinguished component multigraded hilbert scheme unique borel-fixed point. present combinatorial study ideals lying hilbert scheme.",10
"efficient knowledge base management dcsp. dcsp (distributed constraint satisfaction problem) important research area ai (artificial intelligence). many application problems distributed ai formalized dscps. increasing complexity problem size application problems ai, required storage place searching average searching time increasing too. thus, use limited storage place efficiently solving dcsp becomes important problem, help reduce searching time well. paper provides efficient knowledge base management approach based general usage hyper-resolution-rule consistence algorithm. approach minimizes increasing knowledge base eliminate sufficient constraint false nogood. eliminations change completeness original knowledge base increased. proofs given well. example shows approach decrease new nogoods generated knowledge base greatly. thus decreases required storage place simplify searching process.",3
"mutual information fitting deep nonlinear models. deep nonlinear models pose challenge fitting parameters due lack knowledge hidden layer potentially non-affine relation initial observed layers. present work investigate use information theoretic measures mutual information kullback-leibler (kl) divergence objective functions fitting models without knowledge hidden layer. investigate one model proof concept one application cogntive performance. investigate use optimizers methods. mutual information largely successful objective, depending parameters. kl divergence found similarly succesful, given knowledge statistics hidden layer.",10
"developing improved greedy crossover solve symmetric traveling salesman problem. traveling salesman problem (tsp) one famous optimization problems. greedy crossover designed greffenstette et al, used symmetric tsp (stsp) resolved genetic algorithm (ga). researchers proposed several versions greedy crossover. propose improved version it. compare greedy crossover recent crossovers, use greedy crossover recent crossovers ga compare crossovers speed accuracy.",3
"newton sketch: linear-time optimization algorithm linear-quadratic convergence. propose randomized second-order method optimization known newton sketch: based performing approximate newton step using randomly projected sub-sampled hessian. self-concordant functions, prove algorithm super-linear convergence exponentially high probability, convergence complexity guarantees independent condition numbers related problem-dependent quantities. given suitable initialization, similar guarantees also hold strongly convex smooth objectives without self-concordance. implemented using randomized projections based sub-sampled hadamard basis, algorithm typically substantially lower complexity newton's method. also describe extensions methods programs involving convex constraints equipped self-concordant barriers. discuss illustrate applications linear programs, quadratic programs convex constraints, logistic regression generalized linear models, well semidefinite programs.",10
"meta book size-dependent properties written language. evidence given systematic text-length dependence power-law index gamma single book. estimated gamma values consistent monotonic decrease 2 1 increasing length text. direct connection extended heap's law explored. infinite book limit is, consequence, proposed given gamma = 1 instead value gamma=2 expected zipf's law ubiquitously applicable. addition explore idea systematic text-length dependence described meta book concept, abstract representation reflecting word-frequency structure text. according concept word-frequency distribution text, certain length written single author, characteristics text length pulled imaginary complete infinite corpus written author.",14
"non-linear causal inference using gaussianity measures. provide theoretical empirical evidence type asymmetry causes effects present related via linear models contaminated additive non-gaussian noise. assuming causes effects distribution, show distribution residuals linear fit anti-causal direction closer gaussian distribution residuals causal direction. gaussianization effect characterized reduction magnitude high-order cumulants increment differential entropy residuals. problem non-linear causal inference addressed performing embedding expanded feature space, relation causes effects assumed linear. effectiveness method discriminate causes effects based type asymmetry illustrated variety experiments using different measures gaussianity. proposed method shown competitive state-of-the-art techniques causal inference.",18
"comparison several reweighted l1-algorithms solving cardinality minimization problems. reweighted l1-algorithms attracted lot attention field applied mathematics. unified framework algorithms recently proposed zhao li. paper construct new examples reweighted l1-methods. functions certain concave approximations l0-norm function. focus numerical comparison new existing reweighted l1-algorithms. show change parameters reweighted algorithms may affect performance algorithms finding solution cardinality minimization problem. experiments, problem data generated according different statistical distributions, test algorithms different sparsity level solution problem. numerical results demonstrate reweighted l1-method one efficient methods locating solution cardinality minimization problem.",10
"new low-rank tensor model video completion. paper, propose new low-rank tensor model based circulant algebra, namely, twist tensor nuclear norm t-tnn short. twist tensor denotes 3-way tensor representation laterally store 2d data slices order. one hand, t-tnn convexly relaxes tensor multi-rank twist tensor fourier domain, allows efficient computation using fft. other, t-tnn equal nuclear norm block circulant matricization twist tensor original domain, extends traditional matrix nuclear norm block circulant way. test t-tnn model video completion application aims fill missing values experiment results validate effectiveness, especially dealing video recorded non-stationary panning camera. block circulant matricization twist tensor transformed circulant block representation nuclear norm invariance. representation, transformation, exploits horizontal translation relationship frames video, endows t-tnn model powerful ability reconstruct panning videos existing state-of-the-art low-rank models.",3
"accurate combined measurements -- experiment, simulation, theory. paper investigate question much combined measurements increase accuracy additive quantities. therefore, consider set measurements selection possible combinations $n$ labeled masses estimate individual weights $n$ masses linear regression approach. present experimental results motivate comprehensive simulation campaigns. simulations provide valid statistical statements reliable forecasts experimental results. profound analytical treatment turn supports simulation outcomes excellent consistency. one important achievement therein general analytical expression estimate's error, limited two particular weighing schemes presented. turns combined measurements allow estimate weight mass elements accuracy under-runs orders magnitude resolution scale used. error depends amount measurements, one gains higher accuracy increasing effort. broader sense, work wants promote method give inspirations applications various metrological fields beyond high-precision mass determination. moreover, novel simulations analytic formulas enable design optimal experiments.",14
"marrnet: 3d shape reconstruction via 2.5d sketches. 3d object reconstruction single image highly under-determined problem, requiring strong prior knowledge plausible 3d shapes. introduces challenges learning-based approaches, 3d object annotations scarce real images. previous work chose train synthetic data ground truth 3d information, suffered domain adaptation tested real data. work, propose marrnet, end-to-end trainable model sequentially estimates 2.5d sketches 3d object shape. disentangled, two-step formulation three advantages. first, compared full 3d shape, 2.5d sketches much easier recovered 2d image; models recover 2.5d sketches also likely transfer synthetic real data. second, 3d reconstruction 2.5d sketches, systems learn purely synthetic data. easily render realistic 2.5d sketches without modeling object appearance variations real images, including lighting, texture, etc. relieves domain adaptation problem. third, derive differentiable projective functions 3d shape 2.5d sketches; framework therefore end-to-end trainable real images, requiring human annotations. model achieves state-of-the-art performance 3d shape reconstruction.",3
"deterministic annealing approach multiple traveling salesmen related problems. paper presents novel efficient heuristic framework approximating solutions multiple traveling salesmen problem (m-tsp) variants tsp. approach adopted paper extension maximum-entropy-principle (mep) deterministic annealing (da) algorithm. framework presented general tool suitably adapted number variants basic tsp. additionally, unlike heuristics tsp, framework presented paper independent edges defined two pairs nodes. makes algorithm particularly suited variants close-enough traveling salesman problem (cetsp) challenging due added computational complexity. examples presented paper illustrate effectiveness new framework use tsp many variants thereof.",10
"markovian entanglement networks. graphical models probabilistic dependencies extensively investigated context classical uncertainty. however, domains (most notably, computational physics quantum computing) nature relevant uncertainty non-classical, laws classical probability theory superseded quantum mechanics. paper introduce markovian entanglement networks (men), novel class graphical representations quantum-mechanical dependencies context non-classical systems. men quantum-mechanical analogue markovian networks, family undirected graphical representations which, classical domain, exploit notion conditional independence among subsystems. defining notion conditional independence appropriate domain (conditional separability), prove conditional separabilities induced quantum-mechanical wave function effectively reflected graphical structure men. specifically, show wave function exists men perfect map conditional separabilities. next, show graphical structure men used effectively classify pure states three-qubit systems. also demonstrate that, large systems, exploiting conditional independencies may dramatically reduce computational burden various inference tasks. principle, graph-theoretic representation conditional independencies afforded men may facilitate classical simulation quantum systems, also provide guide efficient design complexity analysis quantum algorithms circuits.",17
"machine learning prediction phase-less millimeter-wave beam tracking. future wireless networks may operate millimeter-wave (mmw) sub-terahertz (sub-thz) frequencies enable high data rate requirements. large antenna arrays critical reliable communications mmw sub-thz bands, antenna arrays would also mandate efficient scalable initial beam alignment link maintenance algorithms mobile devices. low-power phased-array architectures phase-less power measurements due high frequency oscillator phase noise pose additional challenges practical beam tracking algorithms. traditional beam tracking protocols require exhaustive sweeps possible beam directions scale poorly high mobility large arrays. compressive sensing machine learning designs proposed improve measurement scaling array size commonly degrade hardware impairments require raw samples respectively. work, introduce novel long short-term memory (lstm) network assisted beam tracking prediction algorithm utilizing phase-less measurements fixed compressive codebooks. demonstrate comparable beam alignment accuracy state-of-the-art phase-less beam alignment algorithms, reducing average number required measurements time.",5
"convergence analysis parallel computing implementation multiagent coordination optimization algorithm. report, novel variation particle swarm optimization (pso) algorithm, called multiagent coordination optimization (mco), implemented parallel computing way practical use introducing matlab built-in function ""parfor"" mco. rigorously analyze global convergence mco means semistability theory. besides sharing global optimal solutions pso algorithm, mco algorithm integrates cooperative swarm behavior multiple agents update formula sharing velocity position information neighbors improve performance. numerical evaluation parallel mco algorithm provided report running proposed algorithm supercomputers high performance computing center texas tech university. particular, optimal value consuming time compared pso serial mco solving several benchmark functions literature, respectively. based simulation results, performance parallel mco superb compared pso solving many nonlinear, noncovex optimization problems, also high efficiency saving computational time.",10
"explaining classification models built high-dimensional sparse data. predictive modeling applications increasingly use data representing people's behavior, opinions, interactions. fine-grained behavior data often different structure traditional data, high-dimensional sparse. models built data quite difficult interpret, since contain many thousands even many millions features. listing features large model coefficients sufficient, model coefficients incorporate information feature presence, key analysing sparse data. paper introduce two alternatives explaining predictive models listing important features. evaluate alternatives terms explanation ""bang buck,"", i.e., many examples' inferences explained given number features listed. bottom line: (i) proposed alternatives double bang-for-the-buck compared listing high-coefficient features, (ii) interestingly, although come different sources motivations, two new alternatives provide strikingly similar rankings important features.",18
"geometric calibration colour stereo surface imaging system esa's trace gas orbiter. many geometric calibration methods ""standard"" cameras. methods, however, cannot used calibration telescopes large focal lengths complex off-axis optics. moreover, specialized calibration methods telescopes scarce literature. describe calibration method developed colour stereo surface imaging system (cassis) telescope, board exomars trace gas orbiter (tgo). although method described context cassis, camera-specific experiments, general applied telescopes. encourage re-use proposed method making calibration code data available on-line.",0
"nips 2016 workshop representation learning artificial biological neural networks (mlini 2016). workshop explores interface cognitive neuroscience recent advances ai fields aim reproduce human performance natural language processing computer vision, specifically deep learning approaches problems. studying cognitive capabilities brain, scientists follow system identification approach present different stimuli subjects try model response different brain areas stimulus. goal understand brain trying find function expresses activity brain areas terms different properties stimulus. experimental stimuli becoming increasingly complex people interested studying real life phenomena perception natural images natural sentences. therefore need rich adequate vector representation properties stimulus, obtain using advances machine learning. parallel, new ml approaches, many deep learning, inspired certain extent human behavior biological principles. neural networks example originally inspired biological neurons. recently, processes attention used inspired human behavior. however, large bulk methods independent findings brain function, unclear whether beneficial machine learning try emulate brain function order achieve tasks brain achieves.",18
"recursive bayesian approach describe retinal vasculature geometry. demographic studies suggest changes retinal vasculature geometry, especially vessel width, associated incidence progression eye-related systemic diseases. date, main information source width estimation fundus images intensity profile vessel edges. however, many factors affecting intensity profile: pathologies, central light reflex local illumination levels, name few. study, introduce three information sources width estimation. probability profiles vessel interior, centreline edge locations generated deep network. probability profiles provide direct access vessel geometry used likelihood calculation bayesian method, particle filtering. also introduce geometric model handle non-ideal conditions probability profiles. experiments conducted review dataset yielded consistent estimates vessel width, even cases one vessel edges difficult identify. moreover, results suggest method better human observers locating edges low contrast vessels.",3
"generative approach zero-shot few-shot action recognition. present generative framework zero-shot action recognition possible action classes occur training data. approach based modeling action class using probability distribution whose parameters functions attribute vector representing action class. particular, assume distribution parameters action class visual space expressed linear combination set basis vectors combination weights given attributes action class. basis vectors learned solely using labeled data known (i.e., previously seen) action classes, used predict parameters probability distributions unseen action classes. consider two settings: (1) inductive setting, use labeled examples seen action classes predict unseen action class parameters; (2) transductive setting leverages unlabeled data unseen action classes. framework also naturally extends few-shot action recognition labeled examples unseen classes available. experiments benchmark datasets (ucf101, hmdb51 olympic) show significant performance improvements compared various baselines, standard zero-shot (disjoint seen unseen classes) generalized zero-shot learning settings.",3
"non-gaussian random generators bacteria foraging algorithm multiobjective optimization. random generators stochastic engines key component structure metaheuristic algorithms. work investigates effects non-gaussian stochastic engines performance metaheuristics solving real-world optimization problem. work, bacteria foraging algorithm (bfa) employed tandem four random generators (stochastic engines). stochastic engines operate using weibull distribution, gamma distribution, gaussian distribution chaotic mechanism. two non-gaussian distributions weibull gamma distributions. work, approaches developed implemented real-world multi-objective resin bonded sand mould problem. pareto frontiers obtained benchmarked using two metrics; hyper volume indicator (hvi) proposed average explorative rate (aer) metric. detail discussions various perspectives effects non-gaussian random generators metaheuristics provided.",3
"playtime measurement survival analysis. maximizing product use central goal many businesses, makes retention monetization two central analytics metrics games. player retention may refer various duration variables quantifying product use: total playtime session playtime popular research targets, active playtime well-suited subscription games. research often goal increasing player retention conversely decreasing player churn. survival analysis framework powerful tools well suited retention type data. paper contributes new methods game analytics playtime analyzed using survival analysis without covariates. survival hazard estimates provide visual analytic interpretation playtime phenomena funnel type nonparametric estimate. metrics based survival curve used aggregate playtime information single statistic. comparison survival curves cohorts provides scientific ab-test. methods work censored data enable computation confidence intervals. especially important time sample limited data occurs game development. throughout paper, illustrate application methods real world game development problems hipster sheep mobile game.",18
"new algorithms learning incoherent overcomplete dictionaries. sparse recovery given matrix $a$ (the dictionary) vector form $a x$ $x$ sparse, goal recover $x$. central notion signal processing, statistics machine learning. applications sparse coding, edge detection, compression super resolution, dictionary $a$ unknown learned random examples form $y = ax$ $x$ drawn appropriate distribution --- dictionary learning problem. settings, $a$ overcomplete: columns rows. paper presents polynomial-time algorithm learning overcomplete dictionaries; previously known algorithm provable guarantees recent work spielman, wang wright gave algorithm full-rank case, rarely case applications. algorithm applies incoherent dictionaries central object study since introduced seminal work donoho huo. particular, dictionary $\mu$-incoherent pair columns inner product $\mu / \sqrt{n}$. algorithm makes natural stochastic assumptions unknown sparse vector $x$, contain $k \leq c \min(\sqrt{n}/\mu \log n, m^{1/2 -\eta})$ non-zero entries (for $\eta > 0$). close best $k$ allowable best sparse recovery algorithms even one knows dictionary $a$ exactly. moreover, running time sample complexity depend $\log 1/\epsilon$, $\epsilon$ target accuracy, algorithms converge quickly true dictionary. algorithm also tolerate substantial amounts noise provided incoherent respect dictionary (e.g., gaussian). noisy setting, running time sample complexity depend polynomially $1/\epsilon$, necessary.",3
"enhancing fractal descriptors images combining boundary interior minkowski dilation. work proposes obtain novel fractal descriptors gray-level texture images combining information interior boundary measures minkowski dilation applied texture surface. first, image converted surface height point gray intensity respective pixel position image. thus, surface morphologically dilated spheres. radius spheres ranged within interval volume external area dilated structure computed radius. final descriptors given measures concatenated subject canonical transform reduce dimensionality. proposal enhancement classical bouligand-minkowski fractal descriptors, volume (interior) information considered. different structures may volume, area, proposal yields rich descriptors confirmed results classification benchmark databases.",14
"streaming architecture large-scale quantized neural networks fpga-based dataflow platform. deep neural networks (dnns) used different applications executed range computer architectures, iot devices supercomputers. footprint networks huge well computational communication needs. order ease pressure resources, research indicates many cases low precision representation (1-2 bit per parameter) weights parameters achieve similar accuracy requiring less resources. using quantized values enables use fpgas run nns, since fpgas well fitted primitives; e.g., fpgas provide efficient support bitwise operations work arbitrary-precision representation numbers. paper presents new streaming architecture running qnns fpgas. proposed architecture scales better alternatives, allowing us take advantage systems multiple fpgas. also included support skip connections, used state-of-the art nns, shown architecture allows add connections almost free. allowed us implement 18-layer resnet 224x224 images classification, achieving 57.5% top-1 accuracy. addition, implemented full-sized quantized alexnet. contrast previous works, use 2-bit activations instead 1-bit ones, improves alexnet's top-1 accuracy 41.8% 51.03% imagenet classification. alexnet resnet handle 1000-class real-time classification fpga. implementation resnet-18 consumes 5x less power 4x slower imagenet, compared nn latest nvidia gpus. smaller nns, fit single fpga, running faster gpus small (32x32) inputs, consuming 20x less energy power.",3
"hilbert space embedding dirichlet process mixtures. paper proposes hilbert space embedding dirichlet process mixture models via stick-breaking construction sethuraman. although bayesian nonparametrics offers powerful approach construct prior avoids need specify model size/complexity explicitly, exact inference often intractable. hand, frequentist approaches kernel machines, suffer model selection/comparison problems, often benefit efficient learning algorithms. paper discusses possibility combine best worlds using dirichlet process mixture model case study.",18
"constructing word similarities meroitic aid decipherment. meroitic still undeciphered language ancient civilization kush. years, various techniques decipherment finding bilingual text cognates modern ancient languages sudan surrounding areas successful. using techniques borrowed information theory natural language statistics, similar words paired attempts made use currently defined words extract least partial meaning unknown words.",3
"learned representation artistic style. diversity painting styles represents rich visual vocabulary construction image. degree one may learn parsimoniously capture visual vocabulary measures understanding higher level features paintings, images general. work investigate construction single, scalable deep network parsimoniously capture artistic style diversity paintings. demonstrate network generalizes across diversity artistic styles reducing painting point embedding space. importantly, model permits user explore new painting styles arbitrarily combining styles learned individual paintings. hope work provides useful step towards building rich models paintings offers window structure learned representation artistic style.",3
"resolvability products. spaces $t_0$ crowded (i.e. isolated points). $n \le \omega$ let $m(n)$ statement $n$ measurable cardinals $\pi(n)$ ($\pi^+(n)$) $n+1$ (0-dimensional $t_2$) spaces whose product irresolvable. prove $m(1),\,\pi(1)$ $\pi^+(1)$ equiconsistent. $1 < n < \omega$ show $con(m(n))$ implies $con(\pi^+(n))$. finally, $con(m(\omega))$ implies consistency infinitely many crowded 0-dimensional $t_2$-spaces product finitely many irresolvable. settle old problems malychin. concerning even older question ceder pearson, show following consistent modulo measurable cardinal: (i) 0-dimensional $t_2$ space $x$ $\omega_2 \le \delta(x) \le 2^{\omega_1}$ whose product countable space $\omega_2$-resolvable, hence maximally resolvable. (ii) monotonically normal space $x$ $\delta(x) = \aleph_\omega$ whose product countable space $\omega_1$-resolvable, hence maximally resolvable. significantly improve result eckertson.",10
"development evaluation deep learning model protein-ligand binding affinity prediction. structure based ligand discovery one successful approaches augmenting drug discovery process. currently, notable shift towards machine learning (ml) methodologies aid procedures. deep learning recently gained considerable attention allows model ""learn"" extract features relevant task hand. developed novel deep neural network estimating binding affinity ligand-receptor complexes. complex represented 3d grid, model utilizes 3d convolution produce feature map representation, treating atoms proteins ligands manner. network tested casf ""scoring power"" benchmark astex diverse set outperformed classical scoring functions. model, together usage instructions examples, available git repository http://gitlab.com/cheminfibb/pafnucy",18
"conformalized kernel ridge regression. general predictive models provide measure confidence predictions without bayesian assumptions. way circumvent potential restrictions use conformal methods constructing non-parametric confidence regions, offer guarantees regarding validity. paper provide detailed description computationally efficient conformal procedure kernel ridge regression (krr), conduct comparative numerical study see well conformal regions perform bayesian confidence sets. results suggest conformalized krr yield predictive confidence regions specified coverage rate, essential constructing anomaly detection systems based predictive models.",18
"handwritten digit recognition committee deep neural nets gpus. competitive mnist handwritten digit recognition benchmark long history broken records since 1998. recent substantial improvement others dates back 7 years (error rate 0.4%) . recently able significantly improve result, using graphics cards greatly speed training simple deep mlps, achieved 0.35%, outperforming previous complex methods. report another substantial improvement: 0.31% obtained using committee mlps.",3
"accelerated variance reduced stochastic admm. recently, many variance reduced stochastic alternating direction method multipliers (admm) methods (e.g.\ sag-admm, sdca-admm svrg-admm) made exciting progress linear convergence rates strongly convex problems. however, best known convergence rate general convex problems o(1/t) opposed o(1/t^2) accelerated batch algorithms, $t$ number iterations. thus, still remains gap convergence rates existing stochastic admm batch algorithms. bridge gap, introduce momentum acceleration trick batch optimization stochastic variance reduced gradient based admm (svrg-admm), leads accelerated (asvrg-admm) method. design two different momentum term update rules strongly convex general convex cases. prove asvrg-admm converges linearly strongly convex problems. besides low per-iteration complexity existing stochastic admm methods, asvrg-admm improves convergence rate general convex problems o(1/t) o(1/t^2). experimental results show effectiveness asvrg-admm.",3
"supersymmetric massive gravity. initiate systematic study self-interactions massive spin-2 ""graviton"" consistent $\mathcal{n}=4$ supersymmetry. using recently developed massive on-shell superspace formalism, construct general set cubic massive graviton amplitudes form supersymmetry lorentz invariance manifest. find $\mathcal{n}\geq 3$ supersymmetry, family consistent interactions coincide ghost-free drgt model. $\mathcal{n}=4$ (maximal) supersymmetry single consistent cubic interaction coincides unique structure required absence asymptotic superluminality. additionally, discuss structure interactions high-energy limit, connections supersymmetric galileons possibility supersymmetric massive double copy.",9
"graph-sparse lda: topic model structured sparsity. originally designed model text, topic modeling become powerful tool uncovering latent structure domains including medicine, finance, vision. goals model vary depending application: cases, discovered topics may used prediction downstream task. cases, content topic may intrinsic scientific interest. unfortunately, even using modern sparse techniques, discovered topics often difficult interpret due high dimensionality underlying space. improve topic interpretability, introduce graph-sparse lda, hierarchical topic model leverages knowledge relationships words (e.g., encoded ontology). model, topics summarized latent concept-words underlying graph explain observed words. graph-sparse lda recovers sparse, interpretable summaries two real-world biomedical datasets matching state-of-the-art prediction performance.",18
"learning vector autoregressive models latent processes. study problem learning support transition matrix random processes vector autoregressive (var) model samples subset processes latent. well known ignoring effect latent processes may lead different estimates influences among observed processes, concerned identifying influences among observed processes, latent ones, latent observed ones. show support transition matrix among observed processes lengths latent paths two observed processes identified successfully conditions var model. lengths latent paths, reconstruct latent subgraph (representing influences among latent processes) minimum number variables uniquely topology directed tree. furthermore, propose algorithm finds possible minimal latent graphs conditions lengths latent paths. results apply non-gaussian gaussian cases, experimental results various synthetic real-world datasets validate theoretical results.",3
"formal calculus international relations computation evaluation. publication presents relation computation calculus international relations using mathematical modeling. examined trust international relations calculus, related bayesian inference, dempster-shafer theory subjective logic. based observation literature, found literature discussing calculus method international relations. bridge research gap, propose relation algebra method international relations computation. proposed method allow relation computation previously subjective incomputable. also present three international relations case studies demonstrate proposed method real-world scenario. method deliver relation computation international relations support decision makers government foreign ministry, defense ministry, presidential prime minister office. department defense (dod) may use method determine nation identified friendly, neutral hostile nation.",3
"ranking biomarkers mutual information. study information theoretic methods ranking biomarkers. clinical trials two, closely related, types biomarkers: predictive prognostic, disentangling key challenge. first step phrase biomarker ranking terms optimizing information theoretic quantity. formalization problem enable us derive rankings predictive/prognostic biomarkers, estimating different, high dimensional, conditional mutual information terms. estimate terms, suggest efficient low dimensional approximations, derive empirical bayes estimator, suitable small sparse datasets. finally, introduce new visualisation tool captures prognostic predictive strength set biomarkers. believe representation prove powerful tool biomarker discovery.",18
"towards automated patient data cleaning using deep learning: feasibility study standardization organ labeling. data cleaning consumes 80% time spent data analysis clinical research projects. much bigger problem era big data machine learning field medicine large volumes data generated. report initial effort towards automated patient data cleaning using deep learning: standardization organ labeling radiation therapy. organs often labeled inconsistently different institutions (sometimes even within institution) different time periods, poses problem clinical research, especially multi-institutional collaborative clinical research acquired patient data used effectively. developed convolutional neural network (cnn) automatically identify organ ct image label standardized nomenclature presented aapm task group 263. tested model ct images 54 patients prostate 100 patients head neck cancer previously received radiation therapy. model achieved 100% accuracy detecting organs assigning standardized labels patients tested. work shows feasibility using deep learning patient data cleaning enables standardized datasets generated effective intra- interinstitutional collaborative clinical research.",14
"learning wasserstein embeddings. wasserstein distance received lot attention recently community machine learning, especially principled way comparing distributions. found numerous applications several hard problems, domain adaptation, dimensionality reduction generative models. however, use still limited heavy computational cost. goal alleviate problem providing approximation mechanism allows break inherent complexity. relies search embedding euclidean distance mimics wasserstein distance. show embedding found siamese architecture associated decoder network allows move embedding space back original input space. embedding found, computing optimization problems wasserstein space (e.g. barycenters, principal directions even archetypes) conducted extremely fast. numerical experiments supporting idea conducted image datasets, show wide potential benefits method.",18
"assessing tuning brain decoders: cross-validation, caveats, guidelines. decoding, ie prediction brain images signals, calls empirical evaluation predictive power. evaluation achieved via cross-validation, method also used tune decoders' hyper-parameters. paper review cross-validation procedures decoding neuroimaging. includes didactic overview relevant theoretical considerations. practical aspects highlighted extensive empirical study common decoders within-and across-subject predictions, multiple datasets --anatomical functional mri meg-- simulations. theory experiments outline popular "" leave-one-out "" strategy leads unstable biased estimates, repeated random splits method preferred. experiments outline large error bars cross-validation neuroimaging settings: typical confidence intervals 10%. nested cross-validation tune decoders' parameters avoiding circularity bias. however find favorable use sane defaults, particular non-sparse decoders.",18
"semi-optimal edge detector based simple standard deviation adjusted thresholding. paper proposes novel method combines median filter simple standard deviation accomplish excellent edge detector image processing. first all, denoising process must applied grey scale image using median filter identify pixels likely contaminated noise. benefit step smooth image get rid noisy pixels. that, simple statistical standard deviation could computed 2x2 window size. value standard deviation inside 2x2 window size greater predefined threshold, upper left pixel 2?2 window represents edge. visual differences proposed edge detector standard known edge detectors shown support contribution paper.",3
"modeling multi-stage decision optimization problems. multi-stage optimization uncertainty techniques used solve long-term management problems. although many optimization modeling language extensions well computational environments proposed, acceptance technique generally low, due inherent complexity modeling solution process. paper simplification annotate multi-stage decision problems uncertainty presented - simplification contrasts common approach create extension top existing optimization modeling language. leads definition meta models, instanced various programming languages. example using statistical computing language r shown.",10
"human shape variation - efficient implementation using skeleton. times important detect human presence automatically secure environments. needs shape recognition algorithm robust, fast low error rates. algorithm needs process camera images quickly detect human range vision, generate alerts, especially object scrutiny moving certain directions. present simple, efficient fast algorithm using skeletons images, simple features like posture length object.",3
"pricing algorithms. paper studies markov perfect equilibria repeated duopoly model sellers choose algorithms. algorithm mapping competitor's price price. set, algorithms respond quickly. customers arrive randomly opportunities revise algorithm. simple game two possible prices, monopoly outcome unique equilibrium standard functional forms profit function. generally, multiple prices, exercise market power rule -- equilibria, expected payoff sellers competitive outcome, least one seller close monopoly outcome. sustenance collusion seems outside scope standard antitrust laws involve direct communication.",4
label-dependent feature extraction social networks node classification. new method feature extraction social network within-network classification proposed paper. method provides new features calculated combination both: network structure information class labels assigned nodes. influence various features classification performance also studied. experiments real-world data shown features created owing proposed method lead significant improvement classification accuracy.,3
"adaptive system optimization using random directions stochastic approximation. present novel algorithms simulation optimization using random directions stochastic approximation (rdsa). include first-order (gradient) well second-order (newton) schemes. incorporate continuous-valued well discrete-valued perturbations algorithms. former chosen independent identically distributed (i.i.d.) symmetric, uniformly distributed random variables (r.v.), latter i.i.d., asymmetric, bernoulli r.v.s. newton algorithm, novel hessian estimation scheme, requires n-dimensional perturbations three loss measurements per iteration, whereas simultaneous perturbation newton search algorithm [1] requires 2n-dimensional perturbations four loss measurements per iteration. prove unbiasedness gradient hessian estimates asymptotic (strong) convergence first-order second-order schemes. also provide asymptotic normality results, particular establish asymmetric bernoulli variant newton rdsa method better 2spsa [1]. numerical experiments used validate theoretical results.",10
"robust anomaly detection using semidefinite programming. paper presents new approach, based polynomial optimization method moments, problem anomaly detection. proposed technique requires information statistical moments normal-state distribution features interest compares favorably existing approaches (such parzen windows 1-class svm). addition, provides succinct description normal state. thus, leads substantial simplification anomaly detection problem working higher dimensional datasets.",10
"probit normal correlated topic models. logistic normal distribution recently adapted via transformation multivariate gaus- sian variables model topical distribution documents presence correlations among topics. paper, propose probit normal alternative approach modelling correlated topical structures. use probit model context topic discovery novel, many authors far con- centrated solely logistic model partly due formidable inefficiency multinomial probit model even case small topical spaces. herein circumvent inefficiency multinomial probit estimation using adaptation diagonal orthant multinomial probit topic models context, resulting ability topic modelling scheme handle corpuses large number latent topics. additional important benefit method lies fact unlike logistic normal model whose non-conjugacy leads need sophisticated sampling schemes, ap- proach exploits natural conjugacy inherent auxiliary formulation probit model achieve greater simplicity. application proposed scheme well known associated press corpus helps discover large number meaningful topics also reveals capturing compellingly intuitive correlations among certain topics. besides, proposed approach lends even scalability thanks various existing high performance algorithms architectures capable handling millions documents.",18
"fast inference sparse coding algorithms applications object recognition. adaptive sparse coding methods learn possibly overcomplete set basis functions, natural image patches reconstructed linearly combining small subset bases. applicability methods visual object recognition tasks limited prohibitive cost optimization algorithms required compute sparse representation. work propose simple efficient algorithm learn basis functions. training, model also provides fast smooth approximator optimal representation, achieving even better accuracy exact sparse coding algorithms visual object recognition tasks.",3
"person re-identification unsupervised video matching. existing person re-identification (reid) methods rely spatial appearance information either one multiple person images, whilst ignore space-time cues readily available video image-sequence data. moreover, often assume availability exhaustively labelled cross-view pairwise data every camera pair, making non-scalable reid applications real-world large scale camera networks. work, introduce novel video based person reid method capable accurately matching people across views arbitrary unaligned image-sequences without labelled pairwise data. specifically, introduce new space-time person representation encoding multiple granularities spatio-temporal dynamics form time series. moreover, time shift dynamic time warping (ts-dtw) model derived performing automatically alignment whilst achieving data selection matching inherently inaccurate incomplete sequences unified way. extend ts-dtw model accommodating multiple feature-sequences image-sequence order fuse information different descriptions. crucially, model require pairwise labelled training data (i.e. unsupervised) therefore readily scalable large scale camera networks arbitrary camera pairs without need exhaustive data annotation every camera pair. show effectiveness advantages proposed method extensive comparisons related state-of-the-art approaches using two benchmarking reid datasets, prid2011 ilids-vid.",3
"towards completely lifted search-based probabilistic inference. promise lifted probabilistic inference carry probabilistic inference relational probabilistic model without needing reason individual separately (grounding representation) treating undistinguished individuals block. current exact methods still need ground cases, typically representation intermediate results closed lifted operations. set answer question whether fundamental reason lifted algorithms would need ground undifferentiated individuals. two main results: (1) completely characterize cases grounding polynomial population size, show lifted inference time polynomial logarithm population size cases. (2) case no-argument single-argument parametrized random variables grounding polynomial population size, present lifted inference polynomial population size whereas grounding exponential. neither cases requires reasoning separately individuals explicitly mentioned.",3
"mean field type equations line bundle closed riemann surface. let $(\mathcal{l},\mathfrak{g})$ line bundle closed riemann surface $(\sigma,g)$, $\gamma(\mathcal{l})$ set smooth sections, $\mathcal{d}:\gamma(\mathcal{l})\rightarrow t^\ast\sigma\otimes \gamma(\mathcal{l})$ connection independent bundle metric $\mathfrak{g}$, $t^\ast\sigma$ cotangent bundle. suppose exists global unit frame $\zeta$ $\gamma({\mathcal{l}})$. precisely $\sigma\in\gamma(\mathcal{l})$, exists unique smooth function $u:\sigma\rightarrow\mathbb{r}$ $\sigma=u\zeta$ $|\zeta|\equiv 1$ $\sigma$. real number $\rho$, define functional $\mathcal{j}_\rho:w^{1,2}(\sigma,\mathcal{l})\rightarrow\mathbb{r}$ $$\mathcal{j}_\rho(\sigma)=\frac{1}{2}\int_\sigma|\mathcal{d} \sigma|^2dv_g+\frac{\rho} {|\sigma|}\int_\sigma\langle\sigma,\zeta\rangle dv_g-\rho\log\int_\sigma h e^{\langle\sigma,\zeta\rangle}dv_g,$$ $w^{1,2}(\sigma,\mathcal{l})$ completion $\gamma(\mathcal{l})$ usual sobolev norm, $|\sigma|$ area $(\sigma,g)$, $h:\sigma\rightarrow\mathbb{r}$ strictly positive smooth function $\langle\cdot,\cdot\rangle$ inner product induced $\mathfrak{g}$. euler-lagrange equations $\mathcal{j}_\rho$ called mean field type equations. write $\mathcal{h}_0=\{\sigma\in w^{1,2}(\sigma,\mathcal{l}):\mathcal{d}\sigma=0\}$ $$\mathcal{h}_1=\left\{\sigma\in w^{1,2}(\sigma,\mathcal{l}):\int_\sigma \langle\sigma,\tau\rangle dv_g=0,\,\,\forall \tau \in \mathcal{h}_0\right\}.$$ based variational method, prove $\mathcal{j}_\rho$ constraint critical point space $\mathcal{h}_1$ $\rho<8\pi$; based blow-up analysis, calculate exact value $\inf_{\sigma\in\mathcal{h}_1}\mathcal{j}_{8\pi}(\sigma)$, provided achieved $\sigma\in\mathcal{h}_1$;",10
"detecting text natural image connectionist text proposal network. propose novel connectionist text proposal network (ctpn) accurately localizes text lines natural image. ctpn detects text line sequence fine-scale text proposals directly convolutional feature maps. develop vertical anchor mechanism jointly predicts location text/non-text score fixed-width proposal, considerably improving localization accuracy. sequential proposals naturally connected recurrent neural network, seamlessly incorporated convolutional network, resulting end-to-end trainable model. allows ctpn explore rich context information image, making powerful detect extremely ambiguous text. ctpn works reliably multi-scale multi- language text without post-processing, departing previous bottom-up methods requiring multi-step post-processing. achieves 0.88 0.61 f-measure icdar 2013 2015 benchmarks, surpass- ing recent results [8, 35] large margin. ctpn computationally efficient 0:14s/image, using deep vgg16 model [27]. online demo available at: http://textdet.com/.",3
"improved bounds quantum learning algorithms. article give several new results complexity algorithms learn boolean functions quantum queries quantum examples. hunziker et al. conjectured class c boolean functions, number quantum black-box queries required exactly identify unknown function c $o(\frac{\log |c|}{\sqrt{{\hat{\gamma}}^{c}}})$, $\hat{\gamma}^{c}$ combinatorial parameter class c. essentially resolve conjecture affirmative giving quantum algorithm that, class c, identifies unknown function c using $o(\frac{\log |c| \log \log |c|}{\sqrt{{\hat{\gamma}}^{c}}})$ quantum black-box queries. consider range natural problems intermediate exact learning problem (in learner must obtain bits information black-box function) usual problem computing predicate (in learner must obtain one bit information black-box function). give positive negative results quantum classical query complexities intermediate problems polynomially related other. finally, improve known lower bounds number quantum examples (as opposed quantum black-box queries) required $(\epsilon,\delta)$-pac learning concept class vapnik-chervonenkis dimension domain $\{0,1\}^n$ $\omega(\frac{d}{n})$ $\omega(\frac{1}{\epsilon}\log \frac{1}{\delta}+d+\frac{\sqrt{d}}{\epsilon})$. new lower bound comes closer matching known upper bounds classical pac learning.",17
"fairness, diversity randomness algorithmic decision making. consider binary decision making process single machine learning classifier replaces multitude humans. raise questions resulting loss diversity decision making process. study potential benefits using random classifier ensembles instead single classifier context fairness-aware learning demonstrate various attractive properties: (i) ensemble fair classifiers guaranteed fair, several different measures fairness, (ii) ensemble unfair classifiers still achieve fair outcomes, (iii) ensemble classifiers achieve better accuracy-fairness trade-offs single classifier. finally, introduce notions distributional fairness characterize potential benefits random classifier ensembles.",18
"evolutionary forces language change. languages genes transmitted generation generation, opportunity differential reproduction survivorship forms. apply rigorous inference framework, drawn population genetics, distinguish two broad mechanisms language change: drift selection. drift change results stochasticity transmission may occur absence intrinsic difference linguistic forms; whereas selection truly evolutionary force arising intrinsic differences -- example, one form preferred members population. using large corpora parsed texts spanning 12th century 21st century, analyze three examples grammatical changes english: regularization past-tense verbs, rise periphrastic `do', syntactic variation verbal negation. show reject stochastic drift favor selective force driving language changes, others. strength drift depends word's frequency, drift provides alternative explanation words prone change others. results suggest important role stochasticity language change, provide null model selective theories language evolution must compared.",15
"towards building knowledge base monetary transactions news collection. address problem extracting structured representations economic events large corpus news articles, using combination natural language processing machine learning techniques. developed techniques allow semi-automatic population financial knowledge base, which, turn, may used support range data mining exploration tasks. key challenge face domain event often reported multiple times, varying correctness details. address challenge first collecting information pertinent given event entire corpus, considering possible representations event, finally, using supervised learning method, rank representations associated confidence scores. main innovative element approach jointly extracts stores attributes event single representation (quintuple). using purpose-built test set demonstrate supervised learning approach achieve 25% improvement f1-score baseline methods consider earliest, latest frequent reporting event.",3
"best-first heuristic search multicore machines. harness modern multicore processors, imperative develop parallel versions fundamental algorithms. paper, compare different approaches parallel best-first search shared-memory setting. present new method, pbnf, uses abstraction partition state space detect duplicate states without requiring frequent locking. pbnf allows speculative expansions necessary keep threads busy. identify fix potential livelock conditions approach, proving correctness using temporal logic. approach general, allowing extend easily suboptimal anytime heuristic search. empirical comparison strips planning, grid pathfinding, sliding tile puzzle problems using 8-core machines, show a*, weighted a* anytime weighted a* implemented using pbnf yield faster search improved versions previous parallel search proposals.",3
"group importance sampling particle filtering mcmc. importance sampling (is) well-known monte carlo technique approximates integrals involving posterior distribution means weighted samples. work, study assignation single weighted sample compresses information contained population weighted samples. part theory present group importance sampling (gis) employed implicitly different works literature. provided analysis yields several theoretical practical consequences. instance, discuss application gis sequential importance resampling framework show independent multiple try metropolis schemes interpreted standard metropolis-hastings algorithm, following gis approach. also introduce two novel markov chain monte carlo (mcmc) techniques based gis. first one, named group metropolis sampling method, produces markov chain sets weighted samples. sets employed obtaining unique global estimator. second one distributed particle metropolis-hastings technique, different parallel particle filters jointly used drive mcmc algorithm. different resampled trajectories compared tested proper acceptance probability. novel schemes tested different numerical experiments learning hyperparameters gaussian processes, localization problem wireless sensor network tracking vegetation parameters given satellite observations, compared several benchmark monte carlo techniques. three illustrative matlab demos also provided.",18
"retinet: automatic amd identification oct volumetric data. optical coherence tomography (oct) provides unique ability image eye retina 3d micrometer resolution gives ophthalmologist ability visualize retinal diseases age-related macular degeneration (amd). visual inspection oct volumes remains main method amd identification, time consuming cross-section within volume must inspected individually clinician. much way, acquiring ground truth information cross-section expensive time consuming. fact heavily limits ability acquire large amounts ground truth, subsequently impacts performance learning-based methods geared automatic pathology identification. avoid burden, propose novel strategy automatic analysis oct volumes volume labels needed. is, train classifier semi-supervised manner conduct task. approach uses novel convolutional neural network (cnn) architecture, needs volume-level labels trained automatically asses whether oct volume healthy contains amd. architecture involves first learning cross-section pathology classifier using pseudo-labels could corrupted leverage towards accurate volume-level classification. show approach provides excellent performances publicly available dataset outperforms number existing automatic techniques.",3
"homological approach gaussian unitary ensemble. study gaussian unitary ensemble (gue) using noncommutative geometry homological framework batalin-vilkovisky (bv) formalism. coefficients correlation functions gue respect rank $n$ described terms ribbon graph feynman diagrams lead counting problem corresponding surfaces. canonical relations provided homological setup determine recurrence relation correlation functions. using recurrence relation properties catalan numbers, determine leading order behavior correlation functions respect rank $n$. application, prove generalization wigner's semicircle law compute large $n$ statistical correlations family random variables gue defined multi-trace functions.",11
"differential angular imaging material recognition. material recognition real-world outdoor surfaces become increasingly important computer vision support operation ""in wild."" computational surface modeling underlies material recognition transitioned reflectance modeling using in-lab controlled radiometric measurements image-based representations based internet-mined images materials captured scene. propose take middle-ground approach material recognition takes advantage rich radiometric cues flexible image capture. realize developing framework differential angular imaging, small angular variations image capture provide enhanced appearance representation significant recognition improvement. build large-scale material database, ground terrain outdoor scenes (gtos) database, geared towards real use autonomous agents. database consists 30,000 images covering 40 classes outdoor ground terrain varying weather lighting conditions. develop novel approach material recognition called differential angular imaging network (dain) fully leverage large dataset. novel network architecture, extract characteristics materials encoded angular spatial gradients appearance. results show dain achieves recognition performance surpasses single view coarsely quantized multiview images. results demonstrate effectiveness differential angular imaging means flexible, in-place material recognition.",3
"analytical shape determination fiber-like objects virtual image correlation. paper reports method allowing determination shape deformed fiber-like objects. compared existing methods, provides analytical results including local slope curvature first importance, instance, beam mechanics. presented vic (virtual image correlation) method consists looking best correlation image fiber-like object virtual beam image, using algorithm close digital image correlation method developed experimental solid mechanics. computation involves part image vicinity fiber: method thus insensitive picture background computational cost remains low. two examples reported: first proves precision method, second ability identify complex shape multiple loops.",3
"abruptly-changing slowly-varying multiarmed bandit problems. study non-stationary stochastic multiarmed bandit (mab) problem propose two generic algorithms, namely, limited memory deterministic sequencing exploration exploitation (lm-dsee) sliding-window upper confidence bound# (sw-ucb#). rigorously analyze algorithms abruptly-changing slowly-varying environments characterize performance. show expected cumulative regret algorithms either environments upper bounded sublinear functions time, i.e., time average regret asymptotically converges zero. complement analytic results numerical illustrations.",18
"initial conditions universe constrained simulations. present new approach recover primordial density fluctuations cosmic web structure underlying galaxy distribution. method based sampling gaussian fields compatible galaxy distribution structure formation model. achieved splitting inversion problem two gibbs-sampling steps: first gaussianisation step transforming distribution point sources lagrangian positions -which priori given- linear alias-free gaussian field. step based hamiltonian sampling gaussian-poisson model. second step consists likelihood comparison set matter tracers initial conditions constrained galaxy distribution assumed structure formation model. computational reasons second order lagrangian perturbation theory used. however, presented approach flexible adopt structure formation model. semi-analytic halo-model based galaxy mock catalog taken demonstrate recovered initial conditions closely unbiased respect actual ones corresponding n-body simulation scales ~ 5 mpc/h. cross-correlation shows substantial gain information, k ~ 0.3 h/mpc doubled. addition initial conditions extremely well gaussian distributed power-spectra follow shape linear power-spectrum close actual one simulation scales k ~ 1 h/mpc.",0
"extremal optimization sensor report pre-processing. describe recently introduced extremal optimization algorithm apply target detection association problems arising pre-processing multi-target tracking. consider problem pre-processing multiple target tracking number sensor reports received large arrives large bursts. case, sometimes necessary pre-process reports sending tracking modules fusion system. pre-processing step associates reports known tracks (or initializes new tracks reports objects seen before). could also used pre-process step clustering, e.g., order test many clusters use. pre-processing done solving approximate version original problem. approximation, pair-wise conflicts calculated. approximation relies knowing many pair-wise conflicts necessary compute. determine this, results phase-transitions occurring coloring (or clustering) large random instances particular graph ensemble used.",3
"tightness branching random walk time-inhomogeneous random environment. consider branching random walk time-inhomogeneous random environment, particles generation $k$ branch random number particles $\mathcal{l}_{k+1}\ge 2$, $\mathcal{l}_k$, $k\in\mathbb{n}$, i.i.d., increments standard normal. let $\mathbb{p}$ denote law $(\mathcal{l}_k)_{k\in\mathbb{n}}$, let $m_n$ denote position maximal particle generation $n$. prove $m_n$, functions $(\mathcal{l}_k)_{k\in\{0,\dots, n\}}$, (with regard $\mathbb{p}$) sequence $(m_n-m_n)_{n\in\mathbb{n}}$ tight high probability.",10
"answering fuzzy conjunctive queries finitely valued fuzzy ontologies. fuzzy description logics (dls) provide means representing vague knowledge application domain. paper, study fuzzy extensions conjunctive queries (cqs) dl $\mathcal{sroiq}$ based finite chains degrees truth. answer queries, extend well-known technique reduces fuzzy ontology classical one, use classical dl reasoners black box. improve complexity previous reduction techniques finitely valued fuzzy dls, allows us prove tight complexity results answering certain kinds fuzzy cqs. conclude experimental evaluation prototype implementation, showing feasibility approach.",3
"joint estimation multiple dependent gaussian graphical models applications mouse genomics. gaussian graphical models widely used represent conditional dependence among random variables. paper, propose novel estimator data arising group gaussian graphical models dependent. motivating example modeling gene expression collected multiple tissues individual: multivariate outcome affected dependencies acting level specific tissues, also level whole body; existing methods assume independence among graphs applicable case. estimate multiple dependent graphs, decompose problem two graphical layers: systemic layer, affects outcomes thereby induces cross- graph dependence, category-specific layer, represents graph-specific variation. propose graphical em technique estimates layers jointly, establish estimation consistency selection sparsistency proposed estimator, confirm simulation em method superior simple one-step method. apply technique mouse genomics data obtain biologically plausible results.",18
"occlusion-aware hand pose estimation using hierarchical mixture density network. hand pose estimation predict pose parameters representing 3d hand model, locations hand joints. problem challenging due large changes viewpoints articulations, intense self-occlusions, etc. many researchers investigated problem aspects input feature learning output prediction modelling. though effective, existing discriminative methods give deterministic estimation target poses. also, due single-value mapping intrinsic, fail adequately handle self-occlusion problems, occluded joints present multiple modes. paper, tackle self-occlusion issue provide complete description observed poses given input depth image hierarchical mixture density network (hmdn) framework. particular, hmdn leverages state-of-the-art cnn module facilitate feature learning, proposes density two-level hierarchy reconcile single-valued multi-valued mapping output. whole framework naturally end-to-end trainable mixture two differentiable density functions. hmdn produces interpretable diverse candidate samples, significantly outperforms state-of-the-art algorithms benchmarks exhibit occlusions.",3
"conservative, proportional optimistic contextual discounting belief functions theory. information discounting plays important role theory belief functions and, generally, information fusion. nevertheless, neither classical uniform discounting contextual cannot model certain use cases, notably temporal discounting. article, new contextual discounting schemes, conservative, proportional optimistic, proposed. properties discounting operations examined. classical discounting shown special case schemes. two motivating cases discussed: modelling source reliability application temporal discounting.",3
"latent semantic word sense disambiguation using global co-occurrence information. paper, propose novel word sense disambiguation method based global co-occurrence information using nmf. calculate dependency relation matrix, existing method tends produce sparse co-occurrence matrix small training set. therefore, nmf algorithm sometimes converge desired solutions. obtain large number co-occurrence relations, propose use co-occurrence frequencies dependency relations word features whole training set. enables us solve data sparseness problem induce effective latent features. evaluate efficiency method word sense disambiguation, make experiments compare result two baseline methods. results experiments show method effective word sense disambiguation comparison baseline methods. moreover, proposed method effective obtaining stable effect analyzing global co-occurrence information.",3
"learning densities conditional many interacting features. learning distribution conditional set discrete-valued features commonly encountered task. becomes challenging high-dimensional feature set possibility interaction features. addition, many frequently applied techniques consider prediction mean, complete conditional density needed answer complex questions. demonstrate novel nonparametric bayes method based upon tensor factorization feature-dependent weights gaussian kernels. method makes use multistage feature selection dimension reduction. resulting conditional density morphs flexibly selected features.",18
"domain adaptation graphs learning aligned graph bases. propose method domain adaptation graphs. given sufficiently many observations label function source graph, study problem transferring label information source graph target graph estimating target label function. assumption relation two domains frequency content label function, regarded graph signal, similar characteristics source target graphs. propose method learn pair coherent bases two graphs, corresponding source target graph basis vectors similar spectral content, ""aligning"" two graphs time reconstructed source target label functions similar coefficients bases. experiments several types data sets suggest proposed method compares quite favorably reference domain adaptation methods. best knowledge, treatment first study domain adaptation problem purely graph-based setting need embedding data ambient space. feature particularly convenient many problems interest concerning learning graphs networks.",18
"weight uncertainty neural networks. introduce new, efficient, principled backpropagation-compatible algorithm learning probability distribution weights neural network, called bayes backprop. regularises weights minimising compression cost, known variational free energy expected lower bound marginal likelihood. show principled kind regularisation yields comparable performance dropout mnist classification. demonstrate learnt uncertainty weights used improve generalisation non-linear regression problems, weight uncertainty used drive exploration-exploitation trade-off reinforcement learning.",18
"classification ecg based hybrid features using cnns wearable applications. sudden cardiac death arrhythmia account large percentage deaths worldwide. electrocardiography (ecg) widely used screening tool cardiovascular diseases. traditionally, ecg signals classified manually, requiring experience great skill, time-consuming prone error. thus machine learning algorithms widely adopted ability perform complex data analysis. features derived points interest ecg - mainly q, r, s, widely used arrhythmia detection. work, demonstrate improved performance ecg classification using hybrid features three different models, building 1-d convolutional neural network (cnn) model proposed past. rr interval features based model proposed work achieved accuracy 98.98%, improvement baseline model. make model immune noise, updated model using frequency features achieved good sustained performance presence noise slightly lower accuracy 98.69%. further, another model combining frequency features rr interval features developed, achieved high accuracy 99% good sustained performance noisy environments. due high accuracy noise immunity, proposed model combines multiple hybrid features, well suited ambulatory wearable sensing applications.",5
"improved strongly adaptive online learning using coin betting. paper describes new parameter-free online learning algorithm changing environments. comparing algorithms time complexity ours, obtain strongly adaptive regret bound factor least $\sqrt{\log(t)}$ better, $t$ time horizon. empirical results show algorithm outperforms state-of-the-art methods learning expert advice metric learning scenarios.",18
"model-lite case-based planning. increasing awareness planning community depending complete models impedes applicability planning technology many real world domains burden specifying complete domain models high. paper, consider novel solution challenge combines generative planning incomplete domain models library plan cases known correct. arguably original motivation case-based planning, existing case-based planners assume (and depend on) from-scratch planners work complete domain models. contrast, approach views plan generated respect incomplete model ""skeletal plan"" augments directed mining plan fragments library cases. present details approach present empirical evaluation method comparison state-of-the-art case-based planner depends complete domain models.",3
"parallel chromatic mcmc spatial partitioning. introduce novel approach parallelizing mcmc inference models spatially determined conditional independence relationships, existing techniques exploiting graphical model structure applicable. approach motivated model seismic events signals, events detected distant regions approximately independent given intermediate regions. perform parallel inference coloring factor graph defined regions latent space, rather individual model variables. evaluating model seismic event detection, achieve significant speedups serial mcmc degradation inference quality.",18
"data recombination neural semantic parsing. modeling crisp logical regularities crucial semantic parsing, making difficult neural models task-specific prior knowledge achieve good results. paper, introduce data recombination, novel framework injecting prior knowledge model. training data, induce high-precision synchronous context-free grammar, captures important conditional independence properties commonly found semantic parsing. train sequence-to-sequence recurrent network (rnn) model novel attention-based copying mechanism datapoints sampled grammar, thereby teaching model structural properties. data recombination improves accuracy rnn model three semantic parsing datasets, leading new state-of-the-art performance standard geoquery dataset models comparable supervision.",3
"mean-field theory meta-learning. discuss mean-field theory cellular automata model meta-learning. meta-learning process combining outcomes individual learning procedures order determine final decision higher accuracy single learning method. method constructed ensemble interacting, learning agents, acquire process incoming information using various types, different versions machine learning algorithms. abstract learning space, agents located, constructed using fully connected model couples agents random strength values. cellular automata network simulates higher level integration information acquired independent learning trials. final classification incoming input data therefore defined stationary state meta-learning system using simple majority rule, yet minority clusters share opposite classification outcome observed system. therefore, probability selecting proper class given input data, estimated even without prior knowledge affiliation. fuzzy logic easily introduced system, even learning agents build simple binary classification machine learning algorithms calculating percentage agreeing agents.",18
"communication efficient distributed optimization using approximate newton-type method. present novel newton-type method distributed optimization, particularly well suited stochastic optimization learning problems. quadratic objectives, method enjoys linear rate convergence provably \emph{improves} data size, requiring essentially constant number iterations reasonable assumptions. provide theoretical empirical evidence advantages method compared approaches, one-shot parameter averaging admm.",3
"weighted tensor decomposition learning latent variables partial data. tensor decomposition methods popular tools learning latent variables given lower-order moments data. however, standard assumption sufficient data estimate moments high accuracy. work, consider case certain dimensions data always observed---common applied settings, measurements may taken observations---resulting moment estimates varying quality. derive weighted tensor decomposition approach computationally efficient non-weighted approach, demonstrate outperforms methods appropriately leverage less-observed dimensions.",18
"sharing deep generative representation perceived image reconstruction human brain activity. decoding human brain activities via functional magnetic resonance imaging (fmri) gained increasing attention recent years. encouraging results reported brain states classification tasks, reconstructing details human visual experience still remains difficult. two main challenges hinder development effective models perplexing fmri measurement noise high dimensionality limited data instances. existing methods generally suffer one issues yield dissatisfactory results. paper, tackle problem casting reconstruction visual stimulus bayesian inference missing view multiview latent variable model. sharing common latent representation, joint generative model external stimulus brain response ""deep"" extracting nonlinear features visual images, also powerful capturing correlations among voxel activities fmri recordings. nonlinearity deep structure endow model strong representation ability, correlations voxel activities critical suppressing noise improving prediction. devise efficient variational bayesian method infer latent variables model parameters. improve reconstruction accuracy, latent representations testing instances enforced close neighbours training set via posterior regularization. experiments three fmri recording datasets demonstrate approach accurately reconstruct visual stimuli.",3
"accuracy latent-variable estimation bayesian semi-supervised learning. hierarchical probabilistic models, gaussian mixture models, widely used unsupervised learning tasks. models consist observable latent variables, represent observable data underlying data-generation process, respectively. unsupervised learning tasks, cluster analysis, regarded estimations latent variables based observable ones. estimation latent variables semi-supervised learning, labels observed, precise unsupervised, one concerns clarify effect labeled data. however, sufficient theoretical analysis accuracy estimation latent variables. previous study, distribution-based error function formulated, asymptotic form calculated unsupervised learning generative models. shown that, estimation latent variables, bayes method accurate maximum-likelihood method. present paper reveals asymptotic forms error function bayesian semi-supervised learning discriminative generative models. results show generative model, uses given data, performs better model well specified.",18
"distribution number unique original items bootstrap sample?. sampling replacement occurs many settings machine learning, notably bagging ensemble technique .632+ validation scheme. number unique original items bootstrap sample important role behaviour prediction models learned it. indeed, uncontrived examples duplicate items effect. purpose report present distribution number unique original items bootstrap sample clearly concisely, view enabling machine learning researchers understand control quantity existing future resampling techniques. describe key characteristics distribution along generalisation case items come distinct categories, classification. cases discuss normal limit, conduct empirical investigation derive heuristic normal approximation permissible.",18
"end-to-end attention-based large vocabulary speech recognition. many current state-of-the-art large vocabulary continuous speech recognition systems (lvcsr) hybrids neural networks hidden markov models (hmms). systems contain separate components deal acoustic modelling, language modelling sequence decoding. investigate direct approach hmm replaced recurrent neural network (rnn) performs sequence prediction directly character level. alignment input features desired character sequence learned automatically attention mechanism built rnn. predicted character, attention mechanism scans input sequence chooses relevant frames. propose two methods speed operation: limiting scan subset promising frames pooling time information contained neighboring frames, thereby reducing source sequence length. integrating n-gram language model decoding process yields recognition accuracies similar hmm-free rnn-based approaches.",3
"learning mixture linear regressions non-realizable setting. mixture linear regressions (mlr) well-studied topic, prior works usually analyze models prediction error. fact, {\em prediction} {\em loss} well-defined context mixtures. paper, first show mlr used prediction instead predicting label, model predicts list values (also known {\em list-decoding}). list size equal number components mixture, loss function defined minimum among losses resulted component models. show definition, solution empirical risk minimization (erm) achieves small probability prediction error. begs algorithm minimize empirical risk mlr, known computationally hard. prior algorithmic works mlr focus {\em realizable} setting, i.e., recovery parameters data probabilistically generated mixed linear (noisy) model. paper show version popular alternating minimization (am) algorithm finds best fit lines dataset even realizable model assumed, regularity conditions dataset initial points, thereby provides solution erm. provide algorithm runs polynomial time number datapoints, recovers good approximation best fit lines. two algorithms experimentally compared.",18
"extractive multi document summarization using dynamical measurements complex networks. due large amount textual information available internet, paramount relevance use techniques find relevant concise content. typical task devoted identification informative sentences documents called extractive document summarization task. paper, use complex network concepts devise extractive multi document summarization (mds) method, extracts central sentences several textual sources. proposed model, texts represented networks, nodes represent sentences edges established based number shared words. differently previous works, identification relevant terms guided characterization nodes via dynamical measurements complex networks, including symmetry, accessibility absorption time. evaluation proposed system revealed excellent results obtained particular dynamical measurements, including based exploration networks via random walks.",3
"study classification imbalanced partially-labelled data streams. domain radio astronomy currently facing significant computational challenges, foremost amongst posed development world's largest radio telescope, square kilometre array (ska). preliminary specifications instrument suggest final design incorporate 2000 3000 individual 15 metre receiving dishes, together expected produce data rate many tb/s. given high data rate, becomes crucial consider information processed stored maximise scientific utility. paper, consider one possible data processing scenario ska, purposes all-sky pulsar survey. particular treat selection promising signals ska processing pipeline data stream classification problem. consider feasibility classifying signals arrive via unlabelled heavily class imbalanced data stream, using currently available algorithms frameworks. results indicate existing stream learners exhibit unacceptably low recall real astronomical data used standard configuration; however, good false positive performance comparable accuracy static learners, suggests definite potential on-line solution particular big data challenge.",0
"mocca: dynamics evolution binary stars multiple stellar populations tidally filling underfilling globular star clusters. present new upgraded version mocca code study dynamical evolution globular clusters first application study evolution multiple stellar populations. explore range initial conditions spanning different structural parameters first (fg) second population (sg) analyze effect binary dynamics survival. set simulations shown represents first phase new mocca-survey-2 project extended. here, focus attention number ratio fg sg binaries, variation distance cluster center, way abundances affected various cluster initial properties. find sg stars abundant clusters initially tidally filling. conversely, fg stars stay abundant clusters initially tidally underfilling. also find ratio binary fractions affected way calculate fractions (e.g. use binaries, main-sequence binaries observational binaries, i.e. main-sequence stars > 0.4 msun, mass ratios > 0.5). find also evolution mixing populations presents features even take account single stars too. implies main-sequence stars good proxy probing entire populations fg sg star clusters. also discuss findings relate observations milky way gcs. show mocca models able reproduce observed range sg fractions milky way gc know fraction. show sg fractions depend initial conditions. provide also explanation could initial conditions star clusters, hubble time, numerous fg, numerous sg stars.",0
"optimization theory hebbian/anti-hebbian networks pca whitening. analyzing information streamed sensory organs, brains face challenges similar solved statistical signal processing. suggests biologically plausible implementations online signal processing algorithms may model neural computation. here, focus workhorses signal processing principal component analysis (pca) whitening maximize information transmission presence noise. adopt similarity matching framework, recently developed principal subspace extraction, modify existing objective functions adding decorrelating term. modified objective functions, derive online pca whitening algorithms implementable neural networks local learning rules, i.e. synaptic weight updates depend activity pre- postsynaptic neurons. theory offers principled model neural computations makes testable predictions dropout underutilized neurons.",15
"stability solitary waves nonlinear klein-gordon equations. stability topological solitary waves pulses one-dimensional nonlinear klein-gordon systems revisited. linearized equation describing small deviations around static solution leads sturm-liouville problem, solved systematic way $-l\,(l+1)\,\sech^2(x)$-potential, showing orthogonality completeness relations fulfilled set solutions values $l\in\mathbb{n}$. approach allows determine linear stability kinks pulses certain nonlinear klein-gordon equations. two families novel nonlinear klein-gordon potentials introduced. exact solutions (kinks pulses) potentials exactly calculated, even nonlinear potential explicitly known. kinks novel models found stable, whereas pulses unstable. stability pulses achieved introducing certain spatial inhomogeneities.",12
"joint modeling text acoustic-prosodic cues neural parsing. conversational speech, acoustic signal provides cues help listeners disambiguate difficult parses. automatically parsing spoken utterance, introduce model integrates transcribed text acoustic-prosodic features using convolutional neural network energy pitch trajectories coupled attention-based recurrent neural network accepts text word-based prosodic features. find different types acoustic-prosodic features individually helpful, together improve parse f1 scores significantly strong text-only baseline. study known sentence boundaries, error analysis shows main benefit acoustic-prosodic features sentences disfluencies attachment errors improved.",3
"multilevel weighted support vector machine classification healthcare data missing values. work motivated needs predictive analytics healthcare data represented electronic medical records. data invariably problematic: noisy, missing entries, imbalance classes interests, leading serious bias predictive modeling. since standard data mining methods often produce poor performance measures, argue development specialized techniques data-preprocessing classification. paper, propose new method simultaneously classify large datasets reduce effects missing values. based multilevel framework cost-sensitive svm expected maximization imputation method missing values, relies iterated regression analyses. compare classification results multilevel svm-based algorithms public benchmark datasets imbalanced classes missing values well real data health applications, show multilevel svm-based method produces fast, accurate robust classification results.",18
"denoising autoencoders fast combinatorial black box optimization. estimation distribution algorithms (edas) require flexible probability models efficiently learned sampled. autoencoders (ae) generative stochastic networks desired properties. integrate special type ae, denoising autoencoder (dae), eda evaluate performance dae-eda several combinatorial optimization problems single objective. asses number fitness evaluations well required cpu times. compare results performance bayesian optimization algorithm (boa) rbm-eda, another eda based generative neural network proven competitive boa. considered problem instances, dae-eda considerably faster boa rbm-eda, sometimes orders magnitude. number fitness evaluations higher boa, competitive rbm-eda. results show daes useful tools problems low non-negligible fitness evaluation costs.",3
"mental sampling multimodal representations. resources natural environment concepts semantic space distributed ""patchily"", large gaps patches. describe people's internal external foraging behavior, various random walk models proposed. particular, internal foraging modeled sampling: order gather relevant information making decision, people draw samples mental representation using random-walk algorithms markov chain monte carlo (mcmc). however, two common empirical observations argue simple sampling algorithms mcmc. first, spatial structure often best described l\'evy flight distribution: probability distance two successive locations follows power-law distances. second, temporal structure sampling humans animals produce long-range, slowly decaying serial correlations characterized $1/f$-like fluctuations. propose mental sampling done simple mcmc, instead adapted multimodal representations implemented metropolis-coupled markov chain monte carlo (mc$^3$), one first algorithms developed sampling multimodal distributions. mc$^3$ involves running multiple markov chains parallel target distributions different temperatures, swaps states chains whenever better location found. heated chains readily traverse valleys probability landscape propose moves far-away peaks, colder chains make local steps explore current peak patch. show mc$^3$ generates distances successive samples follow l\'evy flight distribution $1/f$-like serial correlations, providing single mechanistic account two puzzling empirical phenomena.",3
"proceedings 2017 icml workshop human interpretability machine learning (whi 2017). proceedings 2017 icml workshop human interpretability machine learning (whi 2017), held sydney, australia, august 10, 2017. invited speakers tony jebara, pang wei koh, david sontag.",18
"effectiveness data augmentation image classification using deep learning. paper, explore compare multiple solutions problem data augmentation image classification. previous work demonstrated effectiveness data augmentation simple techniques, cropping, rotating, flipping input images. artificially constrain access data small subset imagenet dataset, compare data augmentation technique turn. one successful data augmentations strategies traditional transformations mentioned above. also experiment gans generate images different styles. finally, propose method allow neural net learn augmentations best improve classifier, call neural augmentation. discuss successes shortcomings method various datasets.",3
"plug-&-play generation non-gaussian states light telecom wavelength. work marks important progress towards practical quantum optical technologies continuous variable regime, shows feasibility experiments non-gaussian state generation entirely relies plug-&-play components guided-wave optics technologies. strategy demonstrated experimentally heralded preparation low amplitude schr\""odinger cat states based single-photon subtraction squeezed vacuum. stages experiment based off-the-shelf fiber components. leads stable, compact, easily re-configurable realization, fully compatible existing fibre networks and, general, future out-of-the-laboratory applications.",17
"deep learning-based numerical methods high-dimensional parabolic partial differential equations backward stochastic differential equations. propose new algorithm solving parabolic partial differential equations (pdes) backward stochastic differential equations (bsdes) high dimension, making analogy bsde reinforcement learning gradient solution playing role policy function, loss function given error prescribed terminal condition solution bsde. policy function approximated neural network, done deep reinforcement learning. numerical results using tensorflow illustrate efficiency accuracy proposed algorithms several 100-dimensional nonlinear pdes physics finance allen-cahn equation, hamilton-jacobi-bellman equation, nonlinear pricing model financial derivatives.",10
"smoothgrad: removing noise adding noise. explaining output deep network remains challenge. case image classifier, one type explanation identify pixels strongly influence final decision. starting point strategy gradient class score function respect input image. gradient interpreted sensitivity map, several techniques elaborate basic idea. paper makes two contributions: introduces smoothgrad, simple method help visually sharpen gradient-based sensitivity maps, discusses lessons visualization maps. publish code experiments website results.",3
"novel scheme binarization vehicle images using hierarchical histogram equalization technique. automatic license plate recognition system challenging area research now-a-days binarization integral important part it. case real life scenario, existing methods fail properly binarize image vehicle congested road, captured ccd camera. current work applied histogram equalization technique complete image also different hierarchy image partitioning. novel scheme formulated giving membership value pixel hierarchy histogram equalization. image binarized depending net membership value pixel. technique exhaustively evaluated vehicle image dataset well license plate dataset, giving satisfactory performances.",3
"actually constrain $f_{\rm nl}$ using scale-dependent bias effect? illustration impact galaxy bias uncertainties using boss dr12 galaxy power spectrum. scale-dependent bias effect galaxy power spectrum promising probe local primordial non-gaussianity (png) parameter $f_{\rm nl}$, amplitude effect proportional $f_{\rm nl}b_{\phi}$, $b_{\phi}$ linear png galaxy bias parameter. knowledge $b_{\phi}$ currently limited, yet nearly existing $f_{\rm nl}$ constraints forecasts assume precise knowledge it. here, use boss dr12 galaxy power spectrum illustrate uncertain knowledge $b_{\phi}$ currently prevents us constraining $f_{\rm nl}$ given statistical precision $\sigma_{f_{\rm nl}}$. assuming different fixed choices relation $b_{\phi}$ linear density bias $b_1$, find $\sigma_{f_{\rm nl}}$ vary much order magnitude. strongest bound $f_{\rm nl} = 16 \pm 16\ (1\sigma)$, loosest $f_{\rm nl} = 230 \pm 226\ (1\sigma)$ boss data. impact $b_{\phi}$ especially pronounced close zero. also show marginalizing $b_{\phi}$ wide priors conservative, leads fact biased constraints parameter space projection effects. independently galaxy bias assumptions, scale-dependent bias effect used detect $f_{\rm nl} \neq 0$ constraining product $f_{\rm nl}b_{\phi}$, error bar $\sigma_{f_{\rm nl}}$ remains undetermined results cannot compared cmb; find $f_{\rm nl}b_{\phi} \neq 0$ $1.6\sigma$ significance. also comment issues important analyses galaxy bispectrum. results strongly motivate simulation-based research programs aimed robust theoretical priors $b_{\phi}$ parameter, without may never able competitively constrain $f_{\rm nl}$ using galaxy data.",0
"simulations validation vision systems. computer vision matures systems science engineering discipline, trend leveraging latest advances computer graphics simulations performance evaluation, learning, inference. however, open question utility graphics simulations vision apparently contradicting views literature. paper, place results recent literature context performance characterization methodology outlined 90's note insights derived simulations qualitative quantitative depending degree fidelity models used simulation nature question posed experimenter. describe simulation platform incorporates latest graphics advances use systematic performance characterization trade-off analysis vision system design. verify utility platform case study validating generative model inspired vision hypothesis, rank-order consistency model, contexts global local illumination changes, bad weather, high-frequency noise. approach establishes link alternative viewpoints, involving models physics based semantics signal perturbation semantics confirms insights literature robust change detection.",3
"detecting dependencies sparse, multivariate databases using probabilistic programming non-parametric bayes. datasets hundreds variables many missing values commonplace. setting, statistically computationally challenging detect true predictive relationships variables also suppress false positives. paper proposes approach combines probabilistic programming, information theory, non-parametric bayes. shows use bayesian non-parametric modeling (i) build ensemble joint probability models variables; (ii) efficiently detect marginal independencies; (iii) estimate conditional mutual information arbitrary subsets variables, subject broad class constraints. users access capabilities using bayesdb, probabilistic programming platform probabilistic data analysis, writing queries simple, sql-like language. paper demonstrates empirically method (i) detect context-specific (in)dependencies challenging synthetic problems (ii) yield improved sensitivity specificity baselines statistics machine learning, real-world database 300 sparsely observed indicators macroeconomic development public health.",18
"deep affordance-grounded sensorimotor object recognition. well-established cognitive neuroscience human perception objects constitutes complex process, object appearance information combined evidence so-called object ""affordances"", namely types actions humans typically perform interacting them. fact recently motivated ""sensorimotor"" approach challenging task automatic object recognition, information sources fused improve robustness. work, aforementioned paradigm adopted, surpassing current limitations sensorimotor object recognition research. specifically, deep learning paradigm introduced problem first time, developing number novel neuro-biologically neuro-physiologically inspired architectures utilize state-of-the-art neural networks fusing available information sources multiple ways. proposed methods evaluated using large rgb-d corpus, specifically collected task sensorimotor object recognition made publicly available. experimental results demonstrate utility affordance information object recognition, achieving 29% relative error reduction inclusion.",3
"trapped mode excitation all-dielectric metamaterials loss gain. non-hermitian photonics based combining loss gain media within single optical system provides number approaches control generate flow light. paper, show introducing non-hermitian perturbation system loss gain constituents, high-quality resonances known trapped modes excited without need change symmetry unit cell geometry. demonstrate idea, consider widely used all-dielectric planar metamaterial whose unit cell consists pair rectangular nanoantennas made ordinal (lossy) doped (gainy) silicon. since quality factor trapped-mode resonance controlled changing geometrical symmetry non-hermiticity, varying loss gain allows us compensate influence asymmetry restore almost perfectly localized trapped mode. results obtained suggest new ways achieve high-quality resonances non-hermitian metasurfaces promising many practical applications nanophotonics.",14
"representing data sparse combination contextual data points classification. paper, study problem using contextual da- ta points data point classification problem. propose represent data point sparse linear reconstruction context, learn sparse context gather linear classifier su- pervised way increase discriminative ability. proposed novel formulation context learning, modeling learning context reconstruction coefficients classifier unified objective. objective, reconstruction error minimized coefficient spar- sity encouraged. moreover, hinge loss classifier minimized complexity classifier reduced. objective opti- mized alternative strategy iterative algorithm. experiments three benchmark data set show advantage state-of-the-art context-based data representation classification methods.",3
"towards new paradigm assistive technology home: research challenges, design issues performance assessment. providing elderly people special needs, including suffering physical disabilities chronic diseases, possibility retaining independence best one important challenges society expected face. assistance models based home care paradigm adopted rapidly almost industrialized emerging countries. paradigms hypothesize necessary ensure so-called activities daily living correctly regularly performed assisted person increase perception improved quality life. chapter describes computational inference engine core arianna, system able understand whether assisted person performs given set adl motivate him/her performing speech-mediated motivational dialogue, using set nearables installed apartment, plus wearable worn fit garments.",3
"integration declarative constraint programming. combining set existing constraint solvers integrated system cooperating solvers useful economic principle solve hybrid constraint problems. paper show approach also used integrate different language paradigms unified framework. furthermore, study syntactic, semantic operational impacts idea amalgamation declarative constraint programming.",3
"coherent pion production neutrino (anti-neutrino)-nucleus interaction. present study coherent pion production neutrino-nucleus interactions using formalism based partially conserved axial current theorem connects neutrino-nucleus cross section pion-nucleus elastic scattering cross section. pion-nucleus elastic scattering cross section calculated using glauber model takes three inputs, nuclear densities, pion-nucleon cross section $\alpha_{\pi n}$ ratio real imaginary part $\pi n$ forward scattering amplitude, parametrizations obtained measured data. calculate differential integrated cross sections charge neutral current coherent pion production neutrino (anti-neutrino)-nucleus scattering range nuclear targets light heavy materials lithium, carbon, hydrocarbon, oxygen, silicon, argon, iron lead. results cross section calculations compared measured data calculations berger-sehgal model genie package. excellent agreement calculated measured cross sections glauber model. genie berger-sehgal model give good description data lower energy range present calculations describe data energy ranges. predictions also made upcoming experiments like ino dune coherent region neutrino cross section.",8
constraints anomalous higgs boson couplings vector bosons fermions production higgs bosons using $ ττ$ final state. study anomalous couplings higgs boson vector bosons fermions presented. data recorded cms experiment center-of-mass energy pp collisions lhc 13 tev correspond integrated luminosity 138 fb$^{-1}$. study uses higgs boson candidates produced mainly gluon fusion electroweak vector boson fusion lhc subsequently decay pair $\tau$ leptons. matrix-element machine-learning techniques employed search anomalous interactions. results combined four-lepton two-photon decay channels yield stringent constraints anomalous higgs boson couplings date. pure $cp$-odd scenario higgs boson coupling gluons excluded 2.4 standard deviations. results consistent standard model predictions.,7
"gravity gapless phase biform symmetries. study effective field theories (efts) enjoying (maximal) biform symmetries. defined presence conserved (electric) current symmetries young tableau two columns equal length. theories also topological (magnetic) biform current, conservation law anomalous. go show mixed anomaly uniquely fixes two-point function electric magnetic currents. perform k\""all\'en-lehmann spectral decomposition current-current correlator, proving massless mode spectrum, whose masslessness protected anomaly. furthermore, anomaly gives rise universal form eft whose relevant term, resembles linear einstein action, dominates infrared physics. applications general formalism, study theories galileon superfluid linearized gravity. thus, one view masslessness graviton protected anomalous biform symmetries. associated eft provides organizing principle gravity low energies terms physical symmetries, allows interactions consistent linearized diffeomorphism invariance. theories ultraviolet-complete, relevant symmetries viewed emergent, include nonlinearities necessary make fully diffeomorphism invariant, contradiction expectation quantum gravity cannot global symmetries.",9
"shape analysis using fractal dimension: curvature based approach. present work shows novel fractal dimension method shape analysis. proposed technique extracts descriptors shape applying multiscale approach calculus fractal dimension shape. fractal dimension obtained application curvature scale-space technique original shape. application multiscale transform dimension calculus, obtained set numbers (descriptors) capable describing high precision shape analysis. obtained descriptors validated classification process. results demonstrate novel technique provides descriptors highly reliable, confirming precision proposed method.",14
"deep learning causal inference. paper, propose deep learning techniques econometrics, specifically causal inference estimating individual well average treatment effects. contribution paper twofold: 1. generalized neighbor matching estimate individual average treatment effects, analyze use autoencoders dimensionality reduction maintaining local neighborhood structure among data points embedding space. deep learning based technique shown perform better simple k nearest neighbor matching estimating treatment effects, especially data points several features/covariates reside low dimensional manifold high dimensional space. also observe better performance manifold learning methods neighbor matching. 2. propensity score matching one specific popular way perform matching order estimate average individual treatment effects. propose use deep neural networks (dnns) propensity score matching, present network called propensitynet this. generalization logistic regression technique traditionally used estimate propensity scores show empirically dnns perform better logistic regression propensity score matching. code methods made available shortly github at: https://github.com/vikas84bf",4
"swaps cost: case optimization-aware qubit routing. despite rapid advances quantum computing technologies, qubit connectivity limitation remains critical challenge. near-term nisq quantum computers relatively long-term scalable quantum architectures offer full connectivity. result, quantum circuits may directly executed quantum hardware, quantum compiler needs perform qubit routing make circuit compatible device layout. qubit routing step, compiler inserts swap gates performs circuit transformations. given connectivity topology target hardware, typically multiple qubit routing candidates. state-of-the-art compilers use cost function evaluate number swap gates different routes select one minimum number swap gates. qubit routing, quantum compiler performs gate optimizations upon circuit newly inserted swap gates. paper, observe aforementioned qubit routing optimal, qubit routing \textit{not} independent subsequent gate optimizations. find consideration gate optimizations, swap gates basis-gate cost. insights lead development qubit routing algorithm, nassc (not swaps cost). nassc first algorithm considers subsequent optimizations routing step. optimization-aware qubit routing leads better routing decisions benefits subsequent optimizations. also propose new optimization-aware decomposition inserted swap gates. experiments show routing overhead compiled routing algorithm reduced $69.30\%$ ($21.30\%$ average) number cnot gates $43.50\%$ ($7.61\%$ average) circuit depth compared state-of-the-art scheme, sabre.",17
"video semantic object segmentation self-adaptation dcnn. paper proposes new framework semantic segmentation objects videos. address label inconsistency problem deep convolutional neural networks (dcnns) exploiting fact videos multiple frames; frames object confidently-estimated (ce) use information improve labels frames. given semantic segmentation results frame obtained dcnn, sample several ce frames adapt dcnn model input video focusing specific instances video rather general objects various circumstances. propose offline online approaches different supervision levels. experiments method achieved great improvement original model previous state-of-the-art methods.",3
"comparisons wavelet functions qrs signal noise ratio enhancement detection accuracy. compare capability wavelet functions used noise removal preprocessing step qrs detection algorithm electrocardiogram (ecg) signal. qrs signal noise ratio enhancement detection accuracy wavelet function evaluated using three measures: (1) ratio maximum beat amplitude minimum beat amplitude (rmm), (2) mean absolute time error (mate), (3) figure merit (fom). three wavelet functions previous well-known publications explored, i.e., bior1.3, db10, mexican hat wavelet functions. results evaluated ecg signal mit-bih arrhythmia database show mexican hat wavelet function better others. scale 8 mexican hat wavelet function provide best enhancement qrs signal noise ratio, scale 4 mexican hat wavelet function provide best detection accuracy. results may combined may enable use single fixed threshold ecg records leading reduction computational complexity qrs detection algorithm.",3
"deep feature flow video recognition. deep convolutional neutral networks achieved great success image recognition tasks. yet, non-trivial transfer state-of-the-art image recognition networks videos per-frame evaluation slow unaffordable. present deep feature flow, fast accurate framework video recognition. runs expensive convolutional sub-network sparse key frames propagates deep feature maps frames via flow field. achieves significant speedup flow computation relatively fast. end-to-end training whole architecture significantly boosts recognition accuracy. deep feature flow flexible general. validated two recent large scale video datasets. makes large step towards practical video recognition.",3
"visual secret sharing scheme using grayscale images. pixel expansion quality reconstructed secret image major issue visual secret sharing (vss) schemes. number probabilistic vss schemes minimum pixel expansion proposed black white (binary) secret images. paper presents probabilistic (2, 3)-vss scheme gray scale images. pixel expansion larger size quality image perfect reconstructed. construction shadow images (transparent shares) based binary operation.",3
"delayed impact fair machine learning. fairness machine learning predominantly studied static classification settings without concern decisions change underlying population time. conventional wisdom suggests fairness criteria promote long-term well-being groups aim protect. study static fairness criteria interact temporal indicators well-being, long-term improvement, stagnation, decline variable interest. demonstrate even one-step feedback model, common fairness criteria general promote improvement time, may fact cause harm cases unconstrained objective would not. completely characterize delayed impact three standard criteria, contrasting regimes exhibit qualitatively different behavior. addition, find natural form measurement error broadens regime fairness criteria perform favorably. results highlight importance measurement temporal modeling evaluation fairness criteria, suggesting range new challenges trade-offs.",3
"automated text categorization framework based hyperparameter optimization. great variety text tasks topic spam identification, user profiling, sentiment analysis posed supervised learning problem tackle using text classifier. text classifier consists several subprocesses, general enough applied supervised learning problem, whereas others specifically designed tackle particular task, using complex computational expensive processes lemmatization, syntactic analysis, etc. contrary traditional approaches, propose minimalistic wide system able tackle text classification tasks independent domain language, namely microtc. composed easy implement text transformations, text representations, supervised learning algorithm. pieces produce competitive classifier even domain informally written text. provide detailed description microtc along extensive experimental comparison relevant state-of-the-art methods. mircotc compared 30 different datasets. regarding accuracy, microtc obtained best performance 20 datasets achieves competitive results remaining 10. compared datasets include several problems like topic polarity classification, spam detection, user profiling authorship attribution. furthermore, important state approach allows usage technology even without knowledge machine learning natural language processing.",3
bayesian networks point view chain graphs. athe paper gives arguments favour use chain graphs description probabilistic conditional independence structures. every bayesian network model equivalently introduced means factorization formula respect chain graph markov equivalent bayesian network. graphical characterization graphs given. class equivalent graphs represented distinguished graph called largest chain graph. factorization formula respect largest chain graph basis proposal represent corresponding (discrete) probability distribution computer (i.e. parametrize it). way depend choice particular bayesian network class equivalent networks seems efficient way point view memory demands. separation criterion reading independency statements chain graph formulated simpler way. resembles well-known d-separation criterion bayesian networks implemented locally.,3
"iterative learning answer set programs context dependent examples. recent years, several frameworks systems proposed extend inductive logic programming (ilp) answer set programming (asp) paradigm. ilp, examples must explained hypothesis together given background knowledge. existing systems, background knowledge examples; however, examples may context-dependent. means examples explained context information, whereas others explained different contexts. paper, capture notion present context-dependent extension learning ordered answer sets framework. extension, contexts used structure background knowledge. propose new iterative algorithm, ilasp2i, exploits feature scale existing ilasp2 system learning tasks large numbers examples. demonstrate gain scalability applying algorithms various learning tasks. results show that, compared ilasp2, newly proposed ilasp2i system two orders magnitude faster use two orders magnitude less memory, whilst preserving average accuracy. paper consideration acceptance tplp.",3
"dynamic vulnerability map assess risk road network traffic utilization. le havre agglomeration (codah) includes 16 establishments classified seveso high threshold. literature, construct vulnerability maps help decision makers assess risk. approaches remain static take account population displacement estimation vulnerability. propose decision making tool based dynamic vulnerability map evaluate difficulty evacuation different sectors codah. use geographic information system (gis) visualize map evolves road traffic state detection communities large graphs algorithm.",3
"enabling multi-source neural machine translation concatenating source sentences multiple languages. paper, propose novel elegant solution ""multi-source neural machine translation"" (msnmt) relies preprocessing n-way multilingual corpus without modifying neural machine translation (nmt) architecture training procedure. simply concatenate source sentences form single long multi-source input sentence keeping target side sentence train nmt system using preprocessed corpus. evaluate method resource poor well resource rich settings show effectiveness (up 4 bleu using 2 source languages 6 bleu using 5 source languages) comparing existing methods msnmt. also provide insights nmt system leverages multilingual information scenario visualizing attention.",3
"conditional generative adversarial networks speech enhancement noise-robust speaker verification. improving speech system performance noisy environments remains challenging task, speech enhancement (se) one effective techniques solve problem. motivated promising results generative adversarial networks (gans) variety image processing tasks, explore potential conditional gans (cgans) se, particular, make use image processing framework proposed isola et al. [1] learn mapping spectrogram noisy speech enhanced counterpart. se cgan consists two networks, trained adversarial manner: generator tries enhance input noisy spectrogram, discriminator tries distinguish enhanced spectrograms provided generator clean ones database using noisy spectrogram condition. evaluate performance cgan method terms perceptual evaluation speech quality (pesq), short-time objective intelligibility (stoi), equal error rate (eer) speaker verification (an example application). experimental results show cgan method overall outperforms classical short-time spectral amplitude minimum mean square error (stsa-mmse) se algorithm, comparable deep neural network-based se approach (dnn-se).",5
"hamiltonian operator spectral shape analysis. many shape analysis methods treat geometry object metric space captured laplace-beltrami operator. paper, propose adapt classical hamiltonian operator quantum mechanics field shape analysis. end study addition potential function laplacian generator dual spaces shape processing performed. present general optimization approach solving variational problems involving basis defined hamiltonian using perturbation theory eigenvectors. suggested operator shown produce better functional spaces operate with, demonstrated different shape analysis tasks.",3
"group delay time fermions graphene tilted potential barrier. group delay time dirac fermions subjected tilting barrier potential along $ x $-axis investigated graphene. start finding eigenspinor solution dirac equation relating incident, reflected, transmitted beam waves. relationship allows us compute group delay time transmission reflection obtaining corresponding phase shifts. discovered barrier width, incident energy, incident angle used modify group delay time, particles travel barrier fermi velocity $ v_f $. findings also show transmission group delay might controlled, gate voltage control could useful graphene-based tilting barriers.",2
"estimation clustering infinite rankings. paper presents natural extension stagewise ranking case infinitely many items. introduce infinite generalized mallows model (igm), describe properties give procedures estimate data. estimation multimodal distributions introduce exponential-blurring-mean-shift nonparametric clustering algorithm. experiments highlight properties new model demonstrate infinite models simple, elegant practical.",3
"asynchronous distributed framework large-scale learning based parameter exchanges. many distributed learning problems, heterogeneous loading computing machines may harm overall performance synchronous strategies. paper, propose effective asynchronous distributed framework minimization sum smooth functions, machine performs iterations parallel local function updates shared parameter asynchronously. way, machines continuously work even though latest version shared parameter. prove convergence consistency general distributed asynchronous method gradient iterations show efficiency matrix factorization problem recommender systems binary classification.",18
"density estimation via discrepancy based adaptive sequential partition. given $iid$ observations unknown absolute continuous distribution defined domain $\omega$, propose nonparametric method learn piecewise constant function approximate underlying probability density function. density estimate piecewise constant function defined binary partition $\omega$. key ingredient algorithm use discrepancy, concept originates quasi monte carlo analysis, control partition process. resulting algorithm simple, efficient, provable convergence rate. empirically demonstrate efficiency density estimation method. present applications wide range tasks, including finding good initializations k-means.",18
"worst-case optimal query answering greedy sets existential rules subclasses. need ontological layer top data, associated advanced reasoning mechanisms able exploit semantics encoded ontologies, acknowledged database knowledge representation communities. focus paper ontological query answering problem, consists querying data taking ontological knowledge account. specifically, establish complexities conjunctive query entailment problem classes existential rules (also called tuple-generating dependencies, datalog+/- rules, forall-exists-rules. contribution twofold. first, introduce class greedy bounded-treewidth sets (gbts) rules, covers guarded rules, well-known generalizations. provide generic algorithm query entailment gbts, worst-case optimal combined complexity without bounded predicate arity, well data complexity query complexity. secondly, classify several gbts classes, whose complexity unknown, respect combined complexity (with unbounded bounded predicate arity) data complexity obtain comprehensive picture complexity existential rule fragments based diverse guardedness notions. upper bounds provided showing proposed algorithm optimal them.",3
"event reconstruction compton telescopes using multi-task neural network. developed neural network model perform event reconstruction compton telescopes. model reconstructs events consist three interactions detector. essential compton telescopes determine time order gamma-ray interactions whether incident photon deposits energy detector escapes detector. model simultaneously predicts two essential factors using multi-task neural network three hidden layers fully connected nodes. verification, conducted numerical experiments using monte carlo simulation, assuming large-area compton telescope using liquid argon measure gamma rays energies $3.0\,\mathrm{mev}$. reconstruction model shows excellent performance event reconstruction multiple scattering events consist eight hits. accuracies hit order prediction around $60\%$ escape flags higher $70\%$ eight-hit events $4\pi$ isotropic photons. compared two algorithms, classical model physics-based probabilistic one, present neural network method shows high performance estimation accuracy particularly number scattering small, 3 4. since simulation data easily optimize network model, model flexibly applied wide variety compton telescopes.",0
"introduction multi-agent simulation. designing systems complex, dynamic stochastic nature, simulation generally recognised one best design support technologies, valuable aid strategic tactical decision making process. simulation model consists set rules define system changes time, given current state. unlike analytical models, simulation model solved run changes system states observed point time. provides insight system dynamics rather predicting output system based specific inputs. simulation decision making tool decision support tool, allowing better informed decisions made. due complexity real world, simulation model approximation target system. essence art simulation modelling abstraction simplification. characteristics important study analysis target system included simulation model.",3
"active tolerant testing. work, give first algorithms tolerant testing nontrivial classes active model: estimating distance target function hypothesis class c respect arbitrary distribution d, using small number label queries polynomial-sized pool unlabeled examples drawn d. specifically, show class unions intervals line, estimate error rate best hypothesis class additive error epsilon $o(\frac{1}{\epsilon^6}\log \frac{1}{\epsilon})$ label queries unlabeled pool size $o(\frac{d}{\epsilon^2}\log \frac{1}{\epsilon})$. key point number labels needed independent vc-dimension class. extends work balcan et al. [2012] solved non-tolerant testing problem class (distinguishing zero-error case case best hypothesis class error greater epsilon). also consider related problem estimating performance given learning algorithm setting. is, given large pool unlabeled examples drawn distribution d, we, label queries, estimate well would perform entire dataset labeled? focus k-nearest neighbor style algorithms, also show results applied problem hyperparameter tuning (selecting best value k given learning problem).",18
"fuzzy logic system analyze student's lifestyle. college student's life primarily categorized domains education, health, social activities may include daily chores travelling time. time management crucial every student. self realisation one's daily time expenditure various domains therefore essential maximize one's effective output. paper presents mobile application using fuzzy logic global positioning system (gps) analyzes student's lifestyle provides recommendations suggestions based results.",3
"noether symmetries first integrals damped harmonic oscillator. noether theorem establishes interesting connection symmetries action integral conservation laws dynamical system. aim present work classify damped harmonic oscillator problem respect noether symmetries construct corresponding conservation laws over-damped, damped critical damped cases. case obtain maximum five linearly independent group generators provide related five conserved quantities. remarkably, obtaining complete set invariant quantities obtain analytical solutions case. current work, also introduce new lagrangian damped harmonic oscillator. though form new lagrangian presented bateman completely different, yet generates set noether symmetries conserved quantities. so, new form lagrangian presenting may seriously interesting physicists. moreover, also find lie algebras noether symmetries point interesting aspects results related noether symmetries first integrals damped harmonic oscillator perhaps reported earlier studies.",11
"lifelong learning crf supervised aspect extraction. paper makes focused contribution supervised aspect extraction. shows system performed aspect extraction many past domains retained results knowledge, conditional random fields (crf) leverage knowledge lifelong learning manner extract new domain markedly better traditional crf without using prior knowledge. key innovation even crf training, model still improve extraction experiences applications.",3
"no-u-turn sampler: adaptively setting path lengths hamiltonian monte carlo. hamiltonian monte carlo (hmc) markov chain monte carlo (mcmc) algorithm avoids random walk behavior sensitivity correlated parameters plague many mcmc methods taking series steps informed first-order gradient information. features allow converge high-dimensional target distributions much quickly simpler methods random walk metropolis gibbs sampling. however, hmc's performance highly sensitive two user-specified parameters: step size {\epsilon} desired number steps l. particular, l small algorithm exhibits undesirable random walk behavior, l large algorithm wastes computation. introduce no-u-turn sampler (nuts), extension hmc eliminates need set number steps l. nuts uses recursive algorithm build set likely candidate points spans wide swath target distribution, stopping automatically starts double back retrace steps. empirically, nuts perform least efficiently sometimes efficiently well tuned standard hmc method, without requiring user intervention costly tuning runs. also derive method adapting step size parameter {\epsilon} fly based primal-dual averaging. nuts thus used hand-tuning all. nuts also suitable applications bugs-style automatic inference engines require efficient ""turnkey"" sampling algorithms.",18
"beyond zipf-mandelbrot law quantitative linguistics. paper zipf-mandelbrot law revisited context linguistics. despite widespread popularity zipf--mandelbrot law describe statistical behaviour rather restricted fraction total number words contained given corpus. particular, focus attention important deviations become statistically relevant larger corpora considered ultimately could understood salient features underlying complex process language generation. finally, shown different observed regimes accurately encompassed within single mathematical framework recently introduced c. tsallis.",2
"kernel regression mode calculation conditional probability distribution. direct way express arbitrary dependencies datasets estimate joint distribution apply afterwards argmax-function obtain mode corresponding conditional distribution. method practice difficult, requires global optimization complicated function, joint distribution fixed input variables. article proposes method finding global maxima joint distribution modeled kernel density estimation. experiments show advantages shortcomings resulting regression method comparison standard nadaraya-watson regression technique, approximates optimum expectation value.",18
"exponential ergodicity branching processes immigration competition. study ergodic property continuous-state branching process immigration competition, extension models studied lambert (2005, ann. appl. probab.), pardoux (2016, springer) berestycki et al. (2018, probab. theory related fields) additional immigration structure. exponential ergodicity weighted total variation distance proved natural assumptions. main theorem applies branching mechanisms full range criticality, including stable types. proof based markov coupling process nonsymmetric control function distance, designed identify take advantage dominating factor branching, immigration competition mechanisms different parts state space. approach provides way finding explicitly ergodicity rate.",10
"flexibly instructable agents. paper presents approach learning situated, interactive tutorial instruction within ongoing agent. tutorial instruction flexible (and thus powerful) paradigm teaching tasks allows instructor communicate whatever types knowledge agent might need whatever situations might arise. support flexibility, however, agent must able learn multiple kinds knowledge broad range instructional interactions. approach, called situated explanation, achieves learning combination analytic inductive techniques. combines form explanation-based learning situated instruction full suite contextually guided responses incomplete explanations. approach implemented agent called instructo-soar learns hierarchies new tasks domain knowledge interactive natural language instructions. instructo-soar meets three key requirements flexible instructability distinguish previous systems: (1) take known unknown commands instruction point; (2) handle instructions apply either current situation hypothetical situation specified language (as in, instance, conditional instructions); (3) learn, instructions, class knowledge uses perform tasks.",3
"actions eye: dynamic gaze datasets learnt saliency models visual recognition. systems based bag-of-words models image features collected maxima sparse interest point operators used successfully computer visual object action recognition tasks. sparse, interest-point based approach recognition inconsistent visual processing biological systems operate `saccade fixate' regimes, methodology emphasis human computer vision communities remains sharply distinct. here, make three contributions aiming bridge gap. first, complement existing state-of-the art large scale dynamic computer vision annotated datasets like hollywood-2 ucf sports human eye movements collected ecological constraints visual action recognition task. knowledge first large human eye tracking datasets collected made publicly available video, vision.imar.ro/eyetracking (497,107 frames, viewed 16 subjects), unique terms (a) large scale computer vision relevance, (b) dynamic, video stimuli, (c) task control, opposed free-viewing. second, introduce novel sequential consistency alignment measures, underline remarkable stability patterns visual search among subjects. third, leverage significant amount collected data order pursue studies build automatic, end-to-end trainable computer vision systems based human eye movements. studies shed light differences computer vision spatio-temporal interest point image sampling strategies human fixations, well impact visual recognition performance, also demonstrate human fixations accurately predicted, used end-to-end automatic system, leveraging advanced computer vision practice, lead state art results.",3
"heterogeneous information network embedding meta path based proximity. network embedding representation large graph low-dimensional space, vertices modeled vectors. objective good embedding preserve proximity vertices original graph. way, typical search mining methods applied embedded space help off-the-shelf multidimensional indexing approaches. existing network embedding techniques focus homogeneous networks, vertices considered belong single class.",3
"facial transformations ancient portraits: face caesar. software solutions used obtain facial transformations help investigating artistic metamorphosis ancient portraits person. analysis freely available software portraitures julius caesar proposed, showing several ""morphs"". software helps enhancing mood artist added portrait.",3
"estimating coefficients mixture two linear regressions expectation maximization. give convergence guarantees estimating coefficients symmetric mixture two linear regressions expectation maximization (em). particular, show convergence empirical iterates guaranteed provided algorithm initialized unbounded cone. is, initializer large cosine angle population coefficient vector signal noise ratio (snr) large, sample-splitting version em algorithm converges true coefficient vector high probability. ""large"" means quantity required least universal constant. finally, show population em operator globally contractive characterizing region fails. give empirical evidence suggests sample based em performs poorly intitializers drawn set. interestingly, analysis borrows tools used problem estimating centers symmetric mixture two gaussians em.",18
"towards multiple kernel principal component analysis integrative analysis tumor samples. personalized treatment patients based tissue-specific cancer subtypes strongly increased efficacy chosen therapies. even though amount data measured cancer patients increased last years, cancer subtypes still diagnosed based individual data sources (e.g. gene expression data). propose unsupervised data integration method based kernel principal component analysis. principal component analysis one widely used techniques data analysis. unfortunately, straight-forward multiple-kernel extension method leads use one input matrices, fit goal gaining information data sources. therefore, present scoring function determine impact input matrix. approach enables visualizing integrated data subsequent clustering cancer subtype identification. due nature method, free parameters set. apply methodology five different cancer data sets demonstrate advantages terms results usability.",18
"probabilistic selection agentspeak(l). agent programming mostly symbolic discipline and, such, draws little benefits probabilistic areas machine learning graphical models. however, greatest objective agent research achievement autonomy dynamical complex environments --- goal implies embracing uncertainty therefore entailed representations, algorithms techniques. paper proposes innovative conflict free two layer approach agent programming uses already established methods tools symbolic probabilistic artificial intelligence. moreover, framework illustrated means widely used agent programming example, goldminers.",3
"network-based methods outcome prediction ""sample space"". thesis present novel semi-supervised network-based algorithm p-net, able rank classify patients respect specific phenotype clinical outcome study. peculiar innovative characteristic method builds network samples/patients, nodes represent samples edges functional genetic relationships individuals (e.g. similarity expression profiles), predict phenotype study. words, constructs network ""sample space"" ""biomarker space"" (where nodes represent biomolecules (e.g. genes, proteins) edges represent functional genetic relationships nodes), usual state-of-the-art methods. assess performances p-net, apply three different publicly available datasets patients afflicted specific type tumor: pancreatic cancer, melanoma ovarian cancer dataset, using data following experimental set-up proposed two recently published papers [barter et al., 2014, winter et al., 2012]. show network-based methods ""sample space"" achieve results competitive classical supervised inductive systems. moreover, graph representation samples easily visualized networks used gain visual clues relationships samples, taking account phenotype associated predicted sample. knowledge one first works proposes graph-based algorithms working ""sample space"" biomolecular profiles patients predict phenotype outcome, thus contributing novel research line framework network medicine.",3
"interactive decomposition multi-objective optimization via progressively learned value functions. decomposition become increasingly popular technique evolutionary multi-objective optimization (emo). decomposition-based emo algorithm usually designed approximate whole pareto-optimal front (pf). however, practice, decision maker (dm) might interested her/his region interest (roi), i.e., part pf. solutions outside might useless even noisy decision-making procedure. furthermore, guarantee find preferred solutions tackling many-objective problems. paper develops interactive framework decomposition-based emo algorithm lead dm preferred solutions her/his choice. consists three modules, i.e., consultation, preference elicitation optimization. specifically, every several generations, dm asked score candidate solutions consultation session. thereafter, approximated value function, models dm's preference information, progressively learned dm's behavior. preference elicitation session, preference information learned consultation module translated form used decomposition-based emo algorithm, i.e., set reference points biased toward roi. optimization module, decomposition-based emo algorithm principle, utilizes biased reference points direct search process. extensive experiments benchmark problems three ten objectives fully demonstrate effectiveness proposed method finding dm's preferred solutions.",3
"cascade residual learning: two-stage convolutional neural network stereo matching. leveraging recent developments convolutional neural networks (cnns), matching dense correspondence stereo pair cast learning problem, performance exceeding traditional approaches. however, remains challenging generate high-quality disparities inherently ill-posed regions. tackle problem, propose novel cascade cnn architecture composing two stages. first stage advances recently proposed dispnet equipping extra up-convolution modules, leading disparity images details. second stage explicitly rectifies disparity initialized first stage; couples first-stage generates residual signals across multiple scales. summation outputs two stages gives final disparity. opposed directly learning disparity second stage, show residual learning provides effective refinement. moreover, also benefits training overall cascade network. experimentation shows cascade residual learning scheme provides state-of-the-art performance matching stereo correspondence. time submission paper, method ranks first kitti 2015 stereo benchmark, surpassing prior works noteworthy margin.",3
"universal representations:the missing link faces, text, planktons, cat breeds. advent large labelled datasets high-capacity models, performance machine vision systems improving rapidly. however, technology still major limitations, starting fact different vision problems still solved different models, trained scratch fine-tuned target data. human visual system, stark contrast, learns universal representation vision early life individual. representation works well enormous variety vision problems, little change, major advantage requiring little training data solve them.",3
"strategic dialogue management via deep reinforcement learning. artificially intelligent agents equipped strategic skills negotiate interactions natural artificial agents still underdeveloped. paper describes successful application deep reinforcement learning (drl) training intelligent agents strategic conversational skills, situated dialogue setting. previous studies modelled behaviour strategic agents using supervised learning traditional reinforcement learning techniques, latter using tabular representations learning linear function approximation. study, apply drl high-dimensional state space strategic board game settlers catan---where players offer resources exchange others also reply offers made players. experimental results report drl-based learnt policies significantly outperformed several baselines including random, rule-based, supervised-based behaviours. drl-based policy 53% win rate versus 3 automated players (`bots'), whereas supervised player trained dialogue corpus setting achieved 27%, versus 3 bots. result supports claim drl promising framework training dialogue systems, strategic agents negotiation abilities.",3
"network lasso accurate?. ""least absolute shrinkage selection operator"" (lasso) method adapted recently networkstructured datasets. particular, network lasso method allows learn graph signals small number noisy signal samples using total variation graph signal regularization. efficient scalable implementations network lasso available, little known conditions underlying network structure ensure network lasso accurate. leveraging concepts compressed sensing, address gap derive precise conditions underlying network topology sampling set guarantee network lasso particular loss function deliver accurate estimate entire underlying graph signal. also quantify error incurred network lasso terms two constants reflect connectivity sampled nodes.",18
ensemble adaptive neuro-fuzzy kohonen networks online data stream fuzzy clustering. new approach data stream clustering help ensemble adaptive neuro-fuzzy systems proposed. proposed ensemble formed adaptive neuro-fuzzy self-organizing kohonen maps parallel processing mode. final result chosen best neuro-fuzzy self-organizing kohonen map.,3
"quantifying prosodic variability middle english alliterative poetry. interest mathematical structure poetry dates back least 19th century: retiring mathematics position, j. j. sylvester wrote book prosody called $\textit{the laws verse}$. today interest computer analysis poems, paper discusses statistical approach applied task. starting definition middle english alliteration is, $\textit{sir gawain green knight}$ william langland's $\textit{piers plowman}$ used illustrate methodology. theory first developed analyzing data riemannian manifold turns applicable strings allowing one compute generalized mean variance textual data, applied poems above. ratio two variances produces analogue f test, resampling allows p-values estimated. consequently, methodology provides way compare prosodic variability two texts.",18
"distributed self management distributed security systems. distributed system e.g. artificial immune systems, complex adaptive systems, multi-agent systems widely used computer science, e.g. network security, optimisations, simulations. systems, small entities move network perform certain tasks. time, entities move another place require therefore information move profitable. common used systems provide information use centralised approach center delegates entities. article discusses whether small information neighbours enhances performance overall system not. therefore, two information-protocols introduced analysed. addition, protocols implemented tested using artificial immune system sana protects network intrusions.",3
"online control false discovery rate decaying memory. online multiple testing problem, p-values corresponding different null hypotheses observed one one, decision whether reject current hypothesis must made immediately, next p-value observed. alpha-investing algorithms control false discovery rate (fdr), formulated foster stine, generalized applied many settings, including quality-preserving databases science multiple a/b multi-armed bandit tests internet commerce. paper improves class generalized alpha-investing algorithms (gai) four ways: (a) show uniformly improve power entire class monotone gai procedures awarding alpha-wealth rejection, giving win-win resolution recent dilemma raised javanmard montanari, (b) demonstrate incorporate prior weights indicate domain knowledge hypotheses likely non-null, (c) allow differing penalties false discoveries indicate hypotheses may important others, (d) define new quantity called decaying memory false discovery rate (mem-fdr) may meaningful truly temporal applications, alleviates problems describe refer ""piggybacking"" ""alpha-death"". gai++ algorithms incorporate four generalizations simultaneously, reduce powerful variants earlier algorithms weights decay set unity. finally, also describe simple method derive new online fdr rules based estimated false discovery proportion.",18
"effect sintering conditions microstructure electrochemical properties 6scandia 1ceria co-doped zirconia (6sc1cezr). current work studies effect sintering conditions densification behavior, microstructure, electrochemical performance 6 mol% sc2o3, 1 mol% ceo2 co-doped zro2 (6sc1cezr) solid-state electrolyte applications. specimen's physical properties electrochemical performance sintered different temperatures ranging 1000 {\deg}c 1700 {\deg}c 6 hours examined diffraction, microscopy, raman electrochemical impedance spectroscopy. activation energy pre-exponential conductivity factor oxygen ions calculated using linear fit arrhenius-type conductivity equation. observe samples sintered 1500 {\deg}c 6h show lower c/a axial ratio tetragonal phase, highest relative density, highest total conductivity among samples investigated. higher total conductivity explained formation higher symmetric tetragonal phase simultaneous existence cubic tetragonal phases (with higher symmetry), supported presence cerium ions partially replaced zirconium ions' sites.",2
"eigenvalue bounds kirchhoff laplacian. prove eigenvalue l(k) kirchhoff laplacian k graph bounded d(k)+d(k-1) k {1,...,n}. l(1),...,l(n) non-decreasing list eigenvalues k d(1),..,d(n) non-decreasing list vertex degrees additional assumption d(0)=0. case k=2 case schur-horn inequality l(1)+l(2)+...+l(k) bounded d(1)+d(2)+...+d(k). case k=n result anderson morley. already corollary l(k) less equal 2 d(k) case stronger gershgorin circle theorem assuring every disk radius d(k) centered d(k) contains eigenvalue k. prove theorem using cauchy interlace theorem.",10
"microstructure representation reconstruction heterogeneous materials via deep belief network computational material design. integrated computational materials engineering (icme) aims accelerate optimal design complex material systems integrating material science design automation. tractable icme, required (1) structural feature space identified allow reconstruction new designs, (2) reconstruction process property-preserving. majority existing structural presentation schemes rely designer's understanding specific material systems identify geometric statistical features, could biased insufficient reconstructing physically meaningful microstructures complex material systems. paper, develop feature learning mechanism based convolutional deep belief network automate two-way conversion microstructures lower-dimensional feature representations, achieves 1000-fold dimension reduction microstructure space. proposed model applied wide spectrum heterogeneous material systems distinct microstructural features including ti-6al-4v alloy, pb63-sn37 alloy, fontainebleau sandstone, spherical colloids, produce material reconstructions close original samples respect 2-point correlation functions mean critical fracture strength. capability achieved existing synthesis methods rely markovian assumption material microstructures.",2
"equatorially balanced c4-face-magic labelings klein bottle grid graphs. graph $g = (v, e)$ embedded klein bottle, let $\mathcal{f}(g)$ denote set faces $g$. then, $g$ called $c_k$-face-magic klein bottle graph exists bijection $f: v(g) \to \{1, 2, \dots, |v(g)|\}$ $f \in \mathcal{f}(g)$ $f \cong c_k$, sum vertex labelings along $c_k$ constant $s$. let $x_v =f(v)$ $v\in v(g)$. call $\{x_v : v\in v(g)\}$ $c_k$-face-magic klein bottle labeling $g$. consider $m \times n$ grid graph, denoted $\mathcal{k}_{m,n}$, embedded klein bottle natural way. show $m,n\ge 2$, $\mathcal{k}_{m,n}$ admits $c_4$-face-magic klein bottle labeling $n$ even. say $c_4$-face-magic klein bottle labeling $\{x_{i,j}: (i,j) \in v(\mathcal{k}_{m,n}) \}$ $\mathcal{k}_{m,n}$ equatorially balanced $x_{i,j} + x_{i,n+1-j} = \tfrac{1}{2} s$ $(i,j) \in v(\mathcal{k}_{m,n})$. show $m$ odd, $c_4$-face-magic klein bottle labeling $\mathcal{k}_{m,n}$ must equatorially balanced. also $m$ odd, show (up symmetries klein bottle) number $c_4$-face-magic klein bottle labelings $m \times 4$ klein bottle grid graph $2^m \, (m-1)! \, \tau(m)$, $\tau(m)$ number positive divisors $m$. furthermore, let $m\ge 3$ odd integer $n \ge 6$ even integer. then, minimum number distinct $c_4$-face-magic klein bottle labelings $x$ $\mathcal{k}_{m,n}$ (up symmetries klein bottle) either $(5\cdot 2^m)(m-1)!$ $n \equiv 0\pmod{4}$, $(6\cdot 2^m)(m-1)!$ $n \equiv 2\pmod{4}$.",10
"skin lesion classification using deep multi-scale convolutional neural networks. present deep learning approach isic 2017 skin lesion classification challenge using multi-scale convolutional neural network. approach utilizes inception-v3 network pre-trained imagenet dataset, fine-tuned skin lesion classification using two different scales input images.",3
"analogy perception applied seven tests word comprehension. argued analogy core cognition. ai research, algorithms analogy often limited need hand-coded high-level representations input. alternative approach use high-level perception, high-level representations automatically generated raw data. analogy perception process recognizing analogies using high-level perception. present pairclass, algorithm analogy perception recognizes lexical proportional analogies using representations automatically generated large corpus raw textual data. proportional analogy analogy form a:b::c:d, meaning ""a b c d"". lexical proportional analogy proportional analogy words, carpenter:wood::mason:stone. pairclass represents semantic relations two words using high-dimensional feature vector, elements based frequencies patterns corpus. pairclass recognizes analogies applying standard supervised machine learning techniques feature vectors. show seven different tests word comprehension framed problems analogy perception apply pairclass seven resulting sets analogy perception problems. achieve competitive results seven tests. first time uniform approach handled range tests word comprehension.",3
"venus exoplanet: i. initial exploration 3-d energy balance co$_{2}$ exoplanetary atmosphere around m-dwarf star. chemical evolution exoplanetary venus-like atmosphere dependent upon ultraviolet near ultraviolet (fuv-nuv) radiation ratio parent star, balance co$_{2}$ photolysis recombination via reactions depend water abundance, various catalytic chemical cycles. study, use three-dimensional (3-d) model simulate conditions venus-like exoplanet orbiting m-dwarf type star gj 436 varying star/planet distance considering resultant effects heating/cooling dynamics. simulation includes middle upper atmosphere (<40 mbar). overall, model comparisons reveal impact extreme ultraviolet ultraviolet (euv-uv) heating energy balance shows radiative dynamical processes responsible driving significant variations zonal winds global temperature profiles < 10$^{-5}$ mbar. specifically, co$_{2}$ 15-${\mu}$m cooling balances euv/uv near infrared (nir) heating altitudes 10$^{-7}$ mbar pressure strong maximum balance pressures ~10$^{-5}$ mbar, thus explaining invariance temperature distribution altitudes 10$^{-5}$mbar pressure cases. model comparisons also show moderate changes nir heating result relatively small changes neutral temperature upper atmosphere, virtually change middle atmosphere. however, larger changes nir heating profile, much greater changes neutral temperature occur entire upper middle atmosphere studied.",0
"boundary-aware instance segmentation. address problem instance-level semantic segmentation, aims jointly detecting, segmenting classifying every individual object image. context, existing methods typically propose candidate objects, usually bounding boxes, directly predict binary mask within proposal. consequence, cannot recover errors object candidate generation process, small shifted boxes. paper, introduce novel object segment representation based distance transform object masks. design object mask network (omn) new residual-deconvolution architecture infers representation decodes final binary object mask. allows us predict masks go beyond scope bounding boxes thus robust inaccurate object candidates. integrate omn multitask network cascade framework, learn resulting boundary-aware instance segmentation (bais) network end-to-end manner. experiments pascal voc 2012 cityscapes datasets demonstrate benefits approach, outperforms state-of-the-art object proposal generation instance segmentation.",3
"evidence algorithm system automated deduction: retrospective view. research project aimed development automated theorem proving system started kiev (ukraine) early 1960s. mastermind project, academician v.glushkov, baptized ""evidence algorithm"", ea. work project lasted, on, 40 years. framework project, russian english versions system automated deduction, sad, constructed. may already seen powerful theorem-proving assistants.",3
"topic modeling using distributed word embeddings. propose new algorithm topic modeling, vec2topic, identifies main topics corpus using semantic information captured via high-dimensional distributed word embeddings. technique unsupervised generates list topics ranked respect importance. find works better existing topic modeling techniques latent dirichlet allocation identifying key topics user-generated content, emails, chats, etc., topics diffused across corpus. also find vec2topic works equally well non-user generated content, papers, reports, etc., small corpora single-document.",3
"many paths equilibrium: gans need decrease divergence every step. generative adversarial networks (gans) family generative models minimize single training criterion. unlike generative models, data distribution learned via game generator (the generative model) discriminator (a teacher providing training signal) minimize cost. gans designed reach nash equilibrium player cannot reduce cost without changing players' parameters. one useful approach theory gans show divergence training distribution model distribution obtains minimum value equilibrium. several recent research directions motivated idea divergence primary guide learning process every step learning decrease divergence. show view overly restrictive. gan training, discriminator provides learning signal situations gradients divergences distributions would useful. provide empirical counterexamples view gan training divergence minimization. specifically, demonstrate gans able learn distributions situations divergence minimization point view predicts would fail. also show gradient penalties motivated divergence minimization perspective equally helpful applied contexts divergence minimization perspective predict would helpful. contributes growing body evidence gan training may usefully viewed approaching nash equilibria via trajectories necessarily minimize specific divergence step.",18
"lower bounds regret reinforcement learning. brief technical note clarify state lower bounds regret reinforcement learning. particular, paper: - reproduces lower bound regret reinforcement learning, similar result theorem 5 journal ucrl2 paper (jaksch et al 2010). - clarifies proposed proof theorem 6 regal paper (bartlett tewari 2009) hold using standard techniques without work. suggest result instead considered conjecture rigorous proof. - suggests conjectured lower bound given (bartlett tewari 2009) incorrect and, fact, possible improve scaling upper bound match weaker lower bounds presented paper. hope note serves clarify existing results field reinforcement learning provides interesting motivation future work.",18
"pseudorehearsal actor-critic agents. catastrophic forgetting serious impact reinforcement learning, data distribution generally sparse non-stationary time. purpose study investigate whether pseudorehearsal increase performance actor-critic agent neural-network based policy selection function approximation pole balancing task compare different pseudorehearsal approaches. expect pseudorehearsal assists learning even simple problems, given proper initialization rehearsal parameters.",3
"exploit imaging opaque wall via deep learning. imaging scattering media encountered many disciplines sciences, ranging biology, mesescopic physics astronomy. still big challenge light suffers multiple scattering media totally decorrelated. here, propose deep-learning-based method retrieve image target behind thick scattering medium. method uses trained deep neural network fit way mapping objects one side thick scattering medium corresponding speckle patterns observed side. demonstration, retrieve images set objects hidden behind 3mm thick white polystyrene slab, optical depth 13.4 times scattering mean free path. work opens new way tackle longstanding challenge using technique deep learning.",3
"estimating nonlinear dynamics convnet smoother. estimating state dynamical system series noise-corrupted observations fundamental many areas science engineering. well-known method, kalman smoother (and related kalman filter), relies assumptions linearity gaussianity rarely met practice. paper, introduced new dynamical smoothing method exploits remarkable capabilities convolutional neural networks approximate complex non-linear functions. main idea generate training set composed latent states observations ensemble simulators train deep network recover former latter. importantly, method requires availability simulators therefore applied situations either latent dynamical model observation model cannot easily expressed closed form. simulation studies, show resulting convnet smoother almost optimal performance gaussian case even parameters unknown. furthermore, method successfully applied extremely non-linear non-gaussian systems. finally, empirically validate approach via analysis measured brain signals.",18
"convolutional feature masking joint object stuff segmentation. topic semantic segmentation witnessed considerable progress due powerful features learned convolutional neural networks (cnns). current leading approaches semantic segmentation exploit shape information extracting cnn features masked image regions. strategy introduces artificial boundaries images may impact quality extracted features. besides, operations raw image domain require compute thousands networks single image, time-consuming. paper, propose exploit shape information via masking convolutional features. proposal segments (e.g., super-pixels) treated masks convolutional feature maps. cnn features segments directly masked maps used train classifiers recognition. propose joint method handle objects ""stuff"" (e.g., grass, sky, water) framework. state-of-the-art results demonstrated benchmarks pascal voc new pascal-context, compelling computational speed.",3
"diabetic retinopathy detection via deep convolutional networks discriminative localization visual explanation. proposed deep learning method interpretable diabetic retinopathy (dr) detection. visual-interpretable feature proposed method achieved adding regression activation map (ram) global averaging pooling layer convolutional networks (cnn). ram, proposed model localize discriminative regions retina image show specific region interest terms severity level. believe advantage proposed deep learning model highly desired dr detection practice, users interested high prediction performance, also keen understand insights dr detection adopted learning model works. experiments conducted large scale retina image dataset, show proposed cnn model achieve high performance dr detection compared state-of-the-art achieving merits providing ram highlight salient regions input image.",3
"identification dynamic functional brain network states tensor decomposition. advances high resolution neuroimaging, growing interest detection functional brain connectivity. complex network theory proposed attractive mathematical representation functional brain networks. however, current studies functional brain networks focused computation graph theoretic indices static networks, i.e. long-time averages connectivity networks. well-known functional connectivity dynamic process construction reorganization networks key understanding human cognition. therefore, growing need track dynamic functional brain networks identify time intervals network quasi-stationary. paper, present tensor decomposition based method identify temporally invariant 'network states' find common topographic representation state. proposed methods applied electroencephalogram (eeg) data study error-related negativity (ern).",3
"local uncertainty sampling large-scale multi-class logistic regression. major challenge building statistical models big data era available data volume may exceed computational capability. common approach solve problem employ subsampled dataset handled available computational resources. paper, propose general subsampling scheme large-scale multi-class logistic regression, examine variance resulting estimator. show asymptotically, proposed method always achieves smaller variance uniform random sampling. moreover, classes conditional imbalanced, significant improvement uniform sampling achieved. empirical performance proposed method compared methods simulated real-world datasets, results match confirm theoretical analysis.",18
"boundedness intrinsic square functions commutators generalized central morrey spaces. paper, authors establish boundedness large class intrinsic square functions $\mathcal{g}_{\alpha}$, $g_{\alpha}$, $g^{\ast}_{\tilde{\lambda},\alpha}$ commutators $[b,\mathcal{g}_{\alpha}]$, $[b,g_{\alpha}]$ $[b,g^{\ast}_{\tilde{\lambda},\alpha}]$ generated $\lambda$-central $bmo$ functions $b\in cbmo^{p,\lambda}(\mathbb{r}^{n})$ generalized central morrey spaces $\mathcal{b}^{q,\varphi}(\mathbb{r}^{n})$ $1<q<\infty,0<\alpha\leq1$, respectively. results new even central morrey spaces $\mathcal{b}^{q,\lambda}(\mathbb{r}^{n})$.",10
"relaxed linearized algorithms faster x-ray ct image reconstruction. statistical image reconstruction (sir) methods studied extensively x-ray computed tomography (ct) due potential acquiring ct scans reduced x-ray dose maintaining image quality. however, longer reconstruction time sir methods hinders use x-ray ct practice. accelerate statistical methods, many optimization techniques investigated. over-relaxation common technique speed convergence iterative algorithms. instance, using relaxation parameter close two alternating direction method multipliers (admm) shown speed convergence significantly. paper proposes relaxed linearized augmented lagrangian (al) method shows theoretical faster convergence rate over-relaxation applies proposed relaxed linearized al method x-ray ct image reconstruction problems. experimental results simulated real ct scan data show proposed relaxed algorithm (with ordered-subsets [os] acceleration) twice fast existing unrelaxed fast algorithms, negligible computation memory overhead.",10
"role harvesting growth rate spatially heterogeneous populations. paper investigates competition two species heterogeneous environment subject effect harvesting. realistic harvesting case connected intrinsic growth rate, harvesting functions developed based clause instead random choice. prove existence uniqueness solution model consider. theoretically, state species coexist, one may drive die out, species extinct, considering possible rational values parameters. results highlight comparative study two harvesting coefficients. finally, solve model using backward-euler, decoupled, linearized time-stepping fully discrete algorithm observe match theoretical numerical findings.",10
"lifestyle vlogs everyday interactions. major stumbling block progress understanding basic human interactions, getting bed opening refrigerator, lack good training data. past efforts gathered data explicitly: starting laundry list action labels, querying search engines videos tagged label. work, reverse search implicitly: start large collection interaction-rich video data annotate analyze it. use internet lifestyle vlogs source surprisingly large diverse interaction data. show collecting data first, able achieve greater scale far greater diversity terms actions actors. additionally, data exposes biases built common explicitly gathered data. make sense data analyzing central component interaction -- hands. benchmark two tasks: identifying semantic object contact video level non-semantic contact state frame level. additionally demonstrate future prediction hands.",3
simple method eliminate blur based lane bates algorithm. simple search method finding blur convolved given image presented. method easily extended large blur. method experimentally tested model blurred image.,3
"identifying mild traumatic brain injury patients mr images using bag visual words. mild traumatic brain injury (mtbi) growing public health problem estimated incidence one million people annually us. neurocognitive tests used assess patient condition monitor patient progress. work aims directly use mr images taken shortly injury detect whether patient suffers mtbi, incorporating machine learning computer vision techniques learn features suitable discriminating mtbi normal patients. focus 3 regions brain, extract multiple patches them, use bag-of-visual-word technique represent subject histogram representative patterns derived patches training subjects. extracting features, use greedy forward feature selection, choose subset features achieves highest accuracy. show experimental studies bow features perform better simple mean value features used previously.",3
"detecting comma-shaped clouds severe weather forecasting using shape motion. meteorologists use shapes movements clouds satellite images indicators several major types severe storms. satellite imaginary data increasingly higher resolution, spatially temporally, making impossible humans fully leverage data forecast. automatic satellite imagery analysis methods find storm-related cloud patterns soon detectable demand. propose machine learning pattern recognition based approach detect ""comma-shaped"" clouds satellite images, specific cloud distribution patterns strongly associated cyclone formulation. order detect regions targeted movement patterns, method trained manually annotated cloud examples represented shape motion-sensitive features. sliding windows different scales used ensure dense clouds captured, implement effective selection rules shrink region interest among sliding windows. finally, evaluate method hold-out annotated comma-shaped cloud dataset cross-match results recorded storm events severe weather database. validated utility accuracy method suggest high potential assisting meteorologists weather forecasting.",3
"improving testsuites via instrumentation. paper explores usefulness technique software engineering, namely code instrumentation, development large-scale natural language grammars. information usage grammar rules test sentences used detect untested rules, redundant test sentences, likely causes overgeneration. results show less half large-coverage grammar german actually tested two large testsuites, 10-30% testing time redundant. methodology applied seen re-use grammar writing knowledge testsuite compilation.",3
"extracting core claims scientific articles. number scientific articles grown rapidly years signs growth slow near future. this, becomes increasingly difficult keep latest developments scientific field. address problem, present approach help researchers learn latest developments findings extracting normalized form core claims scientific articles. normalized representation controlled natural language english sentences called aida, proposed previous work method formally structure organize scientific findings discourse. show aida sentences automatically extracted detecting core claim article, checking aida compliance, - necessary - transforming compliant sentence. algorithm still far perfect, results indicate different steps feasible support claim aida sentences might promising approach improve scientific communication future.",3
"deep matching prior network: toward tighter multi-oriented text detection. detecting incidental scene text challenging task multi-orientation, perspective distortion, variation text size, color scale. retrospective research focused using rectangular bounding box horizontal sliding window localize text, may result redundant background noise, unnecessary overlap even information loss. address issues, propose new convolutional neural networks (cnns) based method, named deep matching prior network (dmpnet), detect text tighter quadrangle. first, use quadrilateral sliding windows several specific intermediate convolutional layers roughly recall text higher overlapping area shared monte-carlo method proposed fast accurate computing polygonal areas. that, designed sequential protocol relative regression exactly predict text compact quadrangle. moreover, auxiliary smooth ln loss also proposed regressing position text, better overall performance l2 loss smooth l1 loss terms robustness stability. effectiveness approach evaluated public word-level, multi-oriented scene text database, icdar 2015 robust reading competition challenge 4 ""incidental scene text localization"". performance method evaluated using f-measure found 70.64%, outperforming existing state-of-the-art method f-measure 63.76%.",3
"individual biases, cultural evolution, statistical nature language universals: case colour naming systems. language universals long attributed innate universal grammar. alternative explanation states linguistic universals emerged independently every language response shared cognitive perceptual biases. computational model recently shown could case, focusing paradigmatic example universal properties colour naming patterns, producing results quantitative agreement experimental data. investigate role individual perceptual bias framework model. study how, extent, structure bias influences corresponding linguistic universal patterns. show cultural history group speakers introduces population-specific constraints act pressure uniformity arising individual bias, clarify interplay two forces.",14
"extracting low-dimensional dynamics multiple large-scale neural population recordings learning predict correlations. powerful approach understanding neural population dynamics extract low-dimensional trajectories population recordings using dimensionality reduction methods. current approaches dimensionality reduction neural data limited single population recordings, identify dynamics embedded across multiple measurements. propose approach extracting low-dimensional dynamics multiple, sequential recordings. algorithm scales data comprising millions observed dimensions, making possible access dynamics distributed across large populations multiple brain areas. building subspace-identification approaches dynamical systems, perform parameter estimation minimizing moment-matching objective using scalable stochastic gradient descent algorithm: model optimized predict temporal covariations across neurons across time. show approach naturally handles missing data multiple partial recordings, identify dynamics predict correlations even presence severe subsampling small overlap recordings. demonstrate effectiveness approach simulated data whole-brain larval zebrafish imaging dataset.",18
"spherical objects dimension two three. paper classifies spherical objects various geometric settings dimension two three, including minimal partial crepant resolutions kleinian singularities, well arbitrary flopping 3-fold contractions gorenstein terminal singularities. main result much general: setting, prove objects associated null category negative ext groups image, action appropriate braid pure braid group, object heart bounded t-structure. corollary objects admit negative exts, self-hom space one dimensional, images simples. variation argument goes further, classifies bounded t-structures. multiple geometric, topological algebraic consequences, primarily autoequivalences stability conditions. main new technique also extends representation theory, establish derived category finite dimensional algebra silting discrete, every object negative ext groups lies heart bounded t-structure. consequence, every semibrick complex completed simple minded collection.",10
"detecting parameter symmetries probabilistic models. probabilistic models often parameters translated, scaled, permuted, otherwise transformed without changing model. symmetries lead strong correlation multimodality posterior distribution model's parameters, pose challenges performing inference interpreting results. work, address automatic detection common problematic model symmetries. so, introduce local symmetries, cover many common cases amenable automatic detection. show derive algorithms detect several broad classes local symmetries. algorithms compatible probabilistic programming constructs arrays, loops, statements, scale models many variables.",18
"view adaptive recurrent neural networks high performance human action recognition skeleton data. skeleton-based human action recognition recently attracted increasing attention due popularity 3d skeleton data. one main challenge lies large view variations captured human actions. propose novel view adaptation scheme automatically regulate observation viewpoints occurrence action. rather re-positioning skeletons based human defined prior criterion, design view adaptive recurrent neural network (rnn) lstm architecture, enables network adapt suitable observation viewpoints end end. extensive experiment analyses show proposed view adaptive rnn model strives (1) transform skeletons various views much consistent viewpoints (2) maintain continuity action rather transforming every frame position body orientation. model achieves significant improvement state-of-the-art approaches three benchmark datasets.",3
"re-initialization free level set evolution via reaction diffusion. paper presents novel reaction-diffusion (rd) method implicit active contours, completely free costly re-initialization procedure level set evolution (lse). diffusion term introduced lse, resulting rd-lse equation, piecewise constant solution derived. order stable numerical solution rd based lse, propose two-step splitting method (tssm) iteratively solve rd-lse equation: first iterating lse equation, solving diffusion equation. second step regularizes level set function obtained first step ensure stability, thus complex costly re-initialization procedure completely eliminated lse. successfully applying diffusion lse, rd-lse model stable means simple finite difference method, easy implement. proposed rd method generalized solve lse variational level set method pde-based level set method. rd-lse method shows good performance boundary anti-leakage, readily extended high dimensional level set method. extensive promising experimental results synthetic real images validate effectiveness proposed rd-lse approach.",3
"tracking deformable parts via dynamic conditional random fields. despite success many advanced tracking methods area, tracking targets drastic variation appearance deformation, view change partial occlusion video sequences still challenge practical applications. letter, take serious tracking problems account simultaneously, proposing dynamic graph based model track object deformable parts multiple resolutions. method introduces well learned structural object detection models object tracking applications prior knowledge deal deformation view change. meanwhile, explicitly formulates partial occlusion integrating spatial potentials temporal potentials unparameterized occlusion handling mechanism dynamic conditional random field framework. empirical results demonstrate method outperforms state-of-the-art trackers different challenging video sequences.",3
fusion qualitative beliefs using dsmt. paper introduces notion qualitative belief assignment model beliefs human experts expressed natural language (with linguistic labels). show qualitative beliefs efficiently combined using extension dezert-smarandache theory (dsmt) plausible paradoxical quantitative reasoning qualitative reasoning. propose new arithmetic linguistic labels allows direct extension classical dsm fusion rule dsm hybrid rules. approximate qualitative pcr5 rule also proposed jointly qualitative average operator. also show crisp interval mappings used deal indirectly linguistic labels. simple example provided illustrate qualitative fusion rules.,3
"algorithmic complexity bounds future prediction errors. bound future loss predicting (computably) stochastic sequence online. solomonoff finitely bounded total deviation universal predictor $m$ true distribution $mu$ algorithmic complexity $mu$. assume time $t>1$ already observed $x=x_1...x_t$. bound future prediction performance $x_{t+1}x_{t+2}...$ new variant algorithmic complexity $mu$ given $x$, plus complexity randomness deficiency $x$. new complexity monotone condition sense complexity decrease condition prolonged. also briefly discuss potential generalizations bayesian model classes classification problems.",3
"jointly learning word embeddings latent topics. word embedding models skip-gram learn vector-space representation word, based local word collocation patterns observed text corpus. latent topic models, hand, take global view, looking word distributions across corpus assign topic word occurrence. two paradigms complementary represent meaning word occurrences. previous works already looked using word embeddings improving quality latent topics, conversely, using latent topics improving word embeddings, ""two-step"" methods cannot capture mutual interaction two paradigms. paper, propose ste, framework learn word embeddings latent topics unified manner. ste naturally obtains topic-specific word embeddings, thus addresses issue polysemy. time, also learns term distributions topics, topic distributions documents. experimental results demonstrate ste model indeed generate useful topic-specific word embeddings coherent latent topics effective efficient way.",3
"aspects univalence holographic axion models. univalent functions complex, analytic (holomorphic) injective functions widely discussed complex analysis. recently proposed stringent constraints univalence imposes growth functions combined sufficient analyticity conditions could used derive rigorous lower upper bounds hydrodynamic dispersion relation, i.e., terms appearing convergent series representations. result exact bounds physical quantities diffusivity speed sound. purpose paper explore ideas, investigate concrete holographic examples, work towards better intuitive understanding role univalence physics. concretely, study diffusive sound modes family holographic axion models offer set observations, arguments tests support applicability univalence methods bounding physical observables described terms effective field theories. work provides insight expected `typical' regions univalence, comparisons tightness bounds corresponding exact values certain quantities characterising transport, tests relations diffusion bounds involve chaotic pole-skipping, well tests condition implies conformal bound speed sound complementary condition implies violation.",9
"supernova dust evolution probed deep-sea 60fe time history. wealth data live, undecayed 60fe ($t_{1/2} = 2.6 \ \rm myr$) deep-sea deposits, lunar regolith, cosmic rays, antarctic snow, interpreted originating recent explosions least two near-earth supernovae. use 60fe profiles deep-sea sediments estimate timescale supernova debris deposition beginning $\sim 3$ myr ago. available data admits variety different profile functions, cases best-fit 60fe pulse durations $> 1.6$ myr data combined. timescale far exceeds $\lesssim 0.1$ myr pulse would expected 60fe entrained supernova blast wave plasma. interpret long signal duration evidence 60fe arrives form supernova dust, whose dynamics separate coupled evolution blast plasma. framework, $> 1.6$ myr timescale dust stopping due drag forces. scenario consistent simulations fry et. al (2020), magnetically trapped supernova remnants thereby confined near regions remnant dominated supernova ejecta, magnetic fields low. picture fits naturally models cosmic-ray injection refractory elements sputtered supernova dust grains, implies recent 60fe detections cosmic rays complement fragments grains survived arrive earth moon. finally, present possible tests scenario.",0
"feed-forward approximations dynamic recurrent network architectures. recurrent neural network architectures useful computational properties, complex temporal dynamics input-sensitive attractor states. however, evaluation recurrent dynamic architectures requires solution systems differential equations, number evaluations required determine response given input vary input, indeterminate altogether case oscillations instability. feed-forward networks, contrast, single pass network needed determine response given input. modern machine-learning systems designed operate efficiently feed-forward architectures. hypothesised two-layer feedforward architectures simple, deterministic dynamics could approximate responses single-layer recurrent network architectures. identifying fixed-point responses given recurrent network, trained two-layer networks directly approximate fixed-point response given input. feed-forward networks embodied useful computations, including competitive interactions, information transformations noise rejection. approach able find useful approximations recurrent networks, evaluated linear deterministic time complexity.",3
"towards optimal learning chain graphs. paper, extend meek's conjecture (meek 1997) directed acyclic graphs chain graphs, prove extended conjecture true. specifically, prove chain graph h independence map independence model induced another chain graph g, (i) g transformed h sequence directed undirected edge additions feasible splits mergings, (ii) operation sequence h remains independence map independence model induced g. result important consequence learning chain graphs data proof meek's conjecture (chickering 2002) learning bayesian networks data: makes possible develop efficient asymptotically correct learning algorithms mild assumptions.",18
"surpassing human-level face verification performance lfw gaussianface. face verification remains challenging problem complex conditions large variations pose, illumination, expression, occlusions. problem exacerbated rely unrealistically single training data source, often insufficient cover intrinsically complex face variations. paper proposes principled multi-task learning approach based discriminative gaussian process latent variable model, named gaussianface, enrich diversity training data. comparison existing methods, model exploits additional data multiple source-domains improve generalization performance face verification unknown target-domain. importantly, model adapt automatically complex data distributions, therefore well capture complex face variations inherent multiple sources. extensive experiments demonstrate effectiveness proposed model learning diverse data sources generalize unseen domain. specifically, accuracy algorithm achieves impressive accuracy rate 98.52% well-known challenging labeled faces wild (lfw) benchmark. first time, human-level performance face verification (97.53%) lfw surpassed.",3
"idnet: smartphone-based gait recognition convolutional neural networks. here, present idnet, user authentication framework smartphone-acquired motion signals. goal recognize target user way walking, using accelerometer gyroscope (inertial) signals provided commercial smartphone worn front pocket user's trousers. idnet features several innovations including: i) robust smartphone-orientation-independent walking cycle extraction block, ii) novel feature extractor based convolutional neural networks, iii) one-class support vector machine classify walking cycles, coherent integration iv) multi-stage authentication technique. idnet first system exploits deep learning approach universal feature extractors gait recognition, combines classification results subsequent walking cycles multi-stage decision making framework. experimental results show superiority approach state-of-the-art techniques, leading misclassification rates (either false negatives positives) smaller 0.15% fewer five walking cycles. design choices discussed motivated throughout, assessing impact user authentication performance.",3
"sentence based semantic similarity measure blog-posts. blogs-online digital diary like application web 2.0 opened new easy way voice opinion, thoughts, like-dislike every internet user world. blogosphere doubt largest user-generated content repository full knowledge. potential knowledge still explored. knowledge discovery new genre quite difficult challenging totally different popular genre web-applications like world wide web (www). blog-posts unlike web documents small size, thus lack context contain relaxed grammatical structures. hence, standard text similarity measure fails provide good results. paper, specialized requirements comparing pair blog-posts thoroughly investigated. based proposed novel algorithm sentence oriented semantic similarity measure pair blog-posts. applied algorithm subset political blogosphere pakistan, cluster blogs different issues political realm identify influential bloggers.",3
"online learning classification low-rank representation features applications audio segment classification. paper, novel framework based trace norm minimization audio segment proposed. framework, feature extraction classification obtained solving corresponding convex optimization problem trace norm regularization. feature extraction, robust principle component analysis (robust pca) via minimization combination nuclear norm $\ell_1$-norm used extract low-rank features robust white noise gross corruption audio segments. low-rank features fed linear classifier weight bias learned solving similar trace norm constrained problems. classifier, methods find weight bias batch-mode learning, makes inefficient large-scale problems. paper, propose online framework using accelerated proximal gradient method. framework main advantage memory cost. addition, result regularization formulation matrix classification, lipschitz constant given explicitly, hence step size estimation general proximal gradient method omitted approach. experiments real data sets laugh/non-laugh applause/non-applause classification indicate novel framework effective noise robust.",3
"language use matters: analysis linguistic structure question texts characterize answerability quora. quora one popular community q&a sites recent times. however, many question posts q&a site often get answered. paper, quantify various linguistic activities discriminates answered question unanswered one. central finding way users use language writing question text effective means characterize answerability. characterization helps us predict early question remaining unanswered specific time period eventually answered achieve accuracy 76.26% (t = 1 month) 68.33% (t = 3 months). notably, features representing language use patterns users discriminative alone account accuracy 74.18%. also compare method similar works (dror et al., yang et al.) achieving maximum improvement ~39% terms accuracy.",3
"weingarten flows riemannian manifolds. given orientable riemannian manifolds $m^n$ $\bar m^{n+1},$ study flows $f_t:m^n\rightarrow\bar m^{n+1},$ called weingarten flows,in hypersurfaces $f_t(m)$ evolve direction normal vectors speed given function $w$ principal curvatures,called weingarten function, homogeneous, monotonic increasing respect variables, positive positive cone. obtain existence results flows isoparametric initial data, hypersurfaces $f_t:m^n\rightarrow\bar m^{n+1}$ parallel, $\bar m^{n+1}$ either simply connected space form rank-one symmetric space noncompact type. prove avoidance principle holds weingarten flows defined odd weingarten functions, also flows embedding preserving.",10
"multi-batch l-bfgs method machine learning. question parallelize stochastic gradient descent (sgd) method received much attention literature. paper, focus instead batch methods use sizeable fraction training set iteration facilitate parallelism, employ second-order information. order improve learning process, follow multi-batch approach batch changes iteration. cause difficulties l-bfgs employs gradient differences update hessian approximations, gradients computed using different data points process unstable. paper shows perform stable quasi-newton updating multi-batch setting, illustrates behavior algorithm distributed computing platform, studies convergence properties convex nonconvex cases.",10
"mining combined causes large data sets. recent years, many methods developed detecting causal relationships observational data. potential tackle large data sets. however, methods fail discover combined cause, i.e. multi-factor cause consisting two component variables individually causes. straightforward approach uncovering combined cause include individual combined variables causal discovery using existing methods, scheme computationally infeasible due huge number combined variables. paper, propose novel approach address practical causal discovery problem, i.e. mining combined causes large data sets. experiments synthetic real world data sets show proposed method obtain high-quality causal discoveries high computational efficiency.",3
"real-time end-to-end action detection two-stream networks. two-stream networks successful solving problem action detection. however, prior work using two-stream networks train streams separately, prevents network exploiting regularities two streams. moreover, unlike visual stream, dominant forms optical flow computation typically maximally exploit gpu parallelism. present real-time end-to-end trainable two-stream network action detection. first, integrate optical flow computation framework using flownet2. second, apply early fusion two streams train whole pipeline jointly end-to-end. finally, better network initialization, transfer task action recognition action detection pre-training framework using recently released large-scale kinetics dataset. experimental results show training pipeline jointly end-to-end fine-tuning optical flow objective action detection improves detection performance significantly. additionally, observe improvement initializing parameters pre-trained using kinetics. last, show integrating optical flow computation, framework efficient, running real-time speeds (up 31 fps).",3
"regularized partial least squares application nmr spectroscopy. high-dimensional data common genomics, proteomics, chemometrics often contains complicated correlation structures. recently, partial least squares (pls) sparse pls methods gained attention areas dimension reduction techniques context supervised data analysis. introduce framework regularized pls solving relaxation simpls optimization problem penalties pls loadings vectors. approach enjoys many advantages including flexibility, general penalties, easy interpretation results, fast computation high-dimensional settings. also outline extensions methods leading novel methods non-negative pls generalized pls, adaption pls structured data. demonstrate utility methods simulations case study proton nuclear magnetic resonance (nmr) spectroscopy data.",18
"embedrank: unsupervised keyphrase extraction using sentence embeddings. keyphrase extraction task automatically selecting small set phrases best describe given free text document. keyphrases used indexing, searching, aggregating summarizing text documents, serving many automatic well human-facing use cases. existing supervised systems keyphrase extraction require large amounts labeled training data generalize poorly outside domain training data. time, unsupervised systems found literature poor accuracy, often generalize well, require input document belong larger corpus also given input. furthermore, supervised unsupervised methods often slow real-time scenarios suffer over-generation. addressing drawbacks, paper, introduce unsupervised method keyphrase extraction single documents leverages sentence embeddings. selecting phrases whose semantic embeddings close embeddings whole document, able separate best candidate phrases rest. show embedding-based method simpler, also effective graph-based state art systems, achieving higher f-scores standard datasets. simplicity significant advantage, especially processing large amounts documents web, resulting considerable speed gains. moreover, describe increase coverage diversity among selected keyphrases introducing embedding-based maximal marginal relevance (mmr) new phrases. user study including 200 votes showed that, although reducing phrase semantic overlap leads gains terms f-score, diversity enriched selection preferred humans.",3
"icdar2017 competition reading chinese text wild (rctw-17). chinese widely used language world. algorithms read chinese text natural images facilitate applications various kinds. despite large potential value, datasets competitions past primarily focus english, bares different characteristics chinese. report introduces rctw, new competition focuses chinese text reading. competition features large-scale dataset 12,000 annotated images. two tasks, namely text localization end-to-end recognition, set up. competition took place january 20 may 31, 2017. 23 valid submissions received 19 teams. report includes dataset description, task definitions, evaluation protocols, results summaries analysis. competition, call future research chinese text reading problem.",3
"clipped action policy gradient. many continuous control tasks bounded action spaces clip out-of-bound actions execution. policy gradient methods often optimize policies actions clipped. propose clipped action policy gradient (capg) alternative policy gradient estimator exploits knowledge actions clipped reduce variance estimation. prove capg unbiased achieves lower variance original estimator ignores action bounds. experimental results demonstrate capg generally outperforms original estimator, indicating promise better policy gradient estimator continuous control tasks.",3
"curvature integration 5d kernel extracting vessel connections retinal images. tree-like structures retinal images widely studied computer-aided diagnosis systems large-scale screening programs. despite several segmentation tracking methods proposed literature, still exist several limitations specifically two curvilinear structures cross bifurcate, presence interrupted lines highly curved blood vessels. paper, propose novel approach based multi-orientation scores augmented contextual affinity matrix, inspired geometry primary visual cortex (v1) contextual connections. connectivity described five-dimensional kernel obtained fundamental solution fokker-planck equation modelling cortical connectivity lifted space positions, orientations, curvatures intensity. used self-tuning spectral clustering step identify main perceptual units stimuli. proposed method validated several easy challenging structures set artificial images actual retinal patches. supported quantitative qualitative results, method capable overcoming limitations current state-of-the-art techniques.",3
"nuclear-electronic orbital approach quantization protons periodic electronic structure calculations. nuclear-electronic orbital (neo) method well-established approach treating nuclei quantum mechanically molecular systems beyond usual born-oppenheimer approximation. work, present strategy implement neo method periodic electronic structure calculations, particularly focused multicomponent density functional theory (dft). neo-dft method implemented all-electron electronic structure code, fhi-aims, using combination analytical numerical integration techniques well resolution identity scheme enhance computational efficiency. validating implementation, proof-of-concept applications presented illustrate effects quantized protons physical properties extended systems two-dimensional materials liquid-semiconductor interfaces. specifically, periodic neo-dft calculations performed trans-polyacetylene chain, hydrogen boride sheet, titanium oxide-water interface. zero-point energy effects protons, well electron-proton correlation, shown noticeably impact density states band structures systems. developments provide foundation application multicomponent dft wide range extended condensed matter systems.",14
"chaotic multi-objective optimization based design fractional order piλdμ controller avr system. paper, fractional order (fo) pi{\lambda}d\mu controller designed take care various contradictory objective functions automatic voltage regulator (avr) system. improved evolutionary non-dominated sorting genetic algorithm ii (nsga ii), augmented chaotic map greater effectiveness, used multi-objective optimization problem. pareto fronts showing trade-off different design criteria obtained pi{\lambda}d\mu pid controller. comparative analysis done respect standard pid controller demonstrate merits demerits fractional order pi{\lambda}d\mu controller.",3
"projection-free online optimization stochastic gradient: convexity submodularity. online optimization successful framework solving large-scale problems computational constraints partial information. current methods online convex optimization require either projection exact gradient computation step, prohibitively expensive large-scale applications. time, growing trend non-convex optimization machine learning community need online methods. continuous submodular functions, exhibit natural diminishing returns condition, recently proposed broad class non-convex functions may efficiently optimized. although online methods introduced, suffer similar problems. work, propose meta-frank-wolfe, first online projectionfree algorithm uses stochastic gradient estimates. algorithm relies careful sampling gradients round achieves optimal $o(\sqrt{t})$ adversarial regret bounds convex continuous submodular optimization. also propose one-shot frank-wolfe, simpler algorithm requires single stochastic gradient estimate round achieves $o(t^{2/3})$ stochastic regret bound convex continuous submodular optimization. apply methods develop novel ""lifting"" framework online discrete submodular maximization also see outperform current state art techniques extensive set experiments.",18
"extended object tracking: introduction, overview applications. article provides elaborate overview current research extended object tracking. provide clear definition extended object tracking problem discuss delimitation types object tracking. next, different aspects extended object modelling extensively discussed. subsequently, give tutorial introduction two basic well used extended object tracking approaches - random matrix approach kalman filter-based approach star-convex shapes. next part treats tracking multiple extended objects elaborates large number feasible association hypotheses tackled using random finite set (rfs) non-rfs multi-object trackers. article concludes summary current applications, four example applications involving camera, x-band radar, light detection ranging (lidar), red-green-blue-depth (rgb-d) sensors highlighted.",3
"local langlands correspondences. first part article review properties expected local langlands correspondence aims considered ""canonical,"" known results establish properties specific groups. absence compatibility global correspondence known general list desirable properties suffices characterize correspondence. remainder article outlines elements strategy prove $l$-packets attached irreducible local parameters finite non-empty, particular, specific parametrization constructed genestier lafforgue $f$ positive characteristic.",10
"generating descriptions grounded co-referenced people. learning generate descriptions images videos received major interest computer vision natural language processing communities. works proposed learn grounding generation process unsupervised way (via attention mechanism), remains unclear good quality grounding whether benefits description quality. work propose movie description model learns generate description jointly ground (localize) mentioned characters well visual co-reference resolution pairs consecutive sentences/clips. also propose use weak localization supervision character mentions provided movie descriptions learn character grounding. training time, first learn localize characters relating visual appearance mentions descriptions via semi-supervised approach. provide (noisy) supervision description model greatly improves performance. proposed description model improves prior work w.r.t. generated description quality additionally provides grounding local co-reference resolution. evaluate mpii movie description dataset using automatic human evaluation measures using newly collected grounding co-reference data characters.",3
"image segmentation adaptive distance based em algorithm. paper introduces bayesian image segmentation algorithm based finite mixtures. em algorithm developed estimate parameters gaussian mixtures. finite mixture flexible powerful probabilistic modeling tool. used provide model-based clustering field pattern recognition. however, application finite mixtures image segmentation presents difficulties; especially sensible noise. paper propose variant method aims resolve problem. approach proceeds characterization pixels two features: first one describes intrinsic properties pixel second characterizes neighborhood pixel. classification made base adaptive distance privileges one features according spatial position pixel image. obtained results shown significant improvement approach compared standard version em algorithm.",3
"image denoising using new adaptive based median filters. noise major issue transferring images kinds electronic communication. one common noise electronic communication impulse noise caused unstable voltage. paper, comparison known image denoising techniques discussed new technique using decision based approach used removal impulse noise. methods primarily preserve image details suppressing impulsive noise. principle techniques first introduced analysed various simulation results using matlab. previously known techniques applicable denoising images corrupted less noise density. new decision based technique presented shows better performances already used. comparisons made based visual appreciation quantitatively mean square error (mse) peak signal noise ratio (psnr) different filtered images..",3
"recovering block-structured activations using compressive measurements. consider problems detection localization contiguous block weak activation large matrix, small number noisy, possibly adaptive, compressive (linear) measurements. closely related problem compressed sensing, task estimate sparse vector using small number linear measurements. contrary results compressed sensing, shown neither adaptivity contiguous structure help much, show reliable localization magnitude weakest signals strongly influenced structure ability choose measurements adaptively detection neither adaptivity structure reduce requirement magnitude signal. characterize precise tradeoffs various problem parameters, signal strength number measurements required reliably detect localize block activation. sufficient conditions complemented information theoretic lower bounds.",18
"cobb angle measurement scoliosis reduced variability. cobb angle, measure spinal curvature standard method quantifying magnitude scoliosis related spinal deformity orthopedics. determining cobb angle manual process subject human errors. work, propose methodology measure magnitude cobb angle, appreciably reduces variability related measurement compared related works. proposed methodology facilitated using suitable new improved version non-local means image denoisation otsus automatic threshold selection canny edge detection. selected nlm preprocessing image one fine states art image denoisation helps retaining image quality. trimmedmean, median robust outliners mean following concept observed nlm denoising quality performance enhanced using euclidean trimmed-mean replacing mean. prove better performance non-local euclidean trimmed-mean denoising filter, provided comparative study results proposed denoising technique traditional nlm nonlocal euclidean medians. experimental results cobb angle measurement intra observer inter observer experimental data reveals better performance superiority proposed approach compared related works. matlab2009b image processing toolbox used purpose simulation verification proposed methodology.",3
"efficient, expressive local minima-free method learning controlled dynamical systems. propose framework modeling estimating state controlled dynamical systems, agent affect system actions receives partial observations. based framework, propose predictive state representation random fourier features (rffpsr). key property rff-psrs state estimate represented conditional distribution future observations given future actions. rff-psrs combine representation moment-matching, kernel embedding local optimization achieve method enjoys several favorable qualities: represent controlled environments affected actions; efficient theoretically justified learning algorithm; uses non-parametric representation expressive power represent continuous non-linear dynamics. provide detailed formulation, theoretical analysis experimental evaluation demonstrates effectiveness method.",18
"decentralized learning wireless communications networking. chapter deals decentralized learning algorithms in-network processing graph-valued data. generic learning problem formulated recast separable form, iteratively minimized using alternating-direction method multipliers (admm) gain desired degree parallelization. without exchanging elements distributed training sets keeping inter-node communications affordable levels, local (per-node) learners consent desired quantity inferred globally, meaning one obtained entire training data set centrally available. impact decentralized learning framework contemporary wireless communications networking tasks illustrated case studies including target tracking using wireless sensor networks, unveiling internet traffic anomalies, power system state estimation, well spectrum cartography wireless cognitive radio networks.",10
"invariant von neumann subalgebras rigidity property. say countable discrete group $\gamma$ satisfies invariant von neumann subalgebras rigidity (isr) property every $\gamma$- invariant von neumann subalgebra $\mathcal{m}$ $l(\gamma)$ form $l(\lambda)$ normal subgroup $\lambda\lhd \gamma$. show many ``negatively curved"" groups, including torsion free non-amenable hyperbolic groups torsion free groups positive first $l^2$-betti number mild assumption, certain finite direct product property. also discuss whether torsion-free assumption relaxed.",10
"face parsing via recurrent propagation. face parsing important problem computer vision finds numerous applications including recognition editing. recently, deep convolutional neural networks (cnns) applied image parsing segmentation state-of-the-art performance. paper, propose face parsing algorithm combines hierarchical representations learned cnn, accurate label propagations achieved spatially variant recurrent neural network (rnn). rnn-based propagation approach enables efficient inference global space guidance semantic edges generated local convolutional model. since convolutional architecture shallow spatial rnn parameters, framework much faster light-weighted state-of-the-art cnns task. apply proposed model coarse-grained fine-grained face parsing. fine-grained face parsing, develop two-stage approach first identifying main regions segmenting detail components, achieves better performance terms accuracy efficiency. single gpu, proposed algorithm parses face images accurately 300 frames per second, facilitates real-time applications.",3
"solution `$1/2$ vs $3/2$' puzzle. using almost complete relativistic method based bethe-salpeter equation, study mixing angle $\theta$, mass splitting $\bigtriangleup m$, strong decay widths $\gamma(d^{({\prime})}_1)$ weak production rates $br(b\to d^{({\prime})}_1\ell\nu_{\ell})$ $d_1(2420)$ $d_1^{\prime}(2430)$. find strong cancellation $^1p_1$ $^3p_1$ partial waves $d_1^{\prime}(2430)$ $\theta \sim-35.3^{\circ}$, leads `$1/2$ vs $3/2$' puzzle. puzzle overcome adding relativistic corrections since large parameter range $\bigtriangleup m$ linear varying small, $\theta$, $\gamma(d^{({\prime})}_1)$ $br(b\to d^{({\prime})}_1\ell\nu_{\ell})$ remain almost unchanged conflict data. special range around mass inverse point $\bigtriangleup m=0$ $\theta =\pm 90^{\circ}$, change rapidly find windows $\bigtriangleup m$, $\gamma(d^{({\prime})}_1)$ $br(b\to d^{({\prime})}_1\ell\nu_{\ell})$ consistent data. small $\bigtriangleup m$ confirmed experiment, crucial solve `$1/2$ vs $3/2$' puzzle.",8
"neutron-induced nuclear recoil background pandax-4t experiment. neutron-induced nuclear recoil background critical dark matter searches pandax-4t liquid xenon experiment. paper studies feature neutron background liquid xenon evaluates contribution single scattering nuclear recoil events three methods. first method fully monte carlo simulation based. last two data-driven methods also use multiple scattering signals high energy signals data, respectively. pandax-4t commissioning data exposure 0.63 tonne-year, methods give consistent result $1.15\pm0.57$ neutron-induced background dark matter signal region within approximated nuclear recoil energy window 5 100 kev.",7
"sequential algorithm training text classifiers. ability cheaply train text classifiers critical use information retrieval, content analysis, natural language processing, tasks involving data partly fully textual. algorithm sequential sampling machine learning statistical classifiers developed tested newswire text categorization task. method, call uncertainty sampling, reduced much 500-fold amount training data would manually classified achieve given level effectiveness.",1
"dark energy survey supernova program results: type ia supernova brightness correlates host galaxy dust. cosmological analyses type ia supernovae (sne ia) often assume single empirical relation color luminosity ($\beta$) account varying host-galaxy dust properties. however, studies dust large samples galaxies, known dust attenuation vary significantly. take advantage state-of-the-art modeling galaxy properties characterize dust parameters (dust attenuation $a_v$, parameter describing dust law slope $r_v$) dark energy survey (des) sn ia host galaxies using publicly available \texttt{bagpipes} code. utilizing optical infrared data hosts alone, find three key aspects host dust impact sn ia cosmology: 1) exists large range ($\sim1-6$) host $r_v$ 2) high stellar mass hosts $r_v$ average $\sim0.7$ lower low-mass hosts 3) significant ($>3\sigma$) correlation hubble diagram residuals red sne ia corrected reduces scatter $\sim13\%$ significance ``mass step'' $\sim1\sigma$. represent independent confirmations recent predictions based dust attempted explain puzzling ``mass step'' intrinsic scatter ($\sigma_{\rm int}$) sn ia analyses. also find red-sequence galaxies lower peaked dust law slope distributions average comparison non red-sequence galaxies. find sn ia $\beta$ $\sigma_{\rm int}$ differ $>3\sigma$ determined separately red-sequence galaxy galaxy hosts. agreement fitted host-$r_v$ sn ia $\beta$ \& $\sigma_{\rm int}$ suggests host dust properties play major role sn ia color-luminosity standardization supports claim sn ia intrinsic scatter driven $r_v$ variation.",0
complete characterization statistical query learning applications evolvability. statistical query (sq) learning model kearns (1993) natural restriction pac learning model learning algorithm allowed obtain estimates statistical properties examples cannot see examples themselves. describe new simple characterization query complexity learning sq learning model. unlike previously known bounds sq learning characterization preserves accuracy efficiency learning. preservation accuracy implies characterization gives first characterization sq learning agnostic learning framework. preservation efficiency achieved using new boosting technique allows us derive new approach design evolutionary algorithms valiant's (2006) model evolvability. use approach demonstrate existence large class monotone evolutionary learning algorithms based square loss performance estimation. results differ significantly known evolutionary algorithms give evidence evolvability valiant's model versatile phenomenon previous reason suspect.,3
"advancing memristive analog neuromorphic networks: increasing complexity, coping imperfect hardware components. experimentally demonstrate classification 4x4 binary images 4 classes, using 3-layer mixed-signal neuromorphic network (""mlp perceptron""), based two passive 20x20 memristive crossbar arrays, board-integrated discrete cmos components. network features 10 hidden-layer 4 output-layer analog cmos neurons 428 metal-oxide memristors, i.e. almost order magnitude complex previously reported functional memristor circuit. moreover, inference operation classifier performed entirely integrated hardware. deal larger crossbar arrays, developed semi-automatic approach forming testing, compared several memristor training schemes coping imperfect behavior devices, well variability analog cmos neurons. effectiveness proposed schemes defect variation tolerance verified experimentally using implemented network and, additionally, modeling operation larger network, 300 hidden-layer neurons, mnist benchmark. finally, propose simple modification implemented memristor-based vector-by-matrix multiplier allow operation wider temperature range.",3
"state art urban reconstruction: street, street network, vegetation, urban feature. world population raising, especially part people living cities. increased population complex roles regarding inhabitants surroundings, cities concentrate difficulties design, planning analysis. tasks require way reconstruct/model city. traditionally, much attention given buildings reconstruction, yet essential part city neglected: streets. streets reconstruction seldom researched. streets also complex compositions urban features, unique role transportation (as comprise roads). aim completing recent state art building reconstruction (musialski2012) considering aspect urban reconstruction. introduce need city models. reconstruction always necessitates data, first analyse data available. expose state art street reconstruction, street network reconstruction, urban features reconstruction/modelling, vegetation , urban objects reconstruction/modelling. although reconstruction strategies vary widely, order role model plays, data driven approach, model-based approach, inverse procedural modelling model catalogue matching. main challenges seems come complex nature urban environment limitations available data. urban features strong relationships, them, surrounding, well hierarchical relations. procedural modelling power express relations, could applied reconstruction urban features via inverse procedural modelling paradigm.",3
"fast nonlinear embeddings via structured matrices. present new paradigm speeding randomized computations several frequently used functions machine learning. particular, paradigm applied improving computations kernels based random embeddings. that, presented framework covers multivariate randomized functions. byproduct, propose algorithmic approach also leads significant reduction space complexity. method based careful recycling gaussian vectors structured matrices share properties fully random matrices. quality proposed structured approach follows combinatorial properties graphs encoding correlations rows structured matrices. framework covers special cases already known structured approaches fast johnson-lindenstrauss transform, much general since applied also highly nonlinear embeddings. provide strong concentration results showing quality presented paradigm.",18
"towards future directions data-integrative supervised prediction human aging-related genes. identification human genes involved aging process critical due incidence many diseases age. state-of-the-art approach purpose infers weighted dynamic aging-specific subnetwork mapping gene expression (ge) levels different ages onto protein-protein interaction network (ppin). then, analyzes subnetwork supervised manner training predictive model learn network topologies known aging- vs. non-aging-related genes change across ages. finally, uses trained model predict novel aging-related genes. however, best current subnetwork resulting approach still yields suboptimal prediction accuracy. could inferred using outdated ge ppin data. here, evaluate whether analyzing weighted dynamic aging-specific subnetwork inferred newer ge ppin data improves prediction accuracy upon analyzing best current subnetwork inferred outdated data. unexpectedly, find case. understand this, perform aging-related pathway gene ontology (go) term enrichment analyses. find suboptimal prediction accuracy, regardless ge ppin data used, may caused current knowledge genes aging-related incomplete, current methods inferring analyzing aging-specific subnetwork unable capture aging-related knowledge. findings potentially guide future directions towards improving supervised prediction aging-related genes via -omics data integration.",15
"r3mc: riemannian three-factor algorithm low-rank matrix completion. exploit versatile framework riemannian optimization quotient manifolds develop r3mc, nonlinear conjugate-gradient method low-rank matrix completion. underlying search space fixed-rank matrices endowed novel riemannian metric tailored least-squares cost. numerical comparisons suggest r3mc robustly outperforms state-of-the-art algorithms across different problem instances, especially combine scarcely sampled ill-conditioned data.",10
"spatially constrained location prior scene parsing. semantic context important useful cue scene parsing complicated natural images substantial amount variations objects environment. paper proposes spatially constrained location prior (sclp) effective modelling global local semantic context scene terms inter-class spatial relationships. unlike existing studies focusing either relative absolute location prior objects, sclp effectively incorporates relative absolute location priors calculating object co-occurrence frequencies spatially constrained image blocks. sclp general used conjunction various visual feature-based prediction models, artificial neural networks support vector machine (svm), enforce spatial contextual constraints class labels. using svm classifiers linear regression model, demonstrate incorporation sclp achieves superior performance compared state-of-the-art methods stanford background sift flow datasets.",3
"policy optimization genetic distillation. genetic algorithms widely used many practical optimization problems. inspired natural selection, operators, including mutation, crossover selection, provide effective heuristics search black-box optimization. however, shown useful deep reinforcement learning, possibly due catastrophic consequence parameter crossovers neural networks. here, present genetic policy optimization (gpo), new genetic algorithm sample-efficient deep policy optimization. gpo uses imitation learning policy crossover state space applies policy gradient methods mutation. experiments mujoco tasks show gpo genetic algorithm able provide superior performance state-of-the-art policy gradient methods achieves comparable higher sample efficiency.",18
"completing $c_2$ completion conjecture $p=2$. $c_2$-invariant arithmetic graph invariant useful understanding feynman periods. brown schnetz conjectured $c_2$-invariant particular symmetry known completion invariance. paper prove completion invariance $c_2$-invariant $p=2$ case, extending previous work one us. methods combinatorial enumerative involving counting certain partitions edges graph.",10
"finding structure text, genome symbolic sequences. statistical methods derived described thesis provide new ways elucidate structural properties text symbolic sequences. generically, methods allow detection difference frequency single feature, detection difference frequencies ensemble features attribution source text. three abstract tasks suffice solve problems wide variety settings. furthermore, techniques described thesis extended provide wide range additional tests beyond ones described here. variety applications methods examined detail. applications drawn area text analysis genetic sequence analysis. textually oriented tasks include finding interesting collocations cooccurent phrases, language identification, information retrieval. biologically oriented tasks include species identification discovery previously unreported long range structure genes. applications reported direct comparison possible, performance new methods substantially exceeds state art. overall, methods described provide new effective ways analyse text symbolic sequences. particular strength deal well situations relatively little data available. since methods abstract nature, applied novel situations relative ease.",3
"self-similarity universal property soap film bubble roll-off stage. children enjoy blowing soap bubbles also show bath wash dishes. analyze thinning breaking soap bubble neck stretched. contrast open film widely studied, concentrate effect volume conservation bubble. like film (f), non-equilibrium state divided three stages bubble (b): (1) roll-off, (2) cusp approach, (3) pinch-off. manage establish existence two major properties soap membrane: self-similarity f-1, b-1, b-3, b-2, universal behavior profile f-1 b-1. based discoveries, show process memory erasing starts earlier stage 3. also study amount inflated air pulling speed affect equally interesting properties soap bubble necking, whether evolution profile minimum radius bubble obey scaling laws.",2
"latent structure dictionaries. many words (and ones) sufficient define words? dictionaries analyzed directed graphs links defining words defined words, reveal latent structure. recursively removing words reachable definition define words reduces dictionary kernel 10%. still smallest number words define rest. 75% kernel turns core, strongly connected subset words definitional path pair words word's definition depending word outside set. core cannot define rest dictionary. 25% kernel surrounding core consists small strongly connected subsets words: satellites. size smallest set words define rest (the graph's minimum feedback vertex set minset) 1% dictionary, 15% kernel, half-core, half-satellite. every dictionary huge number minsets. core words learned earlier, frequent, less concrete satellites, turn learned earlier frequent concrete rest dictionary. principle, one minset's words would need grounded sensorimotor capacity recognize categorize referents. dual-code sensorimotor-symbolic model mental lexicon, symbolic code could rest via re-combinatory definition.",3
"securing transactions: detecting anomalous patterns xml documents. xml transactions used many information systems store data interact systems. abnormal transactions, result either on-going cyber attack actions benign user, potentially harm interacting systems therefore regarded threat. paper address problem anomaly detection localization xml transactions using machine learning techniques. present new xml anomaly detection framework, xml-ad. within framework, automatic method extracting features xml transactions developed well practical method transforming xml features vectors fixed dimensionality. two methods place, xml-ad framework makes possible utilize general learning algorithms anomaly detection. central functioning framework novel multi-univariate anomaly detection algorithm, adifa. framework evaluated four xml transactions datasets, captured real information systems, achieved 89% true positive detection rate less 0.2% false positive rate.",3
"$h^1$-stability l2-type method general nonuniform meshes subdiffusion equation. work $h^1$-stability l2 method general nonuniform meshes established subdiffusion equation. mild constraints time step ratio $\rho_k$, example $0.4573328\leq \rho_k\leq 3.5615528$ $k\geq 2$, crucial bilinear form associated l2 fractional-derivative operator proved positive semidefinite. consequence, $h^1$-stability l2 schemes derived subdiffusion equation. special case graded mesh, positive semidefiniteness holds grading parameter $1<r\leq 3.2016538$ therefore $h^1$-stability l2 schemes holds. best knowledge, first work $h^1$-stability l2 method general nonuniform meshes subdiffusion equation.",10
"evaluation dense 3d reconstruction 2d face images wild. paper investigates evaluation dense 3d face reconstruction single 2d image wild. end, organise competition provides new benchmark dataset contains 2000 2d facial images 135 subjects well 3d ground truth face scans. contrast previous competitions challenges, aim new benchmark dataset evaluate accuracy 3d dense face reconstruction algorithm using real, accurate high-resolution 3d ground truth face scans. addition dataset, provide standard protocol well python script evaluation. last, report results obtained three state-of-the-art 3d face reconstruction systems new benchmark dataset. competition organised along 2018 13th ieee conference automatic face & gesture recognition.",3
optimistic semi-supervised least squares classification. goal semi-supervised learning improve supervised classifiers using additional unlabeled training examples. work study simple self-learning approach semi-supervised learning applied least squares classifier. show soft-label hard-label variant self-learning derived applying block coordinate descent two related slightly different objective functions. resulting soft-label approach related idea dealing missing data dates back 1930s. show soft-label variant typically outperforms hard-label variant benchmark datasets partially explain behaviour studying relative difficulty finding good local minima corresponding objective functions.,18
"argument calculus networks. major reason behind success probability calculus possesses number valuable tools, based notion probabilistic independence. paper, identify notion logical independence makes tools available class propositional databases, called argument databases. specifically, suggest graphical representation argument databases, called argument networks, resemble bayesian networks. also suggest algorithm reasoning argument networks, resembles basic algorithm reasoning bayesian networks. finally, show argument networks several applications: nonmonotonic reasoning, truth maintenance, diagnosis.",3
expectation propagation continuous time stochastic processes. consider inverse problem reconstructing posterior measure trajec- tories diffusion process discrete time observations continuous time constraints. cast problem bayesian framework derive approximations posterior distributions single time marginals using variational approximate inference. show approximation extended wide class discrete-state markov jump pro- cesses making use chemical langevin equation. empirical results show proposed method computationally efficient provides good approximations classes inverse problems.,18
"integrating lexical temporal signals neural ranking models searching social media streams. time important relevance signal searching streams social media posts. distribution document timestamps results initial query leveraged infer distribution relevant documents, used rerank initial results. previous experiments shown kernel density estimation simple yet effective implementation idea. paper explores alternative approach mining temporal signals recurrent neural networks. intuition neural networks provide expressive framework capture temporal coherence neighboring documents time. knowledge, first integrate lexical temporal signals end-to-end neural network architecture, existing neural ranking models used generate query-document similarity vectors feed bidirectional lstm layer temporal modeling. results mixed: existing neural models document ranking alone yield limited improvements simple baselines, integration lexical temporal signals yield significant improvements competitive temporal baselines.",3
"depth assisted full resolution network single image-based view synthesis. researches novel viewpoint synthesis majorly focus interpolation multi-view input images. paper, focus challenging ill-posed problem synthesize novel viewpoints one single input image. achieve goal, propose novel deep learning-based technique. design full resolution network extracts local image features resolution input, contributes derive high resolution prevent blurry artifacts final synthesized images. also involve pre-trained depth estimation network system, thus 3d information able utilized infer flow field input target image. since depth network trained depth order information arbitrary pairs points scene, global image features also involved system. finally, synthesis layer used warp observed pixels desired positions also hallucinate missing pixels recorded pixels. experiments show technique performs well images various scenes, outperforms state-of-the-art techniques.",3
"unified treatment hidden markov switching models. many real-world problems encountered several disciplines deal modeling time-series containing different underlying dynamical regimes, probabilistic approaches often employed. paper describe several approaches common framework graphical models. give unified overview models previously introduced literature, simpler comprehensive previous descriptions enables us highlight commonalities differences among models observed past. addition, present several new models inference routines, naturally derived within unified viewpoint.",18
"statistical computational phase transitions group testing. study group testing problem goal identify set k infected individuals carrying rare disease within population size n, based outcomes pooled tests return positive whenever least one infected individual tested group. consider two different simple random procedures assigning individuals tests: constant-column design bernoulli design. first set results concerns fundamental statistical limits. constant-column design, give new information-theoretic lower bound implies proportion correctly identifiable infected individuals undergoes sharp ""all-or-nothing"" phase transition number tests crosses particular threshold. bernoulli design, determine precise number tests required solve associated detection problem (where goal distinguish group testing instance pure noise), improving upper lower bounds truong, aldridge, scarlett (2020). group testing models, also study power computationally efficient (polynomial-time) inference procedures. determine precise number tests required class low-degree polynomial algorithms solve detection problem. provides evidence inherent computational-statistical gap detection recovery problems small sparsity levels. notably, evidence contrary iliopoulos zadik (2021), predicted absence computational-statistical gap bernoulli design.",18
"killed albert einstein? open data murder mystery games. paper presents framework generating adventure games open data. focusing murder mystery type adventure games, generator able transform open data wikipedia articles, openstreetmap images wikimedia commons wikimysteries. every wikimystery game revolves around murder person wikipedia article populates game suspects must arrested player guilty murder absolved innocent. starting one person victim, extensive generative pipeline finds suspects, alibis, paths connecting open data, transforms open data cities, buildings, non-player characters, locks keys dialog options. paper describes detail generative step, provides specific playthrough one wikimystery albert einstein murdered, evaluates outcomes games generated 100 influential people 20th century.",3
"proximal quasi-newton computationally intensive l1-regularized m-estimators. consider class optimization problems arising computationally intensive l1-regularized m-estimators, function gradient values expensive compute. particular instance interest l1-regularized mle learning conditional random fields (crfs), popular class statistical models varied structured prediction problems sequence labeling, alignment, classification label taxonomy. l1-regularized mles crfs particularly expensive optimize since computing gradient values requires expensive inference step. work, propose use carefully constructed proximal quasi-newton algorithm computationally intensive m-estimation problems, employ aggressive active set selection technique. key contribution paper, show proximal quasi-newton method provably super-linearly convergent, even absence strong convexity, leveraging restricted variant strong convexity. experiments, proposed algorithm converges considerably faster current state-of-the-art problems sequence labeling hierarchical classification.",18
"impossibility learning missing mass. paper shows one cannot learn probability rare events without imposing structural assumptions. event interest obtaining outcome outside coverage i.i.d. sample discrete distribution. probability event referred ""missing mass"". impossibility result stated as: missing mass distribution-free pac-learnable relative error. proof semi-constructive relies coupling argument using dithered geometric distribution. result formalizes folklore order predict rare events, one necessarily needs distributions ""heavy tails"".",18
"kernelized bayesian matrix factorization. extend kernelized matrix factorization fully bayesian treatment ability work multiple side information sources expressed different kernels. kernel functions introduced matrix factorization integrate side information rows columns (e.g., objects users recommender systems), necessary making out-of-matrix (i.e., cold start) predictions. discuss specifically bipartite graph inference, output matrix binary, extensions general matrices straightforward. extend state art two key aspects: (i) fully conjugate probabilistic formulation kernelized matrix factorization problem enables efficient variational approximation, whereas fully bayesian treatments computationally feasible earlier approaches. (ii) multiple side information sources included, treated different kernels multiple kernel learning additionally reveals side information sources informative. method outperforms alternatives predicting drug-protein interactions two data sets. show framework also used solving multilabel learning problems considering samples labels two domains matrix factorization operates on. algorithm obtains lowest hamming loss values 10 14 multilabel classification data sets compared five state-of-the-art multilabel learning algorithms.",18
"variational inference uncertainty inputs gaussian process models. gaussian process latent variable model (gp-lvm) provides flexible approach non-linear dimensionality reduction widely applied. however, current approach training gp-lvms based maximum likelihood, latent projection variables maximized rather integrated out. paper present bayesian method training gp-lvms introducing non-standard variational inference framework allows approximately integrate latent variables subsequently train gp-lvm maximizing analytic lower bound exact marginal likelihood. apply method learning gp-lvm iid observations learning non-linear dynamical systems observations temporally correlated. show benefit variational bayesian procedure robustness overfitting ability automatically select dimensionality nonlinear latent space. resulting framework generic, flexible easy extend purposes, gaussian process regression uncertain inputs semi-supervised gaussian processes. demonstrate method synthetic data standard machine learning benchmarks, well challenging real world datasets, including high resolution video data.",18
"classification finite primitive ibis groups alternating socle. let $g$ finite permutation group $\omega$. ordered sequence $(\omega_1,\ldots,\omega_\ell)$ elements $\omega$ irredundant base $g$ pointwise stabilizer trivial point fixed stabilizer predecessors. irredundant bases $g$ cardinality, $g$ said ibis group. lucchini, morigi moscatiello proved theorem reducing problem classifying finite primitive ibis groups $g$ case socle $g$ either abelian non-abelian simple. paper, classify finite primitive ibis groups socle alternating group. moreover, propose conjecture aiming give classification almost simple primitive ibis groups.",10
"testing differences gaussian graphical models: applications brain connectivity. functional brain networks well described estimated data gaussian graphical models (ggms), e.g. using sparse inverse covariance estimators. comparing functional connectivity subjects two populations calls comparing estimated ggms. goal identify differences ggms known similar structure. characterize uncertainty differences confidence intervals obtained using parametric distribution parameters sparse estimator. sparse penalties enable statistical guarantees interpretable models even high-dimensional low-sample settings. characterizing distributions sparse models inherently challenging penalties produce biased estimator. recent work invokes sparsity assumptions effectively remove bias sparse estimator lasso. distributions used give confidence intervals edges ggms, extension differences. however, case comparing ggms, estimators make use assumed joint structure among ggms. inspired priors brain functional connectivity derive distribution parameter differences joint penalty parameters known sparse difference. leads us introduce debiased multi-task fused lasso, whose distribution characterized efficient manner. show debiased lasso multi-task fused lasso used obtain confidence intervals edge differences ggms. validate techniques proposed set synthetic examples well neuro-imaging dataset created study autism.",18
"dose prediction u-net: feasibility study predicting dose distributions contours using deep learning prostate imrt patients. advancement treatment modalities radiation therapy, outcomes haves greatly improved, cost increased treatment plan complexity planning time. accurate prediction dose distributions would alleviate issue guiding clinical plan optimization save time maintain high quality plans. developed novel application fully convolutional deep network model, u-net, predicting dose patient contours. show model, able accurately predict dose prostate cancer patients, average dice similarity coefficient well 0.9 comparing predicted vs. true isodose volumes 0% 100% prescription dose. average differences mean max dose structures within 2.3% prescription dose.",14
"sphere plane: two-dimensional scattering finite curved region. non-relativistic particles effectively confined two dimensions general move curved surfaces, allowing dynamical phenomena beyond described scalar potentials even vector gauge fields. consider simple case piecewise uniform curvature: particle moves plane spherical extrusion. depending latitude sphere joins plane, extrusion range infinitesimal bump nearly full sphere touches plane. free classical motion surface piecewise uniform curvature follows geodesics independent velocity, quantum mechanical scattering depends energy. compare classical, semi-classical, fully quantum problems, exactly solvable, show semi-classical analysis explains complex quantum differential cross section terms interference two classical trajectories: sphere plane acts kind double slit.",17
"automatic brain tumor segmentation using cascaded anisotropic convolutional neural networks. cascade fully convolutional neural networks proposed segment multi-modal magnetic resonance (mr) images brain tumor background three hierarchical regions: whole tumor, tumor core enhancing tumor core. cascade designed decompose multi-class segmentation problem sequence three binary segmentation problems according subregion hierarchy. whole tumor segmented first step bounding box result used tumor core segmentation second step. enhancing tumor core segmented based bounding box tumor core segmentation result. networks consist multiple layers anisotropic dilated convolution filters, combined multi-view fusion reduce false positives. residual connections multi-scale predictions employed networks boost segmentation performance. experiments brats 2017 validation set show proposed method achieved average dice scores 0.7859, 0.9050, 0.8378 enhancing tumor core, whole tumor tumor core, respectively. corresponding values brats 2017 testing set 0.7831, 0.8739, 0.7748, respectively.",3
"size-to-depth: new perspective single image depth estimation. paper consider problem single monocular image depth estimation. challenging problem due ill-posedness nature found wide application industry. previous efforts belongs roughly two families: learning-based method interactive method. learning-based method, deep convolutional neural network (cnn) widely used, achieve good result. suffer low generalization ability typically perform poorly unfamiliar scenes. besides, data-hungry nature method makes data aquisition expensive time-consuming. interactive method requires human annotation depth which, however, errorneous large variance. overcome problems, propose new perspective single monocular image depth estimation problem: size depth. method require sparse label real-world size object rather raw depth. coarse depth map inferred following geometric relationships according size labels. refine depth map energy function optimization conditional random field(crf). experimentally demonstrate method outperforms traditional depth-labeling methods produce satisfactory depth maps.",3
"online learning mdps side information. study online learning finite markov decision process (mdp) problems side information vector available. problem motivated applications clinical trials, recommendation systems, etc. applications episodic structure, episode corresponds patient/customer. objective compete optimal dynamic policy take side information account. propose computationally efficient algorithm show regret $o(\sqrt{t})$, $t$ number rounds. best knowledge, first regret bound setting.",3
"geometrical morphology. explore inflectional morphology example relationship discrete continuous linguistics. grammar requests form lexeme specifying set feature values, corresponds corner hypercube feature value space. morphology responds request providing morpheme, set morphemes, whose vector sum geometrically closest corner m. short, chosen morpheme $\mu$ morpheme (or set morphemes) maximizes inner product $\mu$ m.",3
"subspace orthogonalization mechanism binding values space. choosing options, must solve important binding problem. values options must associated information, including action needed select them. hypothesized brain solves binding problem use distinct population subspaces. examined responses single neurons five value-sensitive regions rhesus macaques performing risky choice task. areas, neurons encoded values possible options, used semi-orthogonal coding subspaces associated left right options, served link options positions space. also observed covariation subspace orthogonalization behavior: trials less orthogonalized subspaces associated greater likelihood choosing less valued option. semi-orthogonal subspaces arose combination linear non-linear mixed selective neurons. decomposing neural geometry, show combination selectivity achieves code balances binding/separation generalization. results support hypothesis binding operations serve convert high-dimensional codes multiple low-dimensional neural subspaces flexibly solve decision problems.",15
"steal vocal identity internet?: initial investigation cloning obama's voice using gan, wavenet low-quality found data. thanks growing availability spoofing databases rapid advances using them, systems detecting voice spoofing attacks becoming capable, error rates close zero reached asvspoof2015 database. however, speech synthesis voice conversion paradigms considered asvspoof2015 database appearing. examples include direct waveform modelling generative adversarial networks. also need investigate feasibility training spoofing systems using low-quality found data. purpose, developed generative adversarial network-based speech enhancement system improves quality speech data found publicly available sources. using enhanced data, trained state-of-the-art text-to-speech voice conversion models evaluated terms perceptual speech quality speaker similarity. results show enhancement models significantly improved snr low-quality degraded data found publicly available sources significantly improved perceptual cleanliness source speech without significantly degrading naturalness voice. however, results also show limitations generating speech low-quality found data.",5
"hybrid primal heuristic robust multiperiod network design. investigate robust multiperiod network design problem, generalization classical capacitated network design problem additionally considers multiple design periods provides solutions protected traffic uncertainty. given intrinsic difficulty problem, proves challenging even state-of-the art commercial solvers, propose hybrid primal heuristic based combination ant colony optimization exact large neighborhood search. computational experiments set realistic instances sndlib show heuristic find solutions extremely good quality low optimality gap.",10
"slow-roll inflation n3lo. next generation cosmological observations sensitive small deviations pure power law primordial power spectrum curvature perturbations. context slow-roll inflation, deviations expected correspond so-called running spectral index. measurement would bring much information discovery deviations scale invariance. however, robust parameter inference requires marginalize possible higher order uncertainties, been, now, fully determined. tackle issue deriving inflationary scalar tensor slow-roll power spectra next-to-next-to-next leading order (n3lo), fully expanded around observable pivot wavenumber, single field inflationary models minimal non-minimal kinetic terms. result therefore encompasses string-inspired inflationary models varying speed sound.",0
optimal rates regularized learning algorithms general source condition. consider learning algorithms general source condition polynomial decay eigenvalues integral operator vector-valued function setting. discuss upper convergence rates tikhonov regularizer general source condition corresponding increasing monotone index function. convergence issues studied general regularization schemes using concept operator monotone index functions minimax setting. also address minimum possible error learning algorithm.,18
"top-down saliency detection driven visual classification. paper presents approach top-down saliency detection guided visual classification tasks. first learn compute visual saliency specific visual task accomplished, opposed state-of-the-art methods assess saliency merely bottom-up principles. afterwards, investigate extent visual saliency support visual classification nontrivial cases. achieve this, propose salclassnet, cnn framework consisting two networks jointly trained: a) first one computing top-down saliency maps input images, b) second one exploiting computed saliency maps visual classification. test approach, collected dataset eye-gaze maps, using tobii t60 eye tracker, asking several subjects look images stanford dogs dataset, objective distinguishing dog breeds. performance analysis dataset saliency bench-marking datasets, poet, showed salclassnet out-performs state-of-the-art saliency detectors, salnet salicon. finally, analyzed performance salclassnet fine-grained recognition task found generalizes better existing visual classifiers. achieved results, thus, demonstrate 1) conditioning saliency detectors object classes reaches state-of-the-art performance, 2) providing explicitly top-down saliency maps visual classifiers enhances classification accuracy.",3
"bayesian inference posterior regularization applications infinite latent svms. existing bayesian models, especially nonparametric bayesian methods, rely specially conceived priors incorporate domain knowledge discovering improved latent representations. priors affect posterior distributions bayes' rule, imposing posterior regularization arguably direct cases natural general. paper, present regularized bayesian inference (regbayes), novel computational framework performs posterior inference regularization term desired post-data posterior distribution information theoretical formulation. regbayes flexible procedure elicits expert knowledge via priors, covers directed bayesian networks undirected markov networks whose bayesian formulation results hybrid chain graph models. regularization induced linear operator posterior distributions, expectation operator, present general convex-analysis theorem characterize solution regbayes. furthermore, present two concrete examples regbayes, infinite latent support vector machines (ilsvm) multi-task infinite latent support vector machines (mt-ilsvm), explore large-margin idea combination nonparametric bayesian model discovering predictive latent features classification multi-task learning, respectively. present efficient inference methods report empirical studies several benchmark datasets, appear demonstrate merits inherited large-margin learning bayesian nonparametrics. results available now, contribute push forward interface two important subfields, largely treated isolated community.",3
"joint discovery object states manipulation actions. many human activities involve object manipulations aiming modify object state. examples common state changes include full/empty bottle, open/closed door, attached/detached car wheel. work, seek automatically discover states objects associated manipulation actions. given set videos particular task, propose joint model learns identify object states localize state-modifying actions. model formulated discriminative clustering cost constraints. assume consistent temporal order changes object states manipulation actions, introduce new optimization techniques learn model parameters without additional supervision. demonstrate successful discovery seven manipulation actions corresponding object states new dataset videos depicting real-life object manipulations. show joint formulation results improvement object state discovery action recognition vice versa.",3
"flying-qubit control via three-level atom tunable waveguide couplings. control flying qubits core quantum networks. often carried single-photon fields, flying-qubit control involves logical states also shapes. paper, explore variety flying-qubit control problems using three-level atom time-varying tunable couplings two input-output channels. shown one tune couplings $\lambda$-type atom distribute single photon two channels arbitrary shapes, use $v$-type atom catch arbitrary-shape distributed single photon. $\lambda$-type atom also designed transfer flying qubit one channel other, central frequency photon shape converted. $\xi$-type atom, one use tunable coupling shape pair correlated photons via cascaded emission. cases, analytical formulas derived coupling functions fulfil control tasks, physical limitations discussed well. results provide useful control protocols high-fidelity quantum information transmission complex quantum networks.",17
"unsupervised learning depth ego-motion video. present unsupervised learning framework task monocular depth camera motion estimation unstructured video sequences. achieve simultaneously training depth camera pose estimation networks using task view synthesis supervisory signal. networks thus coupled via view synthesis objective training, applied independently test time. empirical evaluation kitti dataset demonstrates effectiveness approach: 1) monocular depth performing comparably supervised methods use either ground-truth pose depth training, 2) pose estimation performing favorably established slam systems comparable input settings.",3
"node-based learning multiple gaussian graphical models. consider problem estimating high-dimensional gaussian graphical models corresponding single set variables several distinct conditions. problem motivated task recovering transcriptional regulatory networks basis gene expression data {containing heterogeneous samples, different disease states, multiple species, different developmental stages}. assume aspects conditional dependence networks shared, structured differences them. rather assuming similarities differences networks driven individual edges, take node-based approach, many cases provides intuitive interpretation network differences. consider estimation two distinct assumptions: (1) differences k networks due individual nodes perturbed across conditions, (2) similarities among k networks due presence common hub nodes shared across k networks. using row-column overlap norm penalty function, formulate two convex optimization problems correspond two assumptions. solve problems using alternating direction method multipliers algorithm, derive set necessary sufficient conditions allows us decompose problem independent subproblems algorithm scaled high-dimensional settings. proposal illustrated synthetic data, webpage data set, brain cancer gene expression data set.",18
"hybrid optimized back propagation learning algorithm multi-layer perceptron. standard neural network based general back propagation learning using delta method gradient descent method great faults like poor optimization error-weight objective function, low learning rate, instability .this paper introduces hybrid supervised back propagation learning algorithm uses trust-region method unconstrained optimization error objective function using quasi-newton method .this optimization leads accurate weight update system minimizing learning error learning phase multi-layer perceptron.[13][14][15] paper augmented line search used finding points satisfies wolfe condition. paper, hybrid back propagation algorithm strong global convergence properties & robust & efficient practice.",3
"petuum: new platform distributed machine learning big data. systematic way efficiently apply wide spectrum advanced ml programs industrial scale problems, using big models (up 100s billions parameters) big data (up terabytes petabytes)? modern parallelization strategies employ fine-grained operations scheduling beyond classic bulk-synchronous processing paradigm popularized mapreduce, even specialized graph-based execution relies graph representations ml programs. variety approaches tends pull systems algorithms design different directions, remains difficult find universal platform applicable wide range ml programs scale. propose general-purpose framework systematically addresses data- model-parallel challenges large-scale ml, observing many ml programs fundamentally optimization-centric admit error-tolerant, iterative-convergent algorithmic solutions. presents unique opportunities integrative system design, bounded-error network synchronization dynamic scheduling based ml program structure. demonstrate efficacy system designs versus well-known implementations modern ml algorithms, allowing ml programs run much less time considerably larger model sizes, even modestly-sized compute clusters.",18
"combining machine learning decision making. present new application covering number bound framework ""machine learning operational costs (mloc),"" exploratory form decision theory. mloc framework incorporates knowledge predictive model used subsequent task, thus combining machine learning decision made afterwards. work, use mloc framework study problem implications power grid reliability maintenance, called machine learning traveling repairman problem ml&trp. goal ml&trp determine route ""repair crew,"" repairs nodes graph. repair crew aims minimize cost failures nodes, many real situations, failure probabilities known must estimated. mloc framework allows us understand uncertainty influences repair route. also present new covering number generalization bounds mloc framework.",10
"concave penalized estimation sparse gaussian bayesian networks. develop penalized likelihood estimation framework estimate structure gaussian bayesian networks observational data. contrast recent methods accelerate learning problem restricting search space, main contribution fast algorithm score-based structure learning restrict search space way works high-dimensional datasets thousands variables. use concave regularization, opposed popular $\ell_0$ (e.g. bic) penalty, new. moreover, provide theoretical guarantees generalize existing asymptotic results underlying distribution gaussian. notably, framework require existence so-called faithful dag representation, result theory must handle inherent nonidentifiability estimation problem novel way. finally, matter independent interest, provide comprehensive comparison approach several standard structure learning methods using open-source packages developed r language. based experiments, show algorithm significantly faster competing methods obtaining higher sensitivity comparable false discovery rates high-dimensional data. particular, total runtime method generate solution path 20 estimates dags 8000 nodes around one hour.",18
"efficiently sampling multiplicative attribute graphs using ball-dropping process. introduce novel efficient sampling algorithm multiplicative attribute graph model (magm - kim leskovec (2010)}). algorithm \emph{strictly} efficient algorithm proposed yun vishwanathan (2012), sense method extends \emph{best} time complexity guarantee algorithm larger fraction parameter space. theory empirical evaluation sparse graphs, new algorithm outperforms previous one. design algorithm, first define stochastic \emph{ball-dropping process} (bdp). although special case process introduced efficient approximate sampling algorithm kronecker product graph model (kpgm - leskovec et al. (2010)}), neither \emph{why} approximation works \emph{what} actual distribution process sampling addressed far best knowledge. rigorous treatment bdp enables us clarify rational behind bdp approximation kpgm, design efficient sampling algorithm magm.",18
"computational approach relative aesthetics. computational visual aesthetics recently become active research area. existing state-of-art methods formulate binary classification task given image predicted beautiful not. many applications image retrieval enhancement, important rank images based aesthetic quality instead binary-categorizing them. furthermore, applications, may possible images belong category. hence determining aesthetic ranking images appropriate. end, formulate novel problem ranking images respect aesthetic quality. construct new dataset image pairs relative labels carefully selecting images popular ava dataset. unlike aesthetics classification, single threshold would determine ranking order images across entire dataset. propose deep neural network based approach trained image pairs incorporating principles relative learning. results show relative training procedure allows network rank images higher accuracy state-of-art network trained set images using binary labels.",3
"deep learning real-time gravitational wave detection parameter estimation ligo data. recent nobel-prize-winning detections gravitational waves merging black holes subsequent detection collision two neutron stars coincidence electromagnetic observations inaugurated new era multimessenger astrophysics. enhance scope emergent science, proposed use deep convolutional neural networks detection characterization gravitational wave signals real-time. method, deep filtering, initially demonstrated using simulated ligo noise. article, present extension deep filtering using real data first observing run ligo, detection parameter estimation gravitational waves binary black hole mergers continuous data streams multiple ligo detectors. show first time machine learning detect estimate true parameters real gw event observed ligo. comparisons show deep filtering far computationally efficient matched-filtering, retaining similar sensitivity lower errors, allowing real-time processing weak time-series signals non-stationary non-gaussian noise, minimal resources, also enables detection new classes gravitational wave sources may go unnoticed existing detection algorithms. approach uniquely suited enable coincident detection campaigns gravitational waves multimessenger counterparts real-time.",6
"convergence proof split bregman method regularized least-squares problems. split bregman (sb) method [t. goldstein s. osher, siam j. imaging sci., 2 (2009), pp. 323-43] fast splitting-based algorithm solves image reconstruction problems general l1, e.g., total-variation (tv) compressed sensing (cs), regularizations introducing single variable split decouple data-fitting term regularization term, yielding simple subproblems separable (or partially separable) easy minimize. several convergence proofs proposed, proofs either impose ""full column rank"" assumption split assume exact updates subproblems. however, assumptions impractical many applications x-ray computed tomography (ct) image reconstructions, inner least-squares problem usually cannot solved efficiently due highly shift-variant hessian. paper, show data-fitting term quadratic, sb method convergent alternating direction method multipliers (admm), straightforward convergence proof inexact updates given using [j. eckstein d. p. bertsekas, mathematical programming, 55 (1992), pp. 293-318, theorem 8]. furthermore, since sb method special case admm algorithm, seems likely admm algorithm faster sb method augmented largangian (al) penalty parameters selected appropriately. concrete example, conduct convergence rate analysis admm algorithm using two splits image restoration problems quadratic data-fitting term regularization term. according analysis, show two-split admm algorithm faster sb method al penalty parameter sb method suboptimal. numerical experiments conducted verify analysis.",10
"stress concentration factors stokes flow two nearly touching rigid particles. paper, mathematical model two adjacent rigid particles immersed viscous incompressible fluid considered. main feature flow cauchy stress tensor consisting strain tensor pressure appear blow-up distance two particles tends zero. purpose making clear high concentration, family unified stress concentration factors precisely captured dimensions, determine whether cauchy stress tensor blow not. direct application, establish optimal gradient estimates asymptotics cauchy stress tensor stokes flow, indicate maximal singularity comes pressure.",10
"continuous 2-category symmetries yang-mills theory. study 4d gauge theory $u(1)^{n-1}\rtimes s_n$ obtained $u(1)^{n-1}$ theory gauging 0-form symmetry $s_n$. show theory global continuous 2-category symmetry, whose structure particularly rich $n>2$. example allows us draw connection higher gauging procedure difference local global fusion, turns key feature higher category symmetries. studying spectrum local extended operators, find mapping gauge invariant operators 4d $su(n)$ yang-mills theory. largest group-like subcategory non-invertible symmetries theory $\mathbb{z}_n^{(1)}$ 1-form symmetry, acting wilson lines way center symmetry yang-mills theory does. supported path-integral argument, propose $u(1)^{n-1}\rtimes s_n$ gauge theory relation ultraviolet limit $su(n)$ yang-mills theory gukov-witten operators become topological, form continuous non-invertible 2-category symmetry, broken center symmetry rg flow.",9
"high-dimensional motion planning using latent variable models via approximate inference. work, present efficient framework generate motion trajectory robot high degree freedom (e.g., humanoid robot). high-dimensionality robot configuration space often leads difficulties utilizing widely-used motion planning algorithms volume decision space increases exponentially number dimensions. handle complications arising large decision space, solve corresponding motion planning problem efficiently, two key concepts adopted. first, gaussian process latent variable model (gp-lvm) utilized low-dimensional representation original configuration space. second, approximate inference algorithm used, exploiting duality control estimation, explore decision space compute high-quality motion trajectory robot. utilizing gp-lvm duality control estimation, construct fully probabilistic generative model transform high-dimensional motion planning problem tractable inference problem. finally, compute optimal motion trajectory via approximate inference algorithm based variant particle filter. numerical experiments presented demonstrate applicability validity proposed approach. resulting motions viewed supplemental video. ( https://youtu.be/dsz1afn6pty )",3
"quantum decision theory quantum theory measurement. present general theory quantum information processing devices, applied human decision makers, atomic multimode registers, molecular high-spin registers. quantum decision theory generalization quantum theory measurement, endowed action ring, prospect lattice probability operator measure. algebra probability operators plays role algebra local observables. composite nature prospects entangling properties probability operators, quantum interference terms appear, make actions noncommutative prospect probabilities non-additive. theory provides basis explaining variety paradoxes typical application classical utility theory real human decision making. principal advantage approach formulated self-consistent mathematical theory, allows us explain one effect actually known paradoxes human decision making. general, approach serve tool characterizing quantum information processing means atomic, molecular, condensed-matter systems.",17
"principled parallel mean-field inference discrete random fields. mean-field variational inference one popular approaches inference discrete random fields. standard mean-field optimization based coordinate descent many situations impractical. thus, practice, various parallel techniques used, either rely ad-hoc smoothing heuristically set parameters, put strong constraints type models. paper, propose novel proximal gradient-based approach optimizing variational objective. naturally parallelizable easy implement. prove convergence, demonstrate that, practice, yields faster convergence often finds better optima traditional mean-field optimization techniques. moreover, method less sensitive choice parameters.",3
"face retrieval using frequency decoded local descriptor. local descriptors backbone computer vision problems. existing local descriptors generated raw input images. order increase discriminative power local descriptors, researchers converted raw image multiple images help high low pass frequency filters, local descriptors computed filtered image finally concatenated single descriptor. so, approaches utilize inter frequency relationship causes less improvement discriminative power descriptor could achieved. paper, problem solved utilizing decoder concept multi-channel decoded local binary pattern multi-frequency patterns. frequency decoded local binary pattern (fdlbp) proposed two decoders. decoder works one low frequency pattern two high frequency pattern. finally, descriptors decoders concatenated form single descriptor. face retrieval experiments conducted four benchmarks challenging databases pasc, lfw, pubfig, essex. experimental results confirm superiority fdlbp descriptor compared state-of-the-art descriptors lbp, sobel_lbp, bof_lbp, svd_s_lbp, mdlbp, etc.",3
"supersparse linear integer models interpretable classification. scoring systems classification models require users add, subtract multiply meaningful numbers make prediction. models often used practical interpretable. paper, introduce off-the-shelf tool create scoring systems accurate interpretable, known supersparse linear integer model (slim). slim discrete optimization problem minimizes 0-1 loss encourage high level accuracy, regularizes l0-norm encourage high level sparsity, constrains coefficients set interpretable values. illustrate practical interpretable nature slim scoring systems applications medicine criminology, show accurate sparse comparison state-of-the-art classification models using numerical experiments.",18
"online em algorithm latent data models. contribution, propose generic online (also sometimes called adaptive recursive) version expectation-maximisation (em) algorithm applicable latent variable models independent observations. compared algorithm titterington (1984), approach directly connected usual em algorithm rely integration respect complete data distribution. resulting algorithm usually simpler shown achieve convergence stationary points kullback-leibler divergence marginal distribution observation model distribution optimal rate, i.e., maximum likelihood estimator. addition, proposed approach also suitable conditional (or regression) models, illustrated case mixture linear regressions model.",18
"thermodynamic unification optimal transport: thermodynamic uncertainty relation, minimum dissipation, thermodynamic speed limits. thermodynamics provides universal means study physical systems energetic perspective. recent years, notions thermodynamics generalized small fluctuating systems, establishing field stochastic quantum thermodynamics. independently developed mathematics statistics, optimal transport theory concerns optimally transport source distribution target distribution, deriving useful metric probability distributions called wasserstein distance. despite seemingly unrelated nature, intimate connection fields unveiled context continuous-state langevin dynamics, providing many important implications nonequilibrium systems. study, elucidate analogous connection discrete case developing thermodynamic framework discrete optimal transport. first introduce novel quantity named dynamical state mobility, significantly improves thermodynamic uncertainty relation provides insight precision currents nonequilibrium markov jump processes. derive variational formulas connect discrete wasserstein distances stochastic quantum thermodynamics discrete markovian dynamics described master equations. specifically, rigorously prove wasserstein distance equals minimum product irreversible entropy production dynamical state mobility admissible markovian dynamics. formulas unify relationship thermodynamics optimal transport theory discrete continuous cases also generalize quantum case. addition, show obtained variational formulas lead remarkable applications stochastic quantum thermodynamics, stringent thermodynamic speed limits finite-time landauer principle.",2
"nonlinear dimensionality reduction framework using smooth geodesics. existing dimensionality reduction methods adept revealing hidden underlying manifolds arising high-dimensional data thereby producing low-dimensional representation. however, smoothness manifolds produced classic techniques presence noise guaranteed. fact, embedding generated using non-smooth, noisy measurements may distort geometry manifold thereby produce unfaithful embedding. herein, propose framework nonlinear dimensionality reduction generates manifold terms smooth geodesics designed treat problems manifold measurements corrupted noise. method generates network structure given high-dimensional data using neighborhood search produces piecewise linear shortest paths defined geodesics. then, fit points geodesic smoothing spline emphasize smoothness. robustness approach noisy sparse datasets demonstrated implementation method synthetic real-world datasets.",18
online handwritten mathematical expressions recognition system using fuzzy neural network. article describes developed information technology online recognition handwritten mathematical expressions based proposed approaches handwritten symbols recognition structural analysis.,3
"necessity mixed models: dynamical frustrations mind. present work present analyze basic processes local global level linguistic derivations seem go beyond limits markovian turing-like computation, require, opinion, quantum processor. first present briefly working hypothesis focus empirical domain. time, argue model appealing one kind computation (be quantum not) necessarily insufficient, thus linear non-linear formal models invoked order pursue fuller understanding mental computations within unified framework.",12
"pixel-wise deep learning contour detection. address problem contour detection via per-pixel classifications edge point. facilitate process, proposed approach leverages densenet, efficient implementation multiscale convolutional neural networks (cnns), extract informative feature vector pixel uses svm classifier accomplish contour detection. experiment contour detection, look effectiveness combining per-pixel features different cnn layers verify performance bsds500.",3
rank-frequency analysis functional style corpora ukrainian language. use rank-frequency analysis estimation kernel vocabulary size within specific corpora ukrainian. extrapolation high-rank behaviour utilized estimation total vocabulary size.,3
"adaptive chemical reaction optimization global numerical optimization. newly proposed chemical-reaction-inspired metaheurisic, chemical reaction optimization (cro), applied many optimization problems discrete continuous domains. alleviate effort tuning parameters, paper reduces number optimization parameters canonical cro develops adaptive scheme evolve them. proposed adaptive cro (acro) adapts better different optimization problems. perform simulations acro widely-used benchmark continuous problems. simulation results show acro superior performance canonical cro.",3
"variance-reduced proximal stochastic gradient descent non-convex composite optimization. study non-convex composite optimization: first, finite-sum smooth non-convex functions, second, general function admits simple proximal mapping. research stochastic methods composite optimization assumes convexity strong convexity function. paper, extend problem non-convex setting using variance reduction techniques, prox-svrg prox-saga. prove that, constant step size, prox-svrg prox-saga suitable non-convex composite optimization, help problem converge stationary point within $o(1/\epsilon)$ iterations. similar convergence rate seen state-of-the-art rsag method faster stochastic gradient descent. analysis also extended min-batch setting, linearly accelerates convergence. best knowledge, first analysis convergence rate variance-reduced proximal stochastic gradient non-convex composite optimization.",18
"set-to-set hashing applications visual recognition. visual data, image sequence video frames, often naturally represented point set. paper, consider fundamental problem finding nearest set collection sets, query set. problem obvious applications large-scale visual retrieval recognition, also applied fields beyond computer vision. one challenge stands solving problem---set representation measure similarity. particularly, query set sets dataset collection varying cardinalities. training collection large enough linear scan impractical. propose simple representation scheme encodes statistical structural information sets. derived representations integrated kernel framework flexible similarity measurement. query set process, adopt learning-to-hash pipeline turns kernel representations hash bits based simple learners, using multiple kernel learning. experiments two visual retrieval datasets show unambiguously set-to-set hashing framework outperforms prior methods take set-to-set search setting.",3
"constructing binary descriptors stochastic hill climbing search. binary descriptors image patches provide processing speed advantages require less storage methods encode patch appearance vector real numbers. provide evidence that, despite simplicity, stochastic hill climbing bit selection procedure descriptor construction defeats recently proposed alternatives standard discriminative power benchmark. method easy implement understand, free parameters need fine tuning, runs fast.",3
"segmenting brain tumors symmetry. explore encoding brain symmetry neural network brain tumor segmentation task. healthy human brain symmetric high level abstraction, high-level asymmetric parts likely tumor regions. paying attention asymmetries potential boost performance brain tumor segmentation. propose method encode brain symmetry existing neural networks apply method state-of-the-art neural network medical imaging segmentation. evaluate symmetry-encoded network dataset brain tumor segmentation challenge verify new model extracts information training images efficiently original model.",3
"characterization deterministic probabilistic sampling patterns finite completability low tensor-train rank tensor. paper, analyze fundamental conditions low-rank tensor completion given separation tensor-train (tt) rank, i.e., ranks unfoldings. exploit algebraic structure tt decomposition obtain deterministic necessary sufficient conditions locations samples ensure finite completability. specifically, propose algebraic geometric analysis tt manifold incorporate whole rank vector simultaneously contrast existing approach based grassmannian manifold incorporate one rank component. proposed technique characterizes algebraic independence set polynomials defined based sampling pattern tt decomposition, instrumental obtaining deterministic condition sampling pattern finite completability. addition, based proposed analysis, assuming entries tensor sampled independently probability $p$, derive lower bound sampling probability $p$, equivalently, number sampled entries ensures finite completability high probability. moreover, also provide deterministic probabilistic conditions unique completability.",3
multiple pattern classification sparse subspace decomposition. robust classification method developed basis sparse subspace decomposition. method tries decompose mixture subspaces unlabeled data (queries) class subspaces possible. query classified class whose subspace significantly contributes decomposed subspace. multiple queries different classes simultaneously classified respective classes. practical greedy algorithm sparse subspace decomposition designed classification. present method achieves high recognition rate robust performance exploiting joint sparsity.,3
"complexity analysis randomized block-coordinate descent methods. paper analyze randomized block-coordinate descent (rbcd) methods proposed [8,11] minimizing sum smooth convex function block-separable convex function. particular, extend nesterov's technique developed [8] analyzing rbcd method minimizing smooth convex function block-separable closed convex set aforementioned general problem obtain sharper expected-value type convergence rate one implied [11]. also, obtain better high-probability type iteration complexity, improves upon one [11] least amount $o(n/\epsilon)$, $\epsilon$ target solution accuracy $n$ number problem blocks. addition, unconstrained smooth convex minimization, develop new technique called {\it randomized estimate sequence} analyze accelerated rbcd method proposed nesterov [11] establish sharper expected-value type convergence rate one given [11].",10
"formal solution grain truth problem. bayesian agent acting multi-agent environment learns predict agents' policies prior assigns positive probability (in words, prior contains \emph{grain truth}). finding reasonably large class policies contains bayes-optimal policies respect class known \emph{grain truth problem}. small classes known grain truth literature contains several related impossibility results. paper present formal general solution full grain truth problem: construct class policies contains computable policies well bayes-optimal policies every lower semicomputable prior class. environment unknown, bayes-optimal agents may fail act optimally even asymptotically. however, agents based thompson sampling converge play {\epsilon}-nash equilibria arbitrary unknown computable multi-agent environments. results purely theoretical, show computationally approximated arbitrarily closely.",3
"missing data reconstruction remote sensing image unified spatial-temporal-spectral deep convolutional neural network. internal malfunction satellite sensors poor atmospheric conditions thick cloud, acquired remote sensing data often suffer missing information, i.e., data usability greatly reduced. paper, novel method missing information reconstruction remote sensing images proposed. unified spatial-temporal-spectral framework based deep convolutional neural network (sts-cnn) employs unified deep convolutional neural network combined spatial-temporal-spectral supplementary information. addition, address fact methods deal single missing information reconstruction task, proposed approach solve three typical missing information reconstruction tasks: 1) dead lines aqua modis band 6; 2) landsat etm+ scan line corrector (slc)-off problem; 3) thick cloud removal. noted proposed model use multi-source data (spatial, spectral, temporal) input unified framework. results simulated real-data experiments demonstrate proposed model exhibits high effectiveness three missing information reconstruction tasks listed above.",3
"tagbook: semantic video representation without supervision event detection. consider problem event detection video scenarios few, even zero examples available training. challenging setting, prevailing solutions literature rely semantic video representation obtained thousands pre-trained concept detectors. different existing work, propose new semantic video representation based freely available social tagged videos only, without need training intermediate concept detectors. introduce simple algorithm propagates tags video's nearest neighbors, similar spirit ones used image retrieval, redesign video event detection including video source set refinement varying video tag assignment. call approach tagbook study construction, descriptiveness detection performance trecvid 2013 2014 multimedia event detection datasets columbia consumer video dataset. despite simple nature, proposed tagbook video representation remarkably effective few-example zero-example event detection, even outperforming recent state-of-the-art alternatives building supervised representations.",3
"appearance descriptors person re-identification: comprehensive review. video-surveillance, person re-identification task recognising whether individual already observed network cameras. typically, achieved exploiting clothing appearance, classical biometric traits like face impractical real-world video surveillance scenarios. clothing appearance represented means low-level \textit{local} and/or \textit{global} features image, usually extracted according part-based body model treat different body parts (e.g. torso legs) independently. paper provides comprehensive review current approaches build appearance descriptors person re-identification. relevant techniques described detail, categorised according body models features used. aim work provide structured body knowledge starting point researchers willing conduct novel investigations challenging topic.",3
"generating retinal flow maps structural optical coherence tomography artificial intelligence. despite significant advances artificial intelligence (ai) computer vision, application medical imaging limited burden limits expert-generated labels. used images optical coherence tomography angiography (octa), relatively new imaging modality measures perfusion retinal vasculature, train ai algorithm generate vasculature maps standard structural optical coherence tomography (oct) images retinae, exceeding ability bypassing need expert labeling. deep learning able infer perfusion microvasculature structural oct images similar fidelity octa significantly better expert clinicians (p < 0.00001). octa suffers need specialized hardware, laborious acquisition protocols, motion artifacts; whereas model works directly standard oct ubiquitous quick obtain, allows unlocking large volumes previously collected standard oct data existing clinical trials clinical practice. finding demonstrates novel application ai medical imaging, whereby subtle regularities different modalities used image body part ai used generate detailed accurate inferences tissue function structure imaging.",3
"multi-view learning structured non-identical outputs. many machine learning problems, labeled training data limited unlabeled data ample. problems instances factored multiple views, nearly sufficent determining correct labels. paper present new algorithm probabilistic multi-view learning uses idea stochastic agreement views regularization. algorithm works structured unstructured problems easily generalizes partial agreement scenarios. full agreement case, algorithm minimizes bhattacharyya distance models view, performs better coboosting two-view perceptron several flat structured classification problems.",3
"neurally implementable semantic networks. propose general principles semantic networks allowing implemented dynamical neural networks. major features scheme include: (a) interpretation node network stands bound integration meanings nodes external events node links with; (b) systematic use nodes stand categories types, separate nodes instances types; (c) implementation relationships use intrinsically typed links nodes.",15
"cross-sensor iris recognition: lg4000-to-lg2200 comparison. cross-sensor comparison experimental results reported show procedure defined simulated cross-sensor comparison competition 2013 team migrating / upgrading lg2200 based lg4000 based biometric systems leads better lg4000-to-lg2200 cross-sensor iris recognition results previously reported, terms user comfort terms system safety. hand, lg2200-to-lg400 migration/upgrade procedure defined implemented us applicable solve interoperability issues lg2200 based lg4000 based systems, also pairs systems shift quality acquired images.",5
"improving cross-domain recommendation probabilistic cluster-level latent factor model--extended version. cross-domain recommendation proposed transfer user behavior pattern pooling together rating data multiple domains alleviate sparsity problem appearing single rating domains. however, previous models assume multiple domains share latent common rating pattern based user-item co-clustering. capture diversities among different domains, propose novel probabilistic cluster-level latent factor (pclf) model improve cross-domain recommendation performance. experiments several real world datasets demonstrate proposed model outperforms state-of-the-art methods cross-domain recommendation task.",3
"multi-horizon solar radiation forecasting mediterranean locations using time series models. considering grid manager's point view, needs terms prediction intermittent energy like photovoltaic resource distinguished according considered horizon: following days (d+1, d+2 d+3), next day hourly step (h+24), next hour (h+1) next minutes (m+5 e.g.). work, identified methodologies using time series models prediction horizon global radiation photovoltaic power. present comparison different predictors developed tested propose hierarchy. horizons d+1 h+1, without advanced ad hoc time series pre-processing (stationarity) find easy differentiate autoregressive moving average (arma) multilayer perceptron (mlp). however observed using exogenous variables improves significantly results mlp . shown mlp adapted horizons h+24 m+5. summary, results complementary improve existing prediction techniques innovative tools: stationarity, numerical weather prediction combination, mlp arma hybridization, multivariate analysis, time index, etc.",14
"near-optimal active learning multi-output gaussian processes. paper addresses problem active learning multi-output gaussian process (mogp) model representing multiple types coexisting correlated environmental phenomena. contrast existing works, active learning problem involves selecting informative sampling locations observed also types measurements selected location minimizing predictive uncertainty (i.e., posterior joint entropy) target phenomenon interest given sampling budget. unfortunately, entropy criterion scales poorly numbers candidate sampling locations selected observations optimized. resolve issue, first exploit structure common sparse mogp models deriving novel active learning criterion. then, exploit relaxed form submodularity property new criterion devising polynomial-time approximation algorithm guarantees constant-factor approximation achieved optimal set selected observations. empirical evaluation real-world datasets shows proposed approach outperforms existing algorithms active learning mogp single-output gp models.",18
"quotus: structure political media coverage revealed quoting patterns. given extremely large pool events stories available, media outlets need focus subset issues aspects convey audience. outlets often accused exhibiting systematic bias selection process, different outlets portraying different versions reality. however, absence objective measures empirical evidence, direction extent systematicity remains widely disputed. paper propose framework based quoting patterns quantifying characterizing degree media outlets exhibit systematic bias. apply framework massive dataset news articles spanning six years obama's presidency speeches, reveal systematic pattern indeed emerge outlet's quoting behavior. moreover, show pattern successfully exploited unsupervised prediction setting, determine new quotes outlet select broadcast. encoding bias patterns low-rank space provide analysis structure political media coverage. reveals latent media bias space aligns surprisingly well political ideology outlet type. linguistic analysis exposes striking differences across latent dimensions, showing different types media outlets portray different realities even reporting events. example, outlets mapped mainstream conservative side latent space focus quotes portray presidential persona disproportionately characterized negativity.",3
"image disguise based generative model. protect image contents, existing encryption algorithms designed transform original image texture-like noise-like image, is, however, obvious visual sign indicating presence encrypted image, results significantly large number attacks. solve problem, paper, propose new image encryption method generate visually image original one sending meaning-normal independent image corresponding well-trained generative model achieve effect disguising original image. image disguise method solves problem obvious visual implication, also guarantees security information.",3
"domain control neural machine translation. machine translation systems sensitive domains trained on. several domain adaptation techniques deeply studied. propose new technique neural machine translation (nmt) call domain control performed runtime using unique neural network covering multiple domains. presented approach shows quality improvements compared dedicated domains translating covered domains even out-of-domain data. addition, model parameters need re-estimated domain, making effective real use cases. evaluation carried english-to-french translation two different testing scenarios. first consider case end-user performs translations known domain. secondly, consider scenario domain known predicted sentence level translating. results show consistent accuracy improvements conditions.",3
"molecular de novo design deep reinforcement learning. work introduces method tune sequence-based generative model molecular de novo design augmented episodic likelihood learn generate structures certain specified desirable properties. demonstrate model execute range tasks generating analogues query structure generating compounds predicted active biological target. proof principle, model first trained generate molecules contain sulphur. second example, model trained generate analogues drug celecoxib, technique could used scaffold hopping library expansion starting single molecule. finally, tuning model towards generating compounds predicted active dopamine receptor type 2, model generates structures 95% predicted active, including experimentally confirmed actives included either generative model activity prediction model.",3
"bsvm: banded suport vector machine. describe novel binary classification technique called banded svm (b-svm). standard c-svm formulation cortes et al. (1995), decision rule encouraged lie interval [1, \infty]. new b-svm objective function contains penalty term encourages decision rule lie user specified range [\rho_1, \rho_2]. addition standard set support vectors (svs) near class boundaries, b-svm results second set svs interior class.",18
"descriptive-complexity based distance fuzzy sets. new distance function dist(a,b) fuzzy sets b introduced. based descriptive complexity, i.e., number bits (on average) needed describe element symmetric difference two sets. distance gives amount additional information needed describe one two sets given other. prove mathematical properties perform pattern clustering data based distance.",3
"estimating undirected graphs weak assumptions. consider problem providing nonparametric confidence guarantees undirected graphs weak assumptions. particular, assume sparsity, incoherence normality. allow dimension $d$ increase sample size $n$. first, prove lower bounds show want accurate inferences low assumptions limitations dimension function sample size. dimension increases slowly sample size, show methods based normal approximations bootstrap lead valid inferences provide berry-esseen bounds accuracy normal approximation. dimension large relative sample size, accurate inferences graphs low assumptions possible. instead propose estimate something less demanding entire partial correlation graph. particular, consider: cluster graphs, restricted partial correlation graphs correlation graphs.",10
"learning fly crashing. learn navigate unmanned aerial vehicle (uav) avoid obstacles? one approach use small dataset collected human experts: however, high capacity learning algorithms tend overfit trained little data. alternative use simulation. gap simulation real world remains large especially perception problems. reason research avoids using large-scale real data fear crashes! paper, propose bite bullet collect dataset crashes itself! build drone whose sole purpose crash objects: samples naive trajectories crashes random objects. crash drone 11,500 times create one biggest uav crash dataset. dataset captures different ways uav crash. use negative flying data conjunction positive data sampled trajectories learn simple yet powerful policy uav navigation. show simple self-supervised model quite effective navigating uav even extremely cluttered environments dynamic obstacles including humans. supplementary video see: https://youtu.be/u151hjagkuo",3
"expectation-propogation generative aspect model. generative aspect model extension multinomial model text allows word probabilities vary stochastically across documents. previous results aspect models promising, hindered computational difficulty carrying inference learning. paper demonstrates simple variational methods blei et al (2001) lead inaccurate inferences biased learning generative aspect model. develop alternative approach leads higher accuracy comparable cost. extension expectation-propagation used inference embedded em algorithm learning. experimental results presented synthetic real data sets.",3
"interactive fuzzy goal programming algorithm solve decentralized bi-level multiobjective fractional programming problem. paper proposes fuzzy goal programming based taylor series solving decentralized bi-level multiobjective fractional programming (dblmofp) problem. proposed approach, membership functions associated fuzzy goals objective levels also fractional membership functions converted linear functions using taylor series approach. fuzzy goal programming proposed reach highest degree membership goals taking satisfactory solution decision makers levels. finally, numerical example presented illustrate effectiveness proposed approach.",10
"homoclinic orbits, multiplier spectrum rigidity theorems complex dynamics. aims paper answering several conjectures questions multiplier spectrum rational maps giving new proofs several rigidity theorems complex dynamics, combining tools complex non-archimedean dynamics. remarkable theorem due mcmullen asserts aside flexible latt\`es family, multiplier spectrum periodic points determines conjugacy class rational maps finitely many choices. proof relies thurston's rigidity theorem post-critically finite maps, teichm\""uller theory essential tool. give new proof mcmullen's theorem without using quasiconformal maps teichm\""uller theory. show aside flexible latt\`es family, length spectrum periodic points determines conjugacy class rational maps finitely many choices. generalize aforementioned mcmullen's theorem. also prove rigidity theorem marked length spectrum. similar ideas also yield simple proof rigidity theorem due zdunik. show rational map exceptional one following holds (i) multipliers periodic points contained integer ring imaginary quadratic field; (ii) finitely many periodic points lyapunov exponent. solves two conjectures milnor.",10
"nonlinear kernel support matrix machine matrix learning. many problems supervised tensor learning (stl), real world data face images mri scans naturally represented matrices, also called second order tensors. existing classifiers based tensor representation, support tensor machine (stm) need solve iteratively occupy much time may suffer local minima. paper, present kernel support matrix machine (ksmm) perform supervised learning data represented matrices. ksmm general framework construction matrix-based hyperplane exploit structural information. analyze unifying optimization problem propose asymptotically convergent algorithm. theoretical analysis generalization bounds derived based rademacher complexity respect probability distribution. demonstrate merits proposed method exhaustive experiments simulation study number real-word datasets variety application domains.",18
"statistical inference incomplete ranking data: case rank-dependent coarsening. consider problem statistical inference ranking data, specifically rank aggregation, assumption samples incomplete sense comprising choice alternatives. contrast existing methods, explicitly model process turning full ranking incomplete one, call coarsening process. end, propose concept rank-dependent coarsening, assumes incomplete rankings produced projecting full ranking random subset ranks. concrete instantiation model, full rankings drawn plackett-luce distribution observations take form pairwise preferences, study performance various rank aggregation methods. addition predictive accuracy finite sample setting, address theoretical question consistency, mean ability recover target ranking sample size goes infinity, despite potential bias observations caused (unknown) coarsening.",18
"multi-scale saliency detection using dictionary learning. saliency detection drawn lot attention researchers various fields past several years. saliency perceptual quality makes object, person draw attention humans sight. salient object detection image used centrally many computational photography computer vision applications like video compression, object recognition classification, object segmentation, adaptive content delivery, motion detection, content aware resizing, camouflage images change blindness images name few. propose method detect saliency objects using multimodal dictionary learning recently used classification image fusion. multimodal dictionary learning task driven gives improved performance counterpart (the one task specific).",3
"exploring syntactic abilities rnns multi-task learning. recent work explored syntactic abilities rnns using subject-verb agreement task, diagnoses sensitivity sentence structure. rnns performed task well common cases, faltered complex sentences (linzen et al., 2016). test whether errors due inherent limitations architecture relatively indirect supervision provided agreement dependencies corpus. trained single rnn perform agreement task additional task, either ccg supertagging language modeling. multi-task training led significantly lower error rates, particular complex sentences, suggesting rnns ability evolve sophisticated syntactic representations shown before. also show easily available agreement training data improve performance syntactic tasks, particular limited amount training data available tasks. multi-task paradigm also leveraged inject grammatical knowledge language models.",3
"disentangling factors variation mixing. propose unsupervised approach learn image representations consist disentangled factors variation. factor variation corresponds image attribute discerned consistently across set images, pose color objects. disentangled representation consists concatenation feature chunks, chunk representing factor variation. supports applications transferring attributes one image another, simply swapping feature chunks, classification retrieval based one several attributes, considering user specified subset feature chunks. learn representation unsupervised manner, without labeling knowledge data domain, using autoencoder architecture two novel training objectives: first, propose invariance objective encourage encoding attribute, decoding chunk, invariant changes attributes chunks, respectively, second, include classification objective, ensures chunk corresponds consistently discernible attribute represented image, hence avoiding shortcut chunks ignored completely. demonstrate effectiveness approach mnist, sprites, celeba datasets.",3
"marginal value adaptive gradient methods machine learning. adaptive optimization methods, perform local optimization metric constructed history iterates, becoming increasingly popular training deep neural networks. examples include adagrad, rmsprop, adam. show simple overparameterized problems, adaptive methods often find drastically different solutions gradient descent (gd) stochastic gradient descent (sgd). construct illustrative binary classification problem data linearly separable, gd sgd achieve zero test error, adagrad, adam, rmsprop attain test errors arbitrarily close half. additionally study empirical generalization capability adaptive methods several state-of-the-art deep learning models. observe solutions found adaptive methods generalize worse (often significantly worse) sgd, even solutions better training performance. results suggest practitioners reconsider use adaptive methods train neural networks.",18
"graph regularized nonnegative matrix factorization hyperspectral data unmixing. spectral unmixing important tool hyperspectral data analysis estimating endmembers abundance fractions mixed pixel. paper examines applicability recently developed algorithm called graph regularized nonnegative matrix factorization (gnmf) aim. proposed approach exploits intrinsic geometrical structure data besides considering positivity full additivity constraints. simulated data based measured spectral signatures, used evaluating proposed algorithm. results terms abundance angle distance (aad) spectral angle distance (sad) show method effectively unmix hyperspectral data.",3
"liberating language research dogmas 20th century. commentary article ""large-scale evidence dependency length minimization 37 languages"" futrell, mahowald & gibson (pnas 2015 112 (33) 10336-10341).",3
"football group draw probabilities corrections. paper considers challenge designing football group draw mechanisms uniform distribution valid draw assignments, also entertaining, practical, transparent. explain simulate fifa sequential draw method, compute non-uniformity draws comparison uniform rejection sampler. propose two practical methods achieving uniform distribution still using balls bowls way suitable televised draw. solutions also tried interactively.",18
"end-to-end fine-grained action segmentation recognition using conditional random field models discriminative sparse coding. fine-grained action segmentation recognition important yet challenging task. given long, untrimmed sequence kinematic data, task classify action time frame segment time series correct sequence actions. paper, propose novel framework combines temporal conditional random field (crf) model powerful frame-level representation based discriminative sparse coding. introduce end-to-end algorithm jointly learning weights crf model, include action classification action transition costs, well overcomplete dictionary mid-level action primitives. results crf model driven sparse coding features obtained using discriminative dictionary shared among different actions adapted task structured output learning. evaluate method three surgical tasks using kinematic data jigsaws dataset, well food preparation task using accelerometer data 50 salads dataset. results show proposed method performs par better state-of-the-art methods.",3
"active learning algorithms graphical model selection. problem learning structure high dimensional graphical model data received considerable attention recent years. many applications sensor networks proteomics often expensive obtain samples variables involved simultaneously. instance, might involve synchronization large number sensors tagging large number proteins. address important issue, initiate study novel graphical model selection problem, goal optimize total number scalar samples obtained allowing collection samples subsets variables. propose general paradigm graphical model selection feedback used guide sampling high degree vertices, obtaining samples ones low degrees. instantiate framework two specific active learning algorithms, one makes mild assumptions computationally expensive, computationally efficient requires stronger (nevertheless standard) assumptions. whereas sample complexity passive algorithms typically function maximum degree graph, show sample complexity algorithms provable smaller depends novel local complexity measure akin average degree graph. finally demonstrate efficacy framework via simulations.",18
"perturbation theory lss $λ$cdm universe: exact time evolution two-loop power spectrum. derive exact analytic solutions density velocity fields orders eulerian perturbation theory $\lambda$cdm cosmology. particular, show density velocity field kernels written separable form time momenta perturbative order. kernel solutions built analytic basis momentum operators time-dependent coefficients, solve set recursive differential equations. also provide exact closed perturbative solution coefficients, expanding around (quasi-)eds approximation. find perturbative solution rapidly converges towards numerically obtained solutions leading order result suffices practical requirements. illustrate findings, compute exact two-loop dark matter density velocity power spectra $\lambda$cdm cosmology. show difference exact $\lambda$cdm (quasi-)eds approximated result reach level several percent. deviation partially mitigated exploiting degeneracy eft counterterms. additional benefit algorithm solutions time-dependent coefficients, computational complexity power spectra loops $\lambda$cdm reduced level eds case. performing two-loop computation, devise explicit method implement so-called ir cancellations, well cancellations arising consequence mass momentum conservation.",0
"convex optimization linear query processing approximate differential privacy. differential privacy enables organizations collect accurate aggregates sensitive data strong, rigorous guarantees individuals' privacy. previous work found differential privacy, computing multiple correlated aggregates batch, using appropriate \emph{strategy}, may yield higher accuracy computing independently. however, finding best strategy maximizes result accuracy non-trivial, involves solving complex constrained optimization program appears non-linear non-convex. hence, past much effort devoted solving non-convex optimization program. existing approaches include various sophisticated heuristics expensive numerical solutions. none them, however, guarantees find optimal solution optimization problem. paper points ($\epsilon$, $\delta$)-differential privacy, optimal solution constrained optimization problem search suitable strategy found, rather surprisingly, solving simple elegant convex optimization program. then, propose efficient algorithm based newton's method, prove always converge optimal solution linear global convergence rate quadratic local convergence rate. empirical evaluations demonstrate accuracy efficiency proposed solution.",3
"advances artificial intelligence: sure, right track?. past decade, ai made remarkable progress. agreed due recently revived deep learning technology. deep learning enables process large amounts data using simplified neuron networks simulate way brain works. however, different point view, posits brain processing information, data. unresolved duality hampered ai progress years. paper, propose notion integrated information hopefully resolve problem. consider integrated information coupling two separate entities - physical information (that implies data processing) semantic information (that provides physical information interpretation). regard, intelligence becomes product information processing. extending line thinking, said information processing require human brain implementation. indeed, bacteria amoebas exhibit intelligent behavior without sign brain. dramatically removes need ai systems emulate human brain complexity! paper tries explore shift ai systems design philosophy.",3
"chains rotating boson stars. boson stars stationary, axially symmetric solutions complex scalar field theory coupled gravity. recently, multi-solitonic configurations interpreted static chains multiple boson stars bound gravity carrying angular momentum reported. propose generalize solutions stationary case constructing chains rotating boson stars analyze properties. non-linear elliptic field equations solved using finite element method. find chains even number constituents exhibit spiral-like frequency dependence mass, angular momentum noether charge single boson stars. contrast, sequences chains odd number constituents show nontrivial loops starting ending flat vacuum. consequence, solutions cannot uniquely parametrized single parameter. conjecture rotating chains correspond excitations single boson stars pairs them. also analyze flat space limit find reduce chains $q$-balls.",6
"strong convergence tamed euler scheme scalar sdes superlinearly growing discontinuous drift coefficient. paper, consider scalar stochastic differential equations (sdes) superlinearly growing piecewise continuous drift coefficient. existence uniqueness strong solutions sdes obtained. furthermore, classical $l_p$-error rate 1/2 $p\in [1, +\infty)$ recovered tamed euler scheme. numerical example provided support conclusion.",10
"linking search space structure, run-time dynamics, problem difficulty: step toward demystifying tabu search. tabu search one effective heuristics locating high-quality solutions diverse array np-hard combinatorial optimization problems. despite widespread success tabu search, researchers poor understanding many key theoretical aspects algorithm, including models high-level run-time dynamics identification search space features influence problem difficulty. consider questions context job-shop scheduling problem (jsp), domain tabu search algorithms shown remarkably effective. previously, demonstrated mean distance random local optima nearest optimal solution highly correlated problem difficulty well-known tabu search algorithm jsp introduced taillard. paper, discuss various shortcomings measure develop new model problem difficulty corrects deficiencies. show taillards algorithm modeled high fidelity simple variant straightforward random walk. random walk model accounts nearly variability cost required locate optimal sub-optimal solutions random jsps, provides explanation differences difficulty random versus structured jsps. finally, discuss empirically substantiate two novel predictions regarding tabu search algorithm behavior. first, method constructing initial solution highly unlikely impact performance tabu search. second, tabu tenure selected small possible simultaneously avoiding search stagnation; values larger necessary lead significant degradations performance.",3
"qualitative approximate behavior composition. behavior composition problem involves automatically building controller able realize desired, unavailable, target system (e.g., house surveillance) suitably coordinating set available components (e.g., video cameras, blinds, lamps, vacuum cleaner, phones, etc.) previous work almost exclusively aimed bringing desired component totality, highly unsatisfactory unsolvable problems. work, develop approach approximate behavior composition without departing classical setting, thus making problem applicable much wider range cases. based notion simulation, characterize maximal controller ""closest"" implementable target module (optimal approximation) are, show computed using atl model checking technology special case. show uniqueness optimal approximations, prove soundness completeness respect imported controllers.",3
"rnade: real-valued neural autoregressive density-estimator. introduce rnade, new model joint density estimation real-valued vectors. model calculates density datapoint product one-dimensional conditionals modeled using mixture density networks shared parameters. rnade learns distributed representation data, tractable expression calculation densities. tractable likelihood allows direct comparison methods training standard gradient-based optimizers. compare performance rnade several datasets heterogeneous perceptual data, finding outperforms mixture models one case.",18
"aspect-based opinion summarization convolutional neural networks. paper considers aspect-based opinion summarization (aos) reviews particular products. enable real applications, aos system needs address two core subtasks, aspect extraction sentiment classification. existing approaches aspect extraction, use linguistic analysis topic modeling, general across different products precise enough suitable particular products. instead take less general precise scheme, directly mapping review sentence pre-defined aspects. tackle aspect mapping sentiment classification, propose two convolutional neural network (cnn) based methods, cascaded cnn multitask cnn. cascaded cnn contains two levels convolutional networks. multiple cnns level 1 deal aspect mapping task, single cnn level 2 deals sentiment classification. multitask cnn also contains multiple aspect cnns sentiment cnn, different networks share word embeddings. experimental results indicate cascaded multitask cnns outperform svm-based methods large margins. multitask cnn generally performs better cascaded cnn.",3
"weighted kendall high-order kernels permutations. propose new positive definite kernels permutations. first introduce weighted version kendall kernel, allows weight unequally contributions different item pairs permutations depending ranks. like kendall kernel, show weighted version invariant relabeling items computed efficiently $o(n \ln(n))$ operations, $n$ number items permutation. second, propose supervised approach learn weights jointly optimizing function estimated kernel machine. third, kendall kernel considers pairwise comparison items, extend considering higher-order comparisons among tuples items show supervised approach learning weights systematically generalized higher-order permutation kernels.",18
"matrix product state higher-order tensor compression classification. paper introduces matrix product state (mps) decomposition new systematic method compress multidimensional data represented higher-order tensors. solves two major bottlenecks tensor compression: computation compression quality. regardless tensor order, mps compresses tensors matrices moderate dimension used classification. mainly based successive sequence singular value decompositions (svd), mps quite simple implement arrives global optimal matrix, bypassing local alternating optimization, computationally expensive cannot yield global solution. benchmark results show mps achieve better classification performance favorable computation cost compared tensor compression methods.",18
"predicting surgery duration neural heteroscedastic regression. scheduling surgeries challenging task due fundamental uncertainty clinical environment, well risks costs associated under- over-booking. investigate neural regression algorithms estimate parameters surgery case durations, focusing issue heteroscedasticity. seek simultaneously estimate duration surgery, well surgery-specific notion uncertainty duration. estimating uncertainty lead nuanced effective scheduling strategies, able schedule surgeries efficiently allowing informed case-specific margin error. using surgery records %from uc san diego health system, large united states health system demonstrate potential improvements order 20% (in terms minutes overbooked) compared current scheduling techniques. moreover, demonstrate surgery durations indeed heteroscedastic. show models estimate case-specific uncertainty better fit data (log likelihood). additionally, show heteroscedastic predictions optimally trade under-booking minutes, especially idle minutes scheduling collisions confer disparate costs.",18
"cell detection microscopy images deep convolutional neural network compressed sensing. ability automatically detect certain types cells cellular subunits microscopy images significant interest wide range biomedical research clinical practices. cell detection methods evolved employing hand-crafted features deep learning-based techniques. essential idea methods cell classifiers detectors trained pixel space, locations target cells labeled. paper, seek different route propose convolutional neural network (cnn)-based cell detection method uses encoding output pixel space. cell detection problem, output space sparsely labeled pixel locations indicating cell centers. employ random projections encode output space compressed vector fixed dimension. then, cnn regresses compressed vector input pixels. furthermore, possible stably recover sparse cell locations output pixel space predicted compressed vector using $l_1$-norm optimization. past, output space encoding using compressed sensing (cs) used conjunction linear non-linear predictors. best knowledge, first successful use cnn cs-based output space encoding. made substantial experiments several benchmark datasets, proposed cnn + cs framework (referred cnncs) achieved highest least top-3 performance terms f1-score, compared state-of-the-art methods.",3
"stock prediction: method based extraction news features recurrent neural networks. paper proposed method stock prediction. terms feature extraction, extract features stock-related news besides stock prices. first select seed words based experience symbols good news bad news. propose optimization method calculate positive polar words. that, construct features news based positive polar words. consideration sequential stock prices continuous news effects, propose recurrent neural network model help predict stock prices. compared svm classifier price features, find proposed method 5% improvement stock prediction accuracy experiments.",16
"steady-state entanglement generation non-degenerate qubits. propose scheme dissipatively produce steady-state entanglement two-qubit system, via interaction bosonic mode. system driven stationary entangled state, compensate mode dissipation injecting energy via coherent pump field. also present scheme allows us adiabatically transfer population desired entangled state. dynamics leading entangled state schemes understood analogy electromagnetically induced transparency (eit) stimulated raman adiabatic passage (stirap), respectively.",17
"stochastic gradient descent approximate bayesian inference. stochastic gradient descent constant learning rate (constant sgd) simulates markov chain stationary distribution. perspective, derive several new results. (1) show constant sgd used approximate bayesian posterior inference algorithm. specifically, show adjust tuning parameters constant sgd best match stationary distribution posterior, minimizing kullback-leibler divergence two distributions. (2) demonstrate constant sgd gives rise new variational em algorithm optimizes hyperparameters complex probabilistic models. (3) also propose sgd momentum sampling show adjust damping coefficient accordingly. (4) analyze mcmc algorithms. langevin dynamics stochastic gradient fisher scoring, quantify approximation errors due finite learning rates. finally (5), use stochastic process perspective give short proof polyak averaging optimal. based idea, propose scalable approximate mcmc algorithm, averaged stochastic gradient sampler.",18
"power localization efficiently learning linear separators noise. introduce new approach designing computationally efficient learning algorithms tolerant noise, demonstrate effectiveness designing algorithms improved noise tolerance guarantees learning linear separators. consider malicious noise model adversarial label noise model. malicious noise, adversary corrupt label features, provide polynomial-time algorithm learning linear separators $\re^d$ isotropic log-concave distributions tolerate nearly information-theoretically optimal noise rate $\eta = \omega(\epsilon)$. adversarial label noise model, distribution feature vectors unchanged, overall probability noisy label constrained $\eta$, also give polynomial-time algorithm learning linear separators $\re^d$ isotropic log-concave distributions handle noise rate $\eta = \omega\left(\epsilon\right)$. show that, active learning model, algorithms achieve label complexity whose dependence error parameter $\epsilon$ polylogarithmic. provides first polynomial-time active learning algorithm learning linear separators presence malicious noise adversarial label noise.",3
"towards theory good sat representations. aim providing foundation theory ""good"" sat representations f boolean functions f. argue hierarchy uc_k unit-refutation complete clause-sets level k, introduced authors, provides basic target classes, is, f uc_k achieved k small feasible. f contain new variables, i.e., f equivalent (as cnf) f, f uc_1 similar ""achieving (generalised) arc consistency"" known literature (it somewhat weaker, theoretically much nicer handle). show polysize representations boolean functions sense, hierarchy uc_k strict. boolean functions separations ""doped"" minimally unsatisfiable clause-sets deficiency 1; functions introduced [sloan, soerenyi, turan, 2007], generalise construction show correspondence strengthened notion irredundant sub-clause-sets. turning lower bounds upper bounds, believe many common cnf representations fit uc_k scheme, give basic tools construct representations uc_1 new variables, based tseitin translation. note regarding new variables uc_1-representations stronger mere ""arc consistency"", since new variables excluded consideration.",3
"structure evolutionary exploration: crossover, buildings blocks estimation-of-distribution algorithms. notion building blocks related structure offspring probability distribution: loci variability strongly correlated constitute building block. call correlated exploration. background analyze structure offspring probability distribution, exploration distribution, ga mutation only, crossover ga, estimation-of-distribution algorithm (eda). results allow precise characterization structure crossover exploration distribution. essentially, crossover operator destroys mutual information loci transforming entropy; inverse correlated exploration. contrast, objective edas model mutual information loci fitness distribution thereby induce correlated exploration.",12
"genetic programming kernel-based learning co-evolving subsets selection. support vector machines (svms) well-established machine learning (ml) algorithms. rely fact i) linear learning formalized well-posed optimization problem; ii) non-linear learning brought linear learning thanks kernel trick mapping initial search space onto high dimensional feature space. kernel designed ml expert governs efficiency svm approach. paper, new approach automatic design kernels genetic programming, called evolutionary kernel machine (ekm), presented. ekm combines well-founded fitness function inspired margin criterion, co-evolution framework ensuring computational scalability approach. empirical validation standard ml benchmark demonstrates ekm competitive using state-of-the-art svms tuned hyper-parameters.",3
"pattern-coupled sparse bayesian learning recovery block-sparse signals. consider problem recovering block-sparse signals whose structures unknown \emph{a priori}. block-sparse signals nonzero coefficients occurring clusters arise naturally many practical scenarios. however, knowledge block structure usually unavailable practice. paper, develop new sparse bayesian learning method recovery block-sparse signals unknown cluster patterns. specifically, pattern-coupled hierarchical gaussian prior model introduced characterize statistical dependencies among coefficients, set hyperparameters employed control sparsity signal coefficients. unlike conventional sparse bayesian learning framework individual hyperparameter associated independently coefficient, paper, prior coefficient involves hyperparameter, also hyperparameters immediate neighbors. way, sparsity patterns neighboring coefficients related hierarchical model potential encourage structured-sparse solutions. hyperparameters, along sparse signal, learned maximizing posterior probability via expectation-maximization (em) algorithm. numerical results show proposed algorithm presents uniform superiority existing methods series experiments.",3
"exact sampling determinantal point processes. determinantal point processes (dpps) important concept random matrix theory combinatorics. also recently attracted interest study numerical methods machine learning, offer elegant ""missing link"" independent monte carlo sampling deterministic evaluation regular grids, applicable general set spaces. helpful whenever algorithm *explores* reduce uncertainty, active learning, bayesian optimization, reinforcement learning, marginalization graphical models. draw samples dpp practice, existing literature focuses approximate schemes low cost, comparably inefficient exact algorithms like rejection sampling. point that, many settings relevance machine learning, also possible draw *exact* samples dpps continuous domains. start intuitive example real line, generalized multivariate real vector spaces. also compare previously studied approximations, showing exact sampling, despite higher cost, preferable precision needed.",3
"bit progress language modeling. past several years, number different language modeling improvements simple trigram models found, including caching, higher-order n-grams, skipping, interpolated kneser-ney smoothing, clustering. present explorations variations on, limits of, techniques, including showing sentence mixture models may potential. techniques studied separately, rarely studied combination. find significant interactions, especially smoothing clustering techniques. compare combination techniques together katz smoothed trigram model count cutoffs. achieve perplexity reductions 38% 50% (1 bit entropy), depending training data size, well word error rate reduction 8.9%. perplexity reductions perhaps highest reported compared fair baseline. extended version paper; contains additional details proofs, designed good introduction state art language modeling.",3
"constraining pseudo-dirac neutrinos galactic core-collapse supernova. neutrinos pseudo-dirac nature -- majorana fermions behaving dirac fermions practical purposes. scenario, active sterile neutrinos quasi-degenerate mass, hence oscillations two, due tiny mass-squared difference $(\delta m^2)$, develop long baselines. hypothesis, analyze neutrino data sn1987a, find mild preference non-zero mass-squared difference. data also used exclude values $\delta m^2\sim 10^{-20}{\rm ev}^2$ - smallest constrained far. also discuss next-generation experiments like dune hyper-kamiokande probe scenario future galactic supernova.",8
"correspondence insertion as-projective-as-possible image stitching. spatially varying warps increasingly popular image alignment. particular, as-projective-as-possible (apap) warps proven effective accurate panoramic stitching, especially cases significant depth parallax defeat standard homographic warps. however, estimating spatially varying warps requires sufficient number feature matches. image regions feature detection matching fail, warp loses guidance unable accurately model true underlying warp, thus resulting poor registration. paper, propose correspondence insertion method apap warps, focus panoramic stitching. method automatically identifies misaligned regions, inserts appropriate point correspondences increase flexibility warp improve alignment. unlike warp varieties, underlying projective regularization apap warps reduces overfitting geometric distortion, despite increases warp complexity. comparisons recent techniques parallax-tolerant image stitching demonstrate effectiveness simplicity approach.",3
"joint spatial-angular sparse coding dmri separable dictionaries. diffusion mri (dmri) provides ability reconstruct neuronal fibers brain, $\textit{in vivo}$, measuring water diffusion along angular gradient directions $q$-space. high angular resolution diffusion imaging (hardi) produce better estimates fiber orientation popularly used diffusion tensor imaging, high number samples needed estimate diffusivity requires lengthy patient scan times. accelerate dmri, compressed sensing (cs) utilized exploiting sparse dictionary representation data, discovered sparse coding. sparser representation, fewer samples needed reconstruct high resolution signal limited information loss, important area research focused finding sparsest possible representation dmri. current reconstruction methods however, rely angular representation $\textit{per voxel}$ added spatial regularization, so, global level sparsity less number voxels. therefore, state-of-the-art dmri cs frameworks may fundamental limit rate acceleration achieved. contrast, propose joint spatial-angular representation dmri allow us achieve levels global sparsity number voxels. major challenge, however, computational complexity solving global sparse coding problem large-scale dmri. work, present novel adaptations popular sparse coding algorithms become better suited solving large-scale problems exploiting spatial-angular separability. experiments show method achieves significantly sparser representations hardi state-of-the-art potential increase hardi acceleration new levels.",18
"rows vs columns linear systems equations - randomized kaczmarz coordinate descent?. paper randomized iterative algorithms solving linear system equations $x \beta = y$ different settings. recent interest topic reignited strohmer vershynin (2009) proved linear convergence rate randomized kaczmarz (rk) algorithm works rows $x$ (data points). following that, leventhal lewis (2010) proved linear convergence randomized coordinate descent (rcd) algorithm works columns $x$ (features). aim paper simplify understanding two algorithms, establish direct relationships (though rk often compared stochastic gradient descent), examine algorithmic commonalities tradeoffs involved working rows columns. also discuss kernel ridge regression present kaczmarz-style algorithm works data points advantage solving problem without ever storing forming gram matrix, one recognized problems encountered scaling kernelized methods.",10
"revealing cluster structure graph path following replicator dynamic. paper, propose path following replicator dynamic, investigate potentials uncovering underlying cluster structure graph. proposed dynamic generalization discrete replicator dynamic. replicator dynamic successfully used extract dense clusters graphs; however, often sensitive degree distribution graph, usually biased vertices large degrees, thus may fail detect densest cluster. overcome problem, introduce dynamic parameter, called path parameter, evolution process. path parameter interpreted maximal possible probability current cluster containing vertex, monotonically increases evolution process proceeds. limiting maximal probability, phenomenon vertices dominating early stage evolution process suppressed, thus making evolution process robust. solve optimization problem fixed path parameter, propose efficient fixed point algorithm. time complexity path following replicator dynamic linear number edges graph, thus analyze graphs millions vertices tens millions edges common pc minutes. besides, naturally generalized hypergraph graph edges different orders. apply four important problems: maximum clique problem, densest k-subgraph problem, structure fitting, discovery high-density regions. extensive experimental results clearly demonstrate advantages, terms robustness, scalability flexility.",3
"investigation applying acoustic feature conversion asr adult child speech. performance child speech recognition generally less satisfactory compared adult speech due limited amount training data. significant performance degradation expected applying automatic speech recognition (asr) system trained adult speech child speech directly, result domain mismatch. present study focused adult-to-child acoustic feature conversion alleviate mismatch. different acoustic feature conversion approaches, including deep neural network based signal processing based, investigated compared fair experimental setting, converted acoustic features amount labeled adult speech used train asr models scratch. experimental results reveal conversion methods lead asr performance gain. specifically, classic unsupervised domain adaptation method, statistic matching show effectiveness. disentanglement-based auto-encoder (dae) conversion framework found useful approach f0 normalization achieves best performance. noted f0 distribution converted features important attribute reflect conversion quality, utilizing adult-child deep classification model make judgment shown inappropriate.",5
"contour manifolds optimal transport. describing shapes suitable measures object segmentation, proposed [24], allows combine advantages representations parametrized contours indicator functions. pseudo-riemannian structure optimal transport used model shapes ways similar contours, kantorovich functional enables application convex optimization methods global optimality segmentation functional. paper provide mathematical study shape measure representation relation contour description. particular show pseudo-riemannian structure optimal transport, restricted set shape measures, yields manifold diffeomorphic manifold closed contours. discussion metric induced optimal transport corresponding geodesic equation given.",10
"toward example-based machine translation written text asl using virtual agent animation. modern computational linguistic software cannot produce important aspects sign language translation. using researches deduce majority automatic sign language translation systems ignore many aspects generate animation; therefore interpretation lost truth information meaning. goals are: translate written text language asl animation; model maximum raw information using machine learning computational techniques; produce adapted expressive form natural looking understandable asl animations. methods include linguistic annotation initial text semantic orientation generate facial expression. use genetic algorithms coupled learning/recognized systems produce natural form. detect emotion based fuzzy logic produce degree interpolation facial expressions. roughly, present new expressive language text adapted sign modeling language tasml describes maximum aspects related natural sign language interpretation. paper organized follow: next section devoted present comprehension effect using space/time/svo form asl animation based experimentation. section 3, describe technical considerations. present general approach adopted develop tool section 4. finally, give perspectives future works.",3
"automatic identification altlexes using monolingual parallel corpora. automatic identification discourse relations still challenging task natural language processing. discourse connectives, ""since"" ""but"", informative cues identify explicit relations; however discourse parsers typically use closed inventory connectives. result, discourse relations signaled markers outside inventories (i.e. altlexes) detected effectively. paper, propose novel method leverage parallel corpora text simplification lexical resources automatically identify alternative lexicalizations signal discourse relation. applied simple wikipedia newsela corpora along wordnet ppdb, method allowed automatic discovery 91 altlexes.",3
"online alternating direction method. online optimization emerged powerful tool large scale optimization. paper, introduce efficient online algorithms based alternating directions method (adm). introduce new proof technique adm batch setting, yields o(1/t) convergence rate adm forms basis regret analysis online setting. consider two scenarios online setting, based whether solution needs lie feasible set not. settings, establish regret bounds objective function well constraint violation general strongly convex functions. preliminary results presented illustrate performance proposed algorithms.",3
"fitness level method tail bounds. fitness-level method, also called method f-based partitions, intuitive widely used technique running time analysis randomized search heuristics. originally defined prove upper lower bounds expected running time. recently, upper tail bounds added technique; however, tail bounds apply running times least twice large expectation. remove restriction supplement fitness-level method sharp tail bounds, including lower tails. exemplary application, prove running time randomized local search onemax sharply concentrated around n ln n - 0.1159 n.",3
"unified convex surrogate schatten-$p$ norm. schatten-$p$ norm ($0<p<1$) widely used replace nuclear norm better approximating rank function. however, existing methods either 1) scalable large scale problems due relying singular value decomposition (svd) every iteration, 2) specific $p$ values, e.g., $1/2$, $2/3$. paper, show $p$, $p_1$, $p_2 >0$ satisfying $1/p=1/p_1+1/p_2$, equivalence schatten-$p$ norm one matrix schatten-$p_1$ schatten-$p_2$ norms two factor matrices. extend equivalence multiple factor matrices show factor norms convex smooth $p>0$. contrast, original schatten-$p$ norm $0<p<1$ non-convex non-smooth. example conduct experiments matrix completion. utilize convexity factor matrix norms, adopt accelerated proximal alternating linearized minimization algorithm establish sequence convergence. experiments synthetic real datasets exhibit superior performance state-of-the-art methods. speed also highly competitive.",18
"multi-shot person re-identification via relational stein divergence. person re-identification particularly challenging due significant appearance changes across separate camera views. order re-identify people, representative human signature effectively handle differences illumination, pose camera parameters. general appearance-based methods modelled euclidean spaces, argued applications image video analysis better modelled via non-euclidean manifold geometry. end, recent approaches represent images covariance matrices, interpret matrices points riemannian manifolds. direct classification manifolds difficult, paper propose represent manifold point vector similarities class representers, via recently introduced form bregman matrix divergence known stein divergence. followed using discriminative mapping similarity vectors final classification. use similarity vectors contrast traditional approach embedding manifolds tangent spaces, suffer representing manifold structure inaccurately. comparative evaluations benchmark ethz ilids datasets person re-identification task show proposed approach obtains better performance recent techniques histogram plus epitome, partial least squares, symmetry-driven accumulation local features.",3
"actively learning makes discrete sequence valid. deep learning techniques hugely successful traditional supervised unsupervised machine learning problems. large part, techniques solve continuous optimization problems. recently however, discrete generative deep learning models successfully used efficiently search high-dimensional discrete spaces. methods work representing discrete objects sequences, powerful sequence-based deep models employed. unfortunately, techniques significantly hindered fact generative models often produce invalid sequences. step towards solving problem, propose learn deep recurrent validator model. given partial sequence, model learns probability sequence occurring beginning full valid sequence. thus identifies valid versus invalid sequences crucially also provides insight individual sequence elements influence validity discrete objects. learn model propose approach inspired seminal work bayesian active learning. synthetic dataset, demonstrate ability model distinguish valid invalid sequences. believe key step toward learning generative models faithfully produce valid discrete objects.",18
"morphologic knowledge dynamics: revision, fusion, abduction. several tasks artificial intelligence require able find models knowledge dynamics. include belief revision, fusion belief merging, abduction. paper exploit algebraic framework mathematical morphology context propositional logic, define operations dilation erosion set formulas. derive concrete operators, based semantic approach, intuitive interpretation formally well behaved, perform revision, fusion abduction. computation tractability addressed, simple examples illustrate typical results obtained.",3
"privacy preserving id3 horizontally, vertically grid partitioned data. consider privacy preserving decision tree induction via id3 case training data horizontally vertically distributed. furthermore, consider problem case data horizontally vertically distributed, situation refer grid partitioned data. give algorithm privacy preserving id3 horizontally partitioned data involving two parties. grid partitioned data, discuss two different evaluation methods preserving privacy id3, namely, first merging horizontally developing vertically first merging vertically next developing horizontally. next introducing privacy preserving data mining grid-partitioned data, main contribution paper show, means complexity analysis former evaluation method efficient.",3
"chain enumeration, partition lattices polynomials real roots. coefficients chain polynomial finite poset enumerate chains poset number elements. chain polynomials partition lattices standard type $b$ analogues shown real roots. real-rootedness chain polynomial conjectured geometric lattices shown preserved pyramid prism operations cohen--macaulay posets. result, new families convex polytopes whose face lattices real-rooted chain polynomials presented. application face enumeration second barycentric subdivision boundary complex simplex also included.",10
"improved style transfer respecting inter-layer correlations. popular series style transfer methods apply style content image controlling mean covariance values early layers feature stack. insufficient transferring styles strong structure across spatial scales like, e.g., textures dots lie long curves. paper demonstrates controlling inter-layer correlations yields visible improvements style transfer methods. achieve control computing cross-layer, rather within-layer, gram matrices. find (a) cross-layer gram matrices sufficient control within-layer statistics. inter-layer correlations improves style transfer texture synthesis. paper shows numerous examples ""hard"" real style transfer problems (e.g. long scale hierarchical patterns); (b) fast approximate style transfer method control cross-layer gram matrices; (c) demonstrate multiplicative, rather additive style content loss, results good style transfer. multiplicative loss produces visible emphasis boundaries, means one hyper-parameter eliminated.",3
"semi-supervised learning deep rendering mixture model. semi-supervised learning algorithms reduce high cost acquiring labeled training data using labeled unlabeled data learning. deep convolutional networks (dcns) achieved great success supervised tasks widely employed semi-supervised learning. paper leverage recently developed deep rendering mixture model (drmm), probabilistic generative model models latent nuisance variation, whose inference algorithm yields dcns. develop em algorithm drmm learn labeled unlabeled data. guided theory drmm, introduce novel non-negativity constraint variational inference term. report state-of-the-art performance mnist svhn competitive results cifar10. also probe deeper drmm trained semi-supervised setting represents latent nuisance variation using synthetically rendered images. taken together, work provides unified framework supervised, unsupervised, semi-supervised learning.",18
"rhetorical relations information retrieval. typically, every part coherent text plausible reason presence, function performs overall semantics text. rhetorical relations, e.g. contrast, cause, explanation, describe parts text linked other. knowledge socalled discourse structure applied successfully several natural language processing tasks. work studies use rhetorical relations information retrieval (ir): correlation certain rhetorical relations retrieval performance? knowledge document's rhetorical relations useful ir? present language model modification considers rhetorical relations estimating relevance document query. empirical evaluation different versions model trec settings shows certain rhetorical relations benefit retrieval effectiveness notably (> 10% mean average precision state-of-the-art baseline).",3
"fast feature reduction intrusion detection datasets. intrusion detection systems (ids), system tries learn characteristics different type attacks analyzing packets sent received network. packets lot features. required analyzed detect specific type attack. detection speed computational cost another vital matter here, types problems, datasets huge regularly. paper tried propose simple fast feature selection method eliminate features helpful information them. result faster learning process redundant feature omission. compared proposed method three successful similarity based feature selection algorithm including correlation coefficient, least square regression error maximal information compression index. used recommended features algorithms two popular classifiers including: bayes knn classifier measure quality recommendations. experimental result shows although proposed method can't outperform evaluated algorithms high differences accuracy, computational cost huge superiority them.",3
"vista: visually, socially, temporally-aware model artistic recommendation. understanding users' interactions highly subjective content---like artistic images---is challenging due complex semantics guide preferences. one hand one overcome `standard' recommender systems challenges, dealing large, sparse, long-tailed datasets. other, several new challenges present themselves, need model content terms visual appearance, even social dynamics, preference toward particular artist independent art create. paper build large-scale recommender systems model dynamics vibrant digital art community, behance, consisting tens millions interactions (clicks `appreciates') users toward digital art. methodologically, main contributions model (a) rich content, especially terms visual appearance; (b) temporal dynamics, terms users prefer `visually consistent' content within across sessions; (c) social dynamics, terms users exhibit preferences towards certain art styles, well artists themselves.",3
"theoretical analysis noisy sparse subspace clustering dimensionality-reduced data. subspace clustering problem partitioning unlabeled data points number clusters data points within one cluster lie approximately low-dimensional linear subspace. many practical scenarios, dimensionality data points clustered compressed due constraints measurement, computation privacy. paper, study theoretical properties popular subspace clustering algorithm named sparse subspace clustering (ssc) establish formal success conditions ssc dimensionality-reduced data. analysis applies general fully deterministic model underlying subspaces data points within subspace deterministically positioned, also wide range dimensionality reduction techniques (e.g., gaussian random projection, uniform subsampling, sketching) fall subspace embedding framework (meng & mahoney, 2013; avron et al., 2014). finally, apply analysis differentially private ssc algorithm established privacy utility guarantees proposed method.",18
"probabilistic semantics defaults. much interest providing probabilistic semantics defaults approaches seem suffer one two problems: either require numbers, problem defaults intended avoid, generate peculiar side effects. rather provide semantics defaults, address problem defaults intended solve: reasoning uncertainty numeric probability distributions available. describe non-numeric formalism called inference graph based standard probability theory, conditional independence sentences favouring favours b - favours(a, b) - p(a|b) > p(a). formalism seems handle examples nonmonotonic literature. importantly, sentences system verified performing appropriate experiment semantic domain.",3
"brst double complex coupling gravity gauge theories. consider (effective) quantum general relativity coupled standard model (qgr-sm) study corresponding brst double complex. double complex generated infinitesimal diffeomorphisms infinitesimal gauge transformations. end, define respective brst differentials $p$ $q$ corresponding anti-brst differentials $\overline{p}$ $\overline{q}$. show differentials mutually anticommute, allows us introduce total brst differential $d := p + q$ total anti-brst differential $\overline{d} := \overline{p} + \overline{q}$. furthermore, characterize non-constant lagrange densities essentially closed respect diffeomorphism differentials $p$ $\overline{p}$ scalar tensor densities weight $w = 1$. consequence thereof, obtain graviton-ghosts decouple matter yang--mills theory gauge fixing ghost lagrange densities tensor densities weight $w = 1$. particular, show every lagrange density modified uniquely satisfy said condition. finally, introduce total gauge fixing fermion $\upsilon$ show generate complete gauge fixing ghost lagrange densities qgr-sm via $d \upsilon$.",11
"overview instrumentation dark energy spectroscopic instrument. dark energy spectroscopic instrument (desi) embarked ambitious five-year survey explore nature dark energy spectroscopy 40 million galaxies quasars. desi determine precise redshifts employ baryon acoustic oscillation method measure distances nearby universe z > 3.5, well measure growth structure probe potential modifications general relativity. paper describe significant instrumentation developed desi survey. new instrumentation includes wide-field, 3.2-deg diameter prime-focus corrector focuses light onto 5020 robotic fiber positioners 0.812 diameter, aspheric focal surface. positioners fibers divided among ten wedge-shaped petals. petal connected one ten spectrographs via contiguous, high-efficiency, nearly 50 fiber cable bundle. ten spectrographs use pair dichroics split light three channels together record light 360 - 980 nm resolution 2000 5000. describe science requirements, technical requirements instrumentation, management project. desi installed 4-m mayall telescope kitt peak, also describe facility upgrades prepare desi installation functional verification process. desi achieved performance goals, desi survey began may 2021. performance highlights include rms positioner accuracy better 0.1"", snr per \sqrt{\aa} > 0.5 z > 2 quasar flux 0.28e-17 erg/s/cm^2/a 380 nm 4000s, median snr = 7 [oii] doublet 8e-17 erg/s/cm^2 1000s exposure emission line galaxies z = 1.4 - 1.6. conclude highlights on-sky validation commissioning instrument, key successes, lessons learned. (abridged)",0
"syntactic enhancement vsimm roadmap based anomalous trajectory detection: natural language processing approach. aim syntactic tracking classify spatio-temporal patterns target's motion using natural language processing models. paper, generalize earlier work considering constrained stochastic context free grammar (cscfg) modeling patterns confined roadmap. constrained grammar facilitates modeling specific directions road names roadmap. present novel particle filtering algorithm exploits cscfg model estimating target's patterns. meta-level algorithm operates conjunction base-level tracking algorithm. extensive numerical results using simulated ground moving target indicator (gmti) radar measurements show substantial improvement target tracking accuracy.",3
"predict responsibly: increasing fairness learning defer. many high-stakes ml applications, multiple decision-makers involved, automated human. interaction agents often goes unaddressed algorithmic development. work, explore simple version interaction two-stage framework containing automated model external decision-maker. model choose say idk, pass decision downstream, explored rejection learning. extend concept proposing learning defer, generalizes rejection learning framework considering effect agents decision-making process. propose learning algorithm accounts potential biases held external decision-makers system. experiments real-world datasets demonstrate learning defer make system accurate also less biased. even operated highly biased users, show deferring models still greatly improve fairness entire system.",18
"classification simple harish-chandra modules ovsienko-roger superalgebra. $\omega$-operators virasoro algebra \cite{bf} super virasoro algebra \cite{cl, cll}, get $\omega$-operators ovsienko-roger superalgebras paper use classify simple cuspidal modules $\bz$-graded $\frac12\bz$-graded ovsienko-roger superalgebras. result, easily classify simple harish-chandra modules related lie superalgebras, including $n=1$ bms$_3$ algebra, super $w(2,2)$, etc.",10
"improving coreference resolution learning entity-level distributed representations. long-standing challenge coreference resolution incorporation entity-level information - features defined clusters mentions instead mention pairs. present neural network based coreference system produces high-dimensional vector representations pairs coreference clusters. using representations, system learns combining clusters desirable. train system learning-to-search algorithm teaches local decisions (cluster merges) lead high-scoring final coreference partition. system substantially outperforms current state-of-the-art english chinese portions conll 2012 shared task dataset despite using hand-engineered features.",3
"good friends, bad news - affect virality twitter. link affect, defined capacity sentimental arousal part message, virality, defined probability sent along, significant theoretical practical importance, e.g. viral marketing. quantitative study emailing articles ny times finds strong link positive affect virality, and, based psychological theories concluded relation universally valid. conclusion appears contrast classic theory diffusion news media emphasizing negative affect promoting propagation. paper explore apparent paradox quantitative analysis information diffusion twitter. twitter interesting context shown present characteristics social news media. basic measure virality twitter probability retweet. twitter different email retweeting depend pre-existing social relations, often occur among strangers, thus respect twitter may similar traditional news media. therefore hypothesize negative news content likely retweeted, non-news tweets positive sentiments support virality. test hypothesis analyze three corpora: complete sample tweets cop15 climate summit, random sample tweets, general text corpus including news. latter allows us train classifier distinguish tweets carry news non-news information. present evidence negative sentiment enhances virality news segment, non-news segment. conclude relation affect virality complex expected based findings berger milkman (2010), short 'if want cited: sweet talk friends serve bad news public'.",3
"bayesian optimization machine learning : practical guidebook. engineering machine learning systems still nascent field; relying seemingly daunting collection quickly evolving tools best practices. hope guidebook serve useful resource machine learning practitioners looking take advantage bayesian optimization techniques. outline four example machine learning problems solved using open source machine learning libraries, highlight benefits using bayesian optimization context common machine learning applications.",3
"clustering using max-norm constrained optimization. suggest using max-norm convex surrogate constraint clustering. show yields better exact cluster recovery guarantee previously suggested nuclear-norm relaxation, study effectiveness method, related convex relaxations, compared clustering approaches.",3
"heuristic method solving problem partitioning graphs supply demand. paper present greedy algorithm solving problem maximum partitioning graphs supply demand (mpgsd). goal method solve mpgsd large graphs reasonable time limit. done using two stage greedy algorithm, two corresponding types heuristics. solutions acquired way improved applying computationally inexpensive, hill climbing like, greedy correction procedure. numeric experiments analyze different heuristic functions stage greedy algorithm, show performance highly dependent properties specific instance. tests show exploring relatively small number solutions generated combining different heuristic functions, applying proposed correction procedure find solutions within percent optimal ones.",3
"equitability, interval estimation, statistical power. analysis high-dimensional dataset, common approach test null hypothesis statistical independence variable pairs using non-parametric measure dependence. however, approach attempts identify non-trivial relationship matter weak, often identifies many relationships useful. needed way identifying smaller set relationships merit detailed analysis. formally present characterize equitability, property measures dependence aims overcome challenge. notionally, equitable statistic statistic that, given measure noise, assigns similar scores equally noisy relationships different types [reshef et al. 2011]. begin formalizing idea via new object called interpretable interval, functions interval estimate amount noise relationship unknown type. define equitable statistic one small interpretable intervals. draw equivalence interval estimation hypothesis testing show moderate assumptions equitable statistic one yields well powered tests distinguishing trivial non-trivial relationships kinds also non-trivial relationships different strengths. means equitability allows us specify threshold relationship strength $x_0$ search relationships kinds strength greater $x_0$. thus, equitability thought strengthening power independence enables fruitful analysis data sets small number strong, interesting relationships large number weaker ones. conclude demonstration two equivalent characterizations equitability used evaluate equitability statistic practice.",10
"accelerated gradient temporal difference learning. family temporal difference (td) methods span spectrum computationally frugal linear methods like td({\lambda}) data efficient least squares methods. least square methods make best use available data directly computing td solution thus require tuning typically highly sensitive learning rate parameter, require quadratic computation storage. recent algorithmic developments yielded several sub-quadratic methods use approximation least squares td solution, incur bias. paper, propose new family accelerated gradient td (atd) methods (1) provide similar data efficiency benefits least-squares methods, fraction computation storage (2) significantly reduce parameter sensitivity compared linear td methods, (3) asymptotically unbiased. illustrate claims proof convergence expectation experiments several benchmark domains large-scale industrial energy allocation domain.",3
"filter-bank-enabled leaky-wave antenna array technique full-band-locked radar system stitched frequency-space domain. inspired filter-bank (fb) concept normally used multi-rate signal processing, fb-enabled array technique leaky-wave antennas (lwas) proposed studied creating full-band-locked frequency-scanning radar (fsr) systems stitched frequency-space domain. mainly addressing coupling dilemma range angle resolutions, historically naturally inherited conventional fsr. first all, frequency-modulated continuous-wave system architecture selected exemplify recall characteristics conventional fsr emphasis resolution coupling. then, radar solution featuring stitched frequency-space domain introduced resolution decoupling, depending array fb-enabled lwa channels. radar equation, fb-related conditions realizing critical frequency-space stitching analytically derived equivalently converted several design specifications array lwas, i.e., engineered beam-scanning functions, beam-crossovers, phase alignments. facilitate practical implementation array technique, detailed generalized design flow developed. finally, simple proof concept, fb-enabled two-channel lwa array modeled, fabricated, measured. simulated measured results reasonable agreement, demonstrate desired frequency-space stitching behavior, i.e., enhanced spectrum bandwidth widened radiation beamwidth. proposed array solution may potentially deployed fsr systems full-band-locked beam illumination decoupled range-angle resolution, may present competitor phased array technique.",14
"near-optimal discrete optimization experimental design: regret minimization approach. experimental design problem concerns selection k points potentially large design pool p-dimensional vectors, maximize statistical efficiency regressed selected k design points. statistical efficiency measured optimality criteria, including a(verage), d(eterminant), t(race), e(igen), v(ariance) g-optimality. except t-optimality, exact optimization np-hard. propose polynomial-time regret minimization framework achieve $(1+\varepsilon)$ approximation $o(p/\varepsilon^2)$ design points, optimality criteria above. contrast, best knowledge, work, polynomial-time algorithm achieves $(1+\varepsilon)$ approximations d/e/g-optimality, best poly-time algorithm achieving $(1+\varepsilon)$-approximation a/v-optimality requires $k = \omega(p^2/\varepsilon)$ design points.",18
"partial membership latent dirichlet allocation. topic models (e.g., plsa, lda, slda) widely used segmenting imagery. models confined crisp segmentation. yet, many images regions cannot assigned crisp label (e.g., transition regions foggy sky ground sand water beach). cases, visual word best represented partial memberships across multiple topics. address this, present partial membership latent dirichlet allocation (pm-lda) model associated parameter estimation algorithms. experimental results two natural image datasets one sonar image dataset show pm-lda produce crisp soft semantic image segmentations; capability existing methods have.",18
"face expression recognition analysis: state art. automatic recognition facial expressions active research topic since early nineties. several advances past years terms face detection tracking, feature extraction mechanisms techniques used expression classification. paper surveys published work since 2001 till date. paper presents time-line view advances made field, applications automatic face expression recognizers, characteristics ideal system, databases used advances made terms standardization detailed summary state art. paper also discusses facial parameterization using facs action units (aus) mpeg-4 facial animation parameters (faps) recent advances face detection, tracking feature extraction methods. notes also presented emotions, expressions facial features, discussion six prototypic expressions recent studies expression classifiers. paper ends note challenges future work. paper written tutorial style intention helping students researchers new field.",3
"structured matrix recovery via generalized dantzig selector. recent years, structured matrix recovery problems gained considerable attention real world applications, recommender systems computer vision. much existing work focused matrices low-rank structure, limited progress made matrices types structure. paper present non-asymptotic analysis estimation generally structured matrices via generalized dantzig selector generic sub-gaussian measurements. show estimation error always succinctly expressed terms geometric measures suitable sets depend structure underlying true matrix. addition, derive general bounds geometric measures structures characterized unitarily invariant norms, large family covering matrix norms practical interest. examples provided illustrate utility theoretical development.",18
"monotonous intuitionistic fuzzy topsis method linear orders two novel admissible distance measures. intuitionistic fuzzy topsis methods contain two key elements: (1) order structure, affect choices positive ideal-points negative ideal-points, construction admissible distance/similarity measures; (2) distance/similarity measure, closely related values relative closeness degrees determines accuracy rationality decision-making. order structure, many efforts devoted constructing score functions, strictly distinguish different intuitionistic fuzzy values (ifvs) preserve natural partial order ifvs. paper proves score function exist, namely application single monotonous continuous function distinguish ifvs. distance similarity measure, examples given show classical similarity measures based normalized euclidean distance normalized minkowski distance meet axiomatic definition intuitionistic fuzzy similarity measures. moreover,some illustrative examples given show classical intuitionistic fuzzy topsis methods ensure monotonicity natural partial order linear orders, may yield counter-intuitive results. overcome limitation, propose novel intuitionistic fuzzy topsis method, using two new admissible distances linear orders measured score degree/similarity function accuracy degree, prove monotonous two linear orders. first result strict mathematical proof monotonicity linear orders intuitionistic fuzzy topsis method. finally, show two practical examples comparative analysis decision-making methods illustrate efficiency developed topsis method.",10
"embedding dimension laplacian eigenfunction maps. closed, connected riemannian manifold $m$ smoothly embedded laplacian eigenfunction maps $\mathbb{r}^m$ $m$. call smallest $m$ maximal embedding dimension $m$. show maximal embedding dimension $m$ bounded constant depending dimension $m$, lower bound injectivity radius, lower bound ricci curvature, volume bound. interpret result case surfaces isometrically immersed $\mathbb{r}^3$, showing maximal embedding dimension depends bounds gaussian curvature, mean curvature, surface area. furthermore, consider relevance results shape registration.",18
"neutralized empirical risk minimization generalization neutrality bound. currently, machine learning plays important role lives individual activities numerous people. accordingly, become necessary design machine learning algorithms ensure discrimination, biased views, unfair treatment result decision making predictions made via machine learning. work, introduce novel empirical risk minimization (erm) framework supervised learning, neutralized erm (nerm) ensures classifiers obtained guaranteed neutral respect viewpoint hypothesis. specifically, given viewpoint hypothesis, nerm works find target hypothesis minimizes empirical risk simultaneously identifying target hypothesis neutral viewpoint hypothesis. within nerm framework, derive theoretical bound empirical generalization neutrality risks. furthermore, realization nerm linear classification, derive max-margin algorithm, neutral support vector machine (svm). experimental results show neutral svm shows improved classification performance real datasets without sacrificing neutrality guarantee.",18
"inferring missing categorical information noisy sparse web markup. embedded markup web pages seen widespread adoption throughout past years driven standards rdfa microdata initiatives schema.org, recent studies show adoption 39% web pages already 2016. constitutes important information source tasks web search, web page classification knowledge graph augmentation, individual markup nodes usually sparsely described often lack essential information. instance, 26 million nodes describing events within common crawl 2016, 59% nodes provide less six statements 257,000 nodes (0.96%) typed specific event subtypes. nevertheless, given scale diversity web markup data, nodes provide missing information obtained web large quantities, particular categorical properties. data constitutes potential training data inferring missing information significantly augment sparsely described nodes. work, introduce supervised approach inferring missing categorical properties web markup. experiments, conducted properties events movies, show performance 79% 83% f1 score correspondingly, significantly outperforming existing baselines.",3
"signature lattice dynamics twisted 2d homo/hetero-bilayers. twisted 2d bilayer materials created artificial stacking two monolayer crystal networks 2d materials desired twisting angle $\theta$. material forms moir\'e superlattice due periodicity top bottom layer crystal structure. optical properties modified lattice reconstruction phonon renormalization, makes optical spectroscopy ideal characterization tool study novel physics phenomena. here, report raman investigation full period twisted bilayer (tb) wse$_2$ moir\'e superlattice (\textit i.e. 0{\deg} $\leq \theta \leq$ 60{\deg}). observe intensity ratio two raman peaks, $b_{2g}$ $e_{2g}/a_{1g}$ correlates evolution moir\'e period. raman intensity ratio function twisting angle follows exponential profile matching moir\'e period two local maxima 0{\deg} 60{\deg} minimum 30{\deg}. using series temperature-dependent raman photoluminescence (pl) measurements well \textit{ab initio} calculations, intensity ratio $i_{b_{2g}}/i_{{e_{2g}}/{a_{1g}}}$ explained signature lattice dynamics tb wse$_2$ moir\'e superlattices. exploring different material combinations twisted hetero-bilayers, results extended kinds mo- w-based tmdcs.",2
"consistent deterministic regression tree non-parametric prediction time series. study online prediction bounded stationary ergodic processes. so, consider setting prediction individual sequences build deterministic regression tree performs asymptotically well best l-lipschitz constant predictors. then, show obtained regret bound entails asymptotical optimality respect class bounded stationary ergodic processes.",10
"relations escape regions parameter space cubic polynomials. describe topological relationship slices parameter space cubic maps. paper \cite{cp1}, milnor defined curves $\mathcal{s}_n$ set cubic polynomials marked critical point period $p$. paper, describe relationship boundaries connectedness loci curves $\mathcal{s}_1$ $\mathcal{s}_2$.",10
"batch normalized recurrent neural networks. recurrent neural networks (rnns) powerful models sequential data potential learn long-term dependencies. however, computationally expensive train difficult parallelize. recent work shown normalizing intermediate representations neural networks significantly improve convergence rates feedforward neural networks . particular, batch normalization, uses mini-batch statistics standardize features, shown significantly reduce training time. paper, show applying batch normalization hidden-to-hidden transitions rnns help training procedure. also show applied input-to-hidden transitions, batch normalization lead faster convergence training criterion seem improve generalization performance language modelling speech recognition tasks. all, applying batch normalization rnns turns challenging applying feedforward networks, certain variants still beneficial.",18
"exploiting parallelism hard problems abstract argumentation: technical report. abstract argumentation framework (\afname) unifying framework able encompass variety nonmonotonic reasoning approaches, logic programming computational argumentation. yet, efficient approaches decision enumeration problems associated \afname missing, thus potentially limiting efficacy argumentation-based approaches real domains. paper, present algorithm enumerating preferred extensions abstract argumentation frameworks exploits parallel computation. purpose, scc-recursive semantics definition schema adopted, extensions defined level specific sub-frameworks. algorithm shows significant performance improvements large frameworks, terms number solutions found speedup.",3
"dissipation-free modes dissipative systems. coupling system environment (or bath) always leads dissipation. show, however, system composed two subsystems dissipation-free mode, bath shared two subsystems. reading reverse, shared bath contribute dissipation modes. key example, consider simple model two-sublattice antiferromagnet, environment modeled bath shared two sublattice magnetizations. model, find n\'eel order parameter dissipation-free mode. antiferromagnets, results offer explanation dissipation rate n\'eel vector typically much lower average magnetization. general, results suggest way reduce dissipation (and decoherence) modes composite systems, could experimental technological applications.",2
"deep canonically correlated lstms. examine deep canonically correlated lstms way learn nonlinear transformations variable length sequences embed correlated, fixed dimensional space. use lstms transform multi-view time-series data non-linearly learning temporal relationships within data. perform correlation analysis outputs neural networks find correlated subspace get final representation via projection. work follows previous work done deep canonical correlation (dcca), deep feed-forward neural networks used learn nonlinear transformations data maximizing correlation.",18
"subspace clustering high-dimensional data: predictive approach. several application domains, high-dimensional observations collected analysed search naturally occurring data clusters might provide insights nature problem. paper describe new approach partitioning high-dimensional data. assumption that, within cluster, data approximated well linear subspace estimated means principal component analysis (pca). proposed algorithm, predictive subspace clustering (psc) partitions data clusters simultaneously estimating cluster-wise pca parameters. algorithm minimises objective function depends upon new measure influence pca models. penalised version algorithm also described carrying simultaneous subspace clustering variable selection. convergence psc discussed detail, extensive simulation results comparisons competing methods presented. comparative performance psc assessed six real gene expression data sets psc often provides state-of-art results.",18
"effective building block design deep convolutional neural networks using search. deep learning shown promising results many machine learning tasks dl models often complex networks large number neurons layers, recently, complex layer structures known building blocks. finding best deep model requires combination finding right architecture correct set parameters appropriate architecture. addition, complexity (in terms layer types, number neurons, number layers) also present problems generalization since larger networks easier overfit data. paper, propose search framework finding effective architectural building blocks convolutional neural networks (cnn). approach much faster finding models close state-of-the-art performance. addition, models discovered approach also smaller models discovered similar techniques. achieve twin advantages designing search space way searches reduced set state-of-the-art building blocks cnns including residual block, inception block, inception-residual block, resnext block many others. apply technique generate models multiple image datasets show models achieve performance comparable state-of-the-art (and even surpassing state-of-the-art one case). also show learned models transferable datasets.",3
"information dropout: learning optimal representations noisy computation. cross-entropy loss commonly used deep learning closely related defining properties optimal representations, enforce key properties. show solved adding regularization term, turn related injecting multiplicative noise activations deep neural network, special case common practice dropout. show regularized loss function efficiently minimized using information dropout, generalization dropout rooted information theoretic principles automatically adapts data better exploit architectures limited capacity. task reconstruction input, show loss function yields variational autoencoder special case, thus providing link representation learning, information theory variational inference. finally, prove promote creation disentangled representations simply enforcing factorized prior, fact observed empirically recent work. experiments validate theoretical intuitions behind method, find information dropout achieves comparable better generalization performance binary dropout, especially smaller models, since automatically adapt noise structure network, well test sample.",18
"parameter selection particle swarm optimization transportation network design problem. transportation planning development, transport network design problem seeks optimize specific objectives (e.g. total travel time) choosing among given set projects keeping consumption resources (e.g. budget) within limits. due numerous cases choosing projects, solving problem difficult time-consuming. based particle swarm optimization (pso) technique, heuristic solution algorithm bi-level problem designed. paper evaluates algorithm performance response changing certain basic pso parameters.",10
"simulation calibration fully bayesian marked multidimensional hawkes process dissimilar decays. propose simulation method multidimensional hawkes processes based superposition theory point processes. formulation allows us design efficient simulations hawkes processes differing exponentially decaying intensities. demonstrate inter-arrival times decomposed simpler auxiliary variables sampled directly, giving exact simulation approximation. establish auxiliary variables provides information parent process event time. algorithm correctness shown verifying simulated intensities theoretical moments. modular inference procedure consisting gibbs samplers auxiliary variable augmentation adaptive rejection sampling presented. finally, compare proposed simulation method existing methods, find significant improvement terms algorithm speed. inference algorithm used discover strengths mutually excitations real dark networks.",18
"learning mid-level words riemannian manifold action recognition. human action recognition remains challenging task due various sources video data large intra-class variations. thus becomes one key issues recent research explore effective robust representation handle challenges. paper, propose novel representation approach constructing mid-level words videos encoding riemannian manifold. specifically, first conduct global alignment densely extracted low-level features build bank corresponding feature groups, statistically modeled mid-level word lying specific riemannian manifold. based mid-level words, construct intrinsic riemannian codebooks employing k-karcher-means clustering riemannian gaussian mixture model, consequently extend riemannian manifold version three well studied encoding methods euclidean space, i.e. bag visual words (bovw), vector locally aggregated descriptors (vlad), fisher vector (fv), obtain final action video representations. method evaluated two tasks four popular realistic datasets: action recognition youtube, ucf50, hmdb51 databases, action similarity labeling aslan database. cases, reported results achieve competitive performance recent state-of-the-art works.",3
"networks utilization improvements service discovery performance. service discovery requests' messages vital role sharing locating resources many service discovery protocols. sending messages link handle may cause congestion loss messages dramatically influences performance protocols. re-send lost messages result latency inefficiency performing tasks user(s) require connected nodes. issue become serious problem two cases: first, number clients performs service discovery request increasing, result increasing number sent discovery messages; second, network resources bandwidth capacity consumed applications. two cases lead network congestion loss messages. paper propose algorithm improve services discovery protocols performance separating consecutive burst messages specific period time calculated regarding available network resources. tested routers connected two configurations; decentralised centralised .in addition, paper explains impact increasing number clients consumed network resources proposed algorithm.",3
"machine learning galaxy morphology classification. work, decision tree learning algorithms fuzzy inferencing systems applied galaxy morphology classification. particular, cart, c4.5, random forest fuzzy logic algorithms studied reliable classifiers developed distinguish spiral galaxies, elliptical galaxies star/unknown galactic objects. morphology information training testing datasets obtained galaxy zoo project corresponding photometric spectra parameters downloaded sdss dr7 catalogue.",0
"elf: extensive, lightweight flexible research platform real-time strategy games. paper, propose elf, extensive, lightweight flexible platform fundamental reinforcement learning research. using elf, implement highly customizable real-time strategy (rts) engine three game environments (mini-rts, capture flag tower defense). mini-rts, miniature version starcraft, captures key game dynamics runs 40k frame-per-second (fps) per core macbook pro notebook. coupled modern reinforcement learning methods, system train full-game bot built-in ais end-to-end one day 6 cpus 1 gpu. addition, platform flexible terms environment-agent communication topologies, choices rl methods, changes game parameters, host existing c/c++-based game environments like arcade learning environment. using elf, thoroughly explore training parameters show network leaky relu batch normalization coupled long-horizon training progressive curriculum beats rule-based built-in ai $70\%$ time full game mini-rts. strong performance also achieved two games. game replays, show agents learn interesting strategies. elf, along rl platform, open-sourced https://github.com/facebookresearch/elf.",3
"identifying hierarchical structure sequences: linear-time algorithm. sequitur algorithm infers hierarchical structure sequence discrete symbols replacing repeated phrases grammatical rule generates phrase, continuing process recursively. result hierarchical representation original sequence, offers insights lexical structure. algorithm driven two constraints reduce size grammar, produce structure by-product. sequitur breaks new ground operating incrementally. moreover, method's simple structure permits proof operates space time linear size input. implementation process 50,000 symbols per second applied extensive range real world sequences.",3
"improving factor-based quantitative investing forecasting company fundamentals. periodic basis, publicly traded companies required report fundamentals: financial data revenue, operating income, debt, among others. data points provide insight financial health company. academic research identified factors, i.e. computed features reported data, known retrospective analysis outperform market average. two popular factors book value normalized market capitalization (book-to-market) operating income normalized enterprise value (ebit/ev). paper: first show simulation could (clairvoyantly) select stocks using factors calculated future fundamentals (via oracle), portfolios would far outperform standard factor approach. motivated analysis, train deep neural networks forecast future fundamentals based trailing 5-years window. quantitative analysis demonstrates significant improvement mse naive strategy. moreover, retrospective analysis using industry-grade stock portfolio simulator (backtester), show improvement compounded annual return 17.1% (mlp) vs 14.4% standard factor model.",18
"building refining abstract planning cases change representation language. ion one promising approaches improve performance problem solvers. several domains abstraction dropping sentences domain description -- used hierarchical planners -- proven useful. paper present examples illustrate significant drawbacks abstraction dropping sentences. overcome drawbacks, propose general view abstraction involving change representation language. developed new abstraction methodology related sound complete learning algorithm allows complete change representation language planning cases concrete abstract. however, achieve powerful change representation language, abstract language well rules describe admissible ways abstracting states must provided domain model. new abstraction approach core paris (plan abstraction refinement integrated system), system abstract planning cases automatically learned given concrete cases. empirical study domain process planning mechanical engineering shows significant advantages proposed reasoning abstract cases classical hierarchical planning.",3
"multiclass diffuse interface models semi-supervised learning graphs. present graph-based variational algorithm multiclass classification high-dimensional data, motivated total variation techniques. energy functional based diffuse interface model periodic potential. augment model introducing alternative measure smoothness preserves symmetry among class labels. modification standard laplacian, construct efficient multiclass method allows sharp transitions classes. experimental results demonstrate approach competitive state art among graph-based algorithms.",18
"deep learning undersampled mri reconstruction. paper presents deep learning method faster magnetic resonance imaging (mri) reducing k-space data sub-nyquist sampling strategies provides rationale proposed approach works well. uniform subsampling used time-consuming phase-encoding direction capture high-resolution image information, permitting image-folding problem dictated poisson summation formula. deal localization uncertainty due image folding, low-frequency k-space data added. training deep learning net involves input output images pairs fourier transforms subsampled fully sampled k-space data. numerous experiments show remarkable performance proposed method; 29% k-space data generate images high quality effectively standard mri reconstruction fully sampled data.",18
"disc reflection model ultra-soft narrow-line seyfert 1 galaxies. present detailed analysis xmm-newton observations five narrow-line seyfert 1 galaxies (nls1s). show soft continuum emission x-ray band photon index $\gamma\gtrsim 2.5$. therefore, referred `ultra-soft' nls1s paper. modeling optical/uv-x-ray spectral energy distribution (sed) reflection-based model, find indications disc surface ultra-soft nls1s higher ionisation state typical seyfert 1 agn. best-fit sed models suggest five ultra-soft nls1s eddington ratio $\lambda_{\rm edd}=1-20$ assuming available black hole mass measurements. addition, models infer significant fraction disc energy ultra-soft nls1s radiated away form non-thermal emission instead thermal emission disc. due extreme properties, x-ray observations sources iron band particularly challenging. future observations, e.g. athena, enable us clearer view spectral shape iron band thus distinguish reflection model interpretations broadband spectra.",0
"theory sexes geodakian advanced iskrin. 1960s v.geodakian proposed theory explains sexes mechanism evolutionary adaptation species changing environmental conditions. 2001 v.iskrin refined augmented concepts geodakian gave new interesting explanation several phenomena involve sex, sex ratio, including war-years phenomena. also introduced new concept ""catastrophic sex ratio."" note attempt digest technical aspects new ideas iskrin.",3
"comparing two trainable grammatical relations finders. grammatical relationships (grs) form important level natural language processing, different sets grs useful different purposes. therefore, one may often time obtain small training corpus desired gr annotations. small training corpus, compare two systems. use different learning techniques, find difference minor effect. larger factor english, different gr length measure appears better suited finding simple argument grs finding modifier grs. also find partitioning data may help memory-based learning.",3
"fuzzy clustering distribution-valued data using adaptive l2 wasserstein distances. distributional (or distribution-valued) data new type data arising several sources considered realizations distributional variables. new set fuzzy c-means algorithms data described distributional variables proposed. algorithms use $l2$ wasserstein distance distributions dissimilarity measures. beside extension fuzzy c-means algorithm distributional data, considering decomposition squared $l2$ wasserstein distance, propose set algorithms using different automatic way compute weights associated variables well components, globally cluster-wise. relevance weights computed clustering process introducing product-to-one constraints. relevance weights induce adaptive distances expressing importance variable component clustering process, acting also variable selection method clustering. tested proposed algorithms artificial real-world data. results confirm proposed methods able better take account cluster structure data respect standard fuzzy c-means, non-adaptive distances.",18
"active learning online recognition human activities streaming videos. recognising human activities streaming videos poses unique challenges learning algorithms: predictive models need scalable, incrementally trainable, must remain bounded size even data stream arbitrarily long. furthermore, parameter tuning problematic streaming setting, suitable approaches parameterless, make assumptions class labels may occur stream. present approach recognition human actions streaming data meets requirements by: (1) incrementally learning model adaptively covers feature space simple local classifiers; (2) employing active learning strategy reduce annotation requests; (3) achieving promising accuracy within fixed model size. extensive experiments standard benchmarks show approach competitive state-of-the-art non-incremental methods, outperforms existing active incremental baselines.",18
"deep learning universal linear embeddings nonlinear dynamics. identifying coordinate transformations make strongly nonlinear dynamics approximately linear central challenge modern dynamical systems. transformations potential enable prediction, estimation, control nonlinear systems using standard linear theory. koopman operator emerged leading data-driven embedding, eigenfunctions operator provide intrinsic coordinates globally linearize dynamics. however, identifying representing eigenfunctions proven mathematically computationally challenging. work leverages power deep learning discover representations koopman eigenfunctions trajectory data dynamical systems. network parsimonious interpretable construction, embedding dynamics low-dimensional manifold intrinsic rank dynamics parameterized koopman eigenfunctions. particular, identify nonlinear coordinates dynamics globally linear using modified auto-encoder. also generalize koopman representations include ubiquitous class systems exhibit continuous spectra, ranging simple pendulum nonlinear optics broadband turbulence. framework parametrizes continuous frequency using auxiliary network, enabling compact efficient embedding intrinsic rank, connecting models half century asymptotics. way, benefit power generality deep learning, retaining physical interpretability koopman embeddings.",10
"win prediction esports: mixed-rank match prediction multi-player online battle arena games. esports emerged popular genre players well spectators, supporting global entertainment industry. esports analytics evolved address requirement data-driven feedback, focused cyber-athlete evaluation, strategy prediction. towards latter, previous work used match data variety player ranks hobbyist professional players. however, professional players shown behave differently lower ranked players. given comparatively limited supply professional data, key question thus whether mixed-rank match datasets used create data-driven models predict winners professional matches provide simple in-game statistic viewers broadcasters. show that, although slightly reduced accuracy, mixed-rank datasets used predict outcome professional matches, suitably optimized configurations.",3
"median-truncated nonconvex approach phase retrieval outliers. paper investigates phase retrieval problem, aims recover signal magnitudes linear measurements. develop statistically computationally efficient algorithms situation measurements corrupted sparse outliers take arbitrary values. propose novel approach robustify gradient descent algorithm using sample median guide pruning spurious samples initialization local search. adopting poisson loss reshaped quadratic loss respectively, obtain two algorithms termed median-twf median-rwf, provably recover signal near-optimal number measurements measurement vectors composed i.i.d. gaussian entries, logarithmic factor, even constant fraction measurements adversarially corrupted. show algorithms stable presence additional dense bounded noise. analysis accomplished developing non-trivial concentration results median-related quantities, may independent interest. provide numerical experiments demonstrate effectiveness approach.",18
"stochastic subsampling factorizing huge matrices. present matrix-factorization algorithm scales input matrices huge number rows columns. learned factors may sparse dense and/or non-negative, makes algorithm suitable dictionary learning, sparse component analysis, non-negative matrix factorization. algorithm streams matrix columns subsampling iteratively learn matrix factors. iteration, row dimension new sample reduced subsampling, resulting lower time complexity compared simple streaming algorithm. method comes convergence guarantees reach stationary point matrix-factorization problem. demonstrate efficiency massive functional magnetic resonance imaging data (2 tb), patches extracted hyperspectral images (103 gb). problems, involve different penalties rows columns, obtain significant speed-ups compared state-of-the-art algorithms.",18
"sparse projections medical images onto manifolds. manifold learning successfully applied variety medical imaging problems. use real-time applications requires fast projection onto low-dimensional space. end, out-of-sample extensions applied constructing interpolation function maps input space low-dimensional manifold. commonly used approaches nystr\""{o}m extension kernel ridge regression require using training points. propose interpolation function depends small subset input training data. consequently, testing phase new point needs compared small number input training data order project point onto low-dimensional space. interpret method out-of-sample extension approximates kernel ridge regression. method involves solving simple convex optimization problem attractive property guaranteeing upper bound approximation error, crucial medical applications. tuning error bound controls sparsity resulting interpolation function. illustrate method two clinical applications require fast mapping input images onto low-dimensional space.",3
"techniques learning binary stochastic feedforward neural networks. stochastic binary hidden units multi-layer perceptron (mlp) network give least three potential benefits compared deterministic mlp networks. (1) allow learn one-to-many type mappings. (2) used structured prediction problems, modeling internal structure output important. (3) stochasticity shown excellent regularizer, makes generalization performance potentially better general. however, training stochastic networks considerably difficult. study training using samples hidden activations per input. show case m=1 leads fundamentally different behavior network tries avoid stochasticity. propose two new estimators training gradient propose benchmark tests comparing training algorithms. experiments confirm training stochastic networks difficult show proposed two estimators perform favorably among five known estimators.",18
"measuring dependence powerfully equitably. given high-dimensional data set often wish find strongest relationships within it. common strategy evaluate measure dependence every variable pair retain highest-scoring pairs follow-up. strategy works well statistic used equitable [reshef et al. 2015a], i.e., if, measure noise, assigns similar scores equally noisy relationships regardless relationship type (e.g., linear, exponential, periodic). paper, introduce characterize population measure dependence called mic*. show three ways mic* viewed: population value mic, highly equitable statistic [reshef et al. 2011], canonical ""smoothing"" mutual information, supremum infinite sequence defined terms optimal one-dimensional partitions marginals joint distribution. based theory, introduce efficient approach computing mic* density pair random variables, define new consistent estimator mice mic* efficiently computable. contrast, known polynomial-time algorithm computing original equitable statistic mic. show simulations mice better bias-variance properties mic. introduce prove consistency second statistic, tice, trivial side-product computation mice whose goal powerful independence testing rather equitability. show simulations mice tice good equitability power independence respectively. analyses complement in-depth empirical evaluation several leading measures dependence [reshef et al. 2015b] shows state-of-the-art performance mice tice.",18
"less more: culling training set improve robustness deep neural networks. deep neural networks vulnerable adversarial examples. prior defenses attempted make deep networks robust either improving network architecture adding adversarial examples training set, respective limitations. propose new direction. motivated recent research shows outliers training set high negative influence trained model, approach makes model robust detecting removing outliers training set without modifying network architecture requiring adversarial examples. propose two methods detecting outliers based canonical examples training errors, respectively. removing outliers, train classifier remaining examples obtain sanitized model. evaluation shows sanitized model improves classification accuracy forces attacks generate adversarial examples higher distortions. moreover, kullback-leibler divergence output original model sanitized model allows us distinguish normal adversarial examples reliably.",3
"kernel methods covariates rankings. permutation-valued features arise variety applications, either direct way preferences elicited collection items, indirect way numerical ratings converted ranking. date, relatively limited study regression, classification, testing problems based permutation-valued features, opposed permutation-valued responses. paper studies use reproducing kernel hilbert space methods learning permutation-valued features. methods embed rankings implicitly defined function space, allow efficient estimation regression test functions richer space. first contribution characterize feature spaces spectral properties associated two kernels rankings, kendall mallows kernels. using tools representation theory, explain limited expressive power kendall kernel characterizing degenerate spectrum, sharp contrast, prove mallows' kernel universal characteristic. also introduce families polynomial kernels interpolate kendall (degree one) mallows' (infinite degree) kernels. show practical effectiveness methods via applications eurobarometer survey data well movielens ratings dataset.",18
"rational distributed process-level account independence judgment. inconceivable chaotic world would look humans, faced innumerable decisions day made uncertainty, lacking capacity distinguish relevant irrelevant---a capacity computationally amounts handling probabilistic independence relations. highly parallel distributed computational machinery brain suggests satisfying process-level account human independence judgment also mimic features. work, present first rational, distributed, message-passing, process-level account independence judgment, called $\mathcal{d}^\ast$. interestingly, $\mathcal{d}^\ast$ shows curious, normatively-justified tendency quick detection dependencies, whenever hold. furthermore, $\mathcal{d}^\ast$ outperforms previously proposed algorithms ai literature terms worst-case running time, salient aspect supported recent work neuroscience investigating possible implementations bayes nets neural level. $\mathcal{d}^\ast$ nicely exemplifies pursuit cognitive plausibility lead discovery state-of-the-art algorithms appealing properties, simplicity makes $\mathcal{d}^\ast$ potentially good candidate pedagogical purposes.",3
"digital ecosystems: optimisation distributed intelligence. intelligence optimise digital ecosystems? could distributed intelligence interact ecosystem dynamics? software components part genetic selection intelligent themselves, adaptive technology? consider effect distributed intelligence mechanism evolutionary ecological dynamics digital ecosystem, digital counterpart biological ecosystem evolving software services distributed network. investigate neural networks support vector machine learning based pattern recognition functionality distributed intelligence. simulation results imply digital ecosystem performs better application distributed intelligence, marginally effectively powered support vector machine neural networks, suggest contribute optimising operation digital ecosystem.",3
"probabilistic numerics uncertainty computations. deliver call arms probabilistic numerical methods: algorithms numerical tasks, including linear algebra, integration, optimization solving differential equations, return uncertainties calculations. uncertainties, arising loss precision induced numerical calculation limited time hardware, important much contemporary science industry. within applications climate science astrophysics, need make decisions basis computations large complex data led renewed focus management numerical uncertainty. describe several seminal classic numerical methods interpreted naturally probabilistic inference. show probabilistic view suggests new algorithms flexibly adapted suit application specifics, delivering improved empirical performance. provide concrete illustrations benefits probabilistic numeric algorithms real scientific problems astrometry astronomical imaging, highlighting open problems new algorithms. finally, describe probabilistic numerical methods provide coherent framework identifying uncertainty calculations performed combination numerical algorithms (e.g. numerical optimisers differential equation solvers), potentially allowing diagnosis (and control) error sources computations.",10
"stochastic bound majorization. recently majorization method optimizing partition functions log-linear models proposed alongside novel quadratic variational upper-bound. batch setting, outperformed state-of-the-art first- second-order optimization methods various learning tasks. propose stochastic version bound majorization method well low-rank modification high-dimensional data-sets. resulting stochastic second-order method outperforms stochastic gradient descent (across variations various tunings) terms number iterations computation time till convergence finding better quality parameter setting. proposed method bridges first- second-order stochastic optimization methods maintaining computational complexity linear data dimension exploiting second order information pseudo-global curvature objective function (as opposed local curvature hessian).",3
"incorporating knowledge structural equation models using auxiliary variables. paper, extend graph-based identification methods allowing background knowledge form non-zero parameter values. information could obtained, example, previously conducted randomized experiment, substantive understanding domain, even identification technique. incorporate information systematically, propose addition auxiliary variables model, constructed certain paths conveniently cancelled. cancellation allows auxiliary variables help conventional methods identification (e.g., single-door criterion, instrumental variables, half-trek criterion), well model testing (e.g., d-separation, over-identification). moreover, iteratively alternating steps identification adding auxiliary variables, improve power existing identification methods via bootstrapping approach require external knowledge. operationalize method simple instrumental sets (a generalization instrumental variables) show resulting method able identify least many models general identification method linear systems known date. discuss application auxiliary variables tasks model testing z-identification.",18
"relational learning feature extraction querying heterogeneous information networks. many real world systems need operate heterogeneous information networks consist numerous interacting components different types. examples include systems perform data analysis biological information networks; social networks; information extraction systems processing unstructured data convert raw text knowledge graphs. many previous works describe specialized approaches perform specific types analysis, mining learning networks. work, propose unified framework consisting data model -a graph first order schema along declarative language constructing, querying manipulating networks ways facilitate relational structured machine learning. particular, provide initial prototype relational graph traversal query language queries directly used relational features structured machine learning models. feature extraction performed making declarative graph traversal queries. learning inference models directly operate relational representation augment new data knowledge that, turn, integrated seamlessly relational structure support new predictions. demonstrate system's capabilities showcasing tasks natural language processing computational biology domains.",3
"temporal topic modeling assess associations news trends infectious disease outbreaks. retrospective assessments, internet news reports shown capture early reports unknown infectious disease transmission prior official laboratory confirmation. general, media interest reporting peaks wanes course outbreak. study, quantify extent media interest infectious disease outbreaks indicative trends reported incidence. introduce approach uses supervised temporal topic models transform large corpora news articles temporal topic trends. key advantages approach include, applicability wide range diseases, ability capture disease dynamics - including seasonality, abrupt peaks troughs. evaluated method using data multiple infectious disease outbreaks reported united states america (u.s.), china india. noted temporal topic trends extracted disease-related news reports successfully captured dynamics multiple outbreaks whooping cough u.s. (2012), dengue outbreaks india (2013) china (2014). observations also suggest efficient modeling temporal topic trends using time-series regression techniques estimate disease case counts increased precision official reports health organizations.",3
"learning boltzmann machine em-like method. propose expectation-maximization-like(emlike) method train boltzmann machine unconstrained connectivity. adopts monte carlo approximation e-step, replaces intractable likelihood objective efficiently computed objectives directly approximates gradient likelihood objective m-step. em-like method modification alternating minimization. prove em-like method exactly contrastive divergence restricted boltzmann machine m-step method adopts special approximation. also propose new measure assess performance boltzmann machine generative models data, computational complexity o(rmn). finally, demonstrate performance em-like method using numerical experiments.",3
"finding academic experts multisensor approach using shannon's entropy. expert finding information retrieval task concerned search knowledgeable people, topic, basis documents describing peoples activities. task involves taking user query input returning list people sorted level expertise regarding user query. paper introduces novel approach combining multiple estimators expertise based multisensor data fusion framework together dempster-shafer theory evidence shannon's entropy. specifically, defined three sensors detect heterogeneous information derived textual contents, graph structure citation patterns community experts, profile information academic experts. given evidences collected, sensor may define different candidates experts consequently agree final ranking decision. deal conflicts, applied dempster-shafer theory evidence combined shannon's entropy formula fuse information come accurate reliable final ranking list. experiments made two datasets academic publications computer science domain attest adequacy proposed approach traditional state art approaches. also made experiments representative supervised state art algorithms. results revealed proposed method achieved similar performance compared supervised techniques, confirming capabilities proposed framework.",3
"imposing hard constraints deep networks: promises limitations. imposing constraints output deep neural net one way improve quality predictions loosening requirements labeled training data. constraints usually imposed soft constraints adding new terms loss function minimized training. alternative impose hard constraints, number theoretical benefits explored far due perceived intractability problem. paper, show imposing hard constraints fact done computationally feasible way delivers reasonable results. however, theoretical benefits materialize resulting technique better existing ones relying soft constraints. analyze reasons hope spur researchers proposing better solutions.",3
"discovering causal signals images. paper establishes existence observable footprints reveal ""causal dispositions"" object categories appearing collections images. achieve goal two steps. first, take learning approach observational causal discovery, build classifier achieves state-of-the-art performance finding causal direction pairs random variables, given samples joint distribution. second, use causal direction classifier effectively distinguish features objects features contexts collections static images. experiments demonstrate existence relation direction causality difference objects contexts, token, existence observable signals reveal causal dispositions objects.",18
"teleoperation system using past image records considering narrow communication band. teleoperation necessary robot applied real missions, example surveillance, search rescue. proposed teleoperation system using past image records (spir). spir virtually generates bird's-eye view image overlaying cg model robot corresponding current position background image captured camera mounted robot past time. problem spir communication bandwidth often narrow teleoperation tasks. case, candidates background image spir position robot often delayed. study, propose zoom function insufficiency candidates background image additional interpolation lines delay position data robot. evaluate proposed system, outdoor experiments carried out. outdoor experiment conducted training course driving school.",3
"learning temporal transformations time-lapse videos. based life-long observations physical, chemical, biologic phenomena natural world, humans often easily picture minds object look like future. but, computers? paper, learn computational models object transformations time-lapse videos. particular, explore use generative models create depictions objects future times. models explore several different prediction tasks: generating future state given single depiction object, generating future state given two depictions object different times, generating future states recursively recurrent framework. provide qualitative quantitative evaluations generated results, also conduct human evaluation compare variations models.",3
"optimal weighting multi-view data low dimensional hidden states. natural language processing (nlp) tasks, data often following two properties: first, data chopped multi-views successfully used dimension reduction purposes. example, topic classification, every paper chopped title, main text references. however, common views less noisier views supervised learning problems. second, unlabeled data easy obtain labeled data relatively rare. example, articles occurred new york times recent 10 years easy grab classified 'politics', 'finance' 'sports' need human labor. hence less noisy features preferred running supervised learning methods. paper propose unsupervised algorithm optimally weights features different views views generated low dimensional hidden state, occurs widely used models like mixture gaussian model, hidden markov model (hmm) latent dirichlet allocation (lda).",18
"quantum reinforcement learning. key approaches machine learning, especially learning unknown probabilistic environments new representations computation mechanisms. paper, novel quantum reinforcement learning (qrl) method proposed combining quantum theory reinforcement learning (rl). inspired state superposition principle quantum parallelism, framework value updating algorithm introduced. state (action) traditional rl identified eigen state (eigen action) qrl. state (action) set represented quantum superposition state eigen state (eigen action) obtained randomly observing simulated quantum state according collapse postulate quantum measurement. probability eigen action determined probability amplitude, parallelly updated according rewards. related characteristics qrl convergence, optimality balancing exploration exploitation also analyzed, shows approach makes good tradeoff exploration exploitation using probability amplitude speed learning quantum parallelism. evaluate performance practicability qrl, several simulated experiments given results demonstrate effectiveness superiority qrl algorithm complex problems. present work also effective exploration application quantum computation artificial intelligence.",17
"twisted self-similarity einstein vacuum equations. previous works [i. rodnianski y. shlapentokh-rothman, naked singularities einstein vacuum equations: exterior solution, arxiv:1912.08478 y. shlapentokh-rothman, naked singularities einstein vacuum equations: interior solution, arxiv:2204.09891] introduced new type self-similarity einstein vacuum equations characterized fact homothetic vector field may spacelike past light cone singularity. work give systematic treatment new self-similarity. particular, provide geometric characterizations spacetimes admitting new symmetry show existence uniqueness formal expansions around past null cone singularity may considered analogues well-known fefferman--graham expansions. combination previous results, analysis show twisted self-similar solutions sufficiently general describe possible asymptotic behaviors spacetimes small data regime self-similar whose homothetic vector field everywhere spacelike initial spacelike hypersurface. present application later fact understanding global structure fefferman--graham spacetimes naked singularity exteriors [i. rodnianski y. shlapentokh-rothman, naked singularities einstein vacuum equations: exterior solution, arxiv:1912.08478]. lastly, observe amalgamation techniques previous works, one may associate true solutions einstein vacuum equations formal expansions suitable region spacetime.",6
"mapping dialog act annotations lego corpus communicative functions iso 24617-2. paper present strategies mapping dialog act annotations lego corpus communicative functions iso 24617-2 standard. using strategies, obtained additional 347 dialogs annotated according standard. particularly important given reduced amount existing data conditions due recency standard. furthermore, dialogs widely explored corpus dialog related tasks. however, dialog annotations neglected due high domain-dependency, renders unuseful outside context corpus. thus, mapping process, obtain data annotated according recent standard provide useful dialog act annotations widely explored corpus context dialog research.",3
"effectiveness pre- inprocessing cdcl-based sat solving. applying pre- inprocessing techniques simplify cnf formulas search considerably improve performance modern sat solvers. algorithms mostly aim reducing number clauses, literals, variables formula. however, worthwhile, necessary additional runtime exceed runtime saved subsequent sat solver execution. paper investigate efficiency practicability selected simplification algorithms cdcl-based sat solving. first analyze means expected impact cnf formula sat solving all. testing real-world combinatorial sat instances, show techniques combinations yield desirable speedup ones avoided.",3
"compact molecular simulation quantum computers via combinatorial mapping variational state preparation. compact representations fermionic hamiltonians necessary perform calculations quantum computers lack error-correction. fermionic system typically defined within subspace fixed particle number spin unnecessary states projected hilbert space. provide bijective mapping using combinatoric ranking bijectively map fermion basis states qubit basis states express operators standard spin representation. evaluate compact mapping using variational quantum eigensolver (vqe) unitary coupled cluster singles doubles excitations (uccsd) ansatz compact representation. compactness beneficial orbital filling well away half, show 30 spin orbital $h_{2}$ calculation 8 qubits. find gate depth needed prepare compact wavefunction much greater full configuration space practice. notable observation regards number calls optimizer needed compact simulation compared full simulation. find compact representation converges faster full representation using adam optimizer cases. analysis demonstrates effect compact mapping practice.",17
"optimizing return secure disposal prescription opioids reduce diversion secondary users black market. opioid use disorder (oud) reached epidemic level us. unintended flow unused prescription opioids secondary users black market exacerbated public health risk. paper, aim address critical public health problem designing strategies return safe disposal unused prescription opioids. propose optimization framework determine optimal incentive disbursement plans locations easily accessible opioid disposal kiosks motivate prescription opioid users diverse patient profiles returning unused opioids. present mixed-integer non-linear programming (minlp) model solve decision problem, followed reformulation using benders decomposition results computationally efficient two-stage mixed-integer linear programming (milp) model. present case study show benefits usability model using dataset created massachusetts payer claims data (ma apcd). solve model prescription opioid user profiles, parameters associated incentive thresholds willingness return profile constructed dataset. numerical experiments demonstrate ability model usefulness determining optimal locations opioid disposal kiosks incentive disbursement plans maximizing disposal unused opioids. proposed optimization framework provide estimate total cost various trade strategies instrumental state agencies assess feasibility devise pragmatic policies reducing diversion unused prescription opioids.",10
"objective variables probabilistic revenue maximization second-price auctions reserve. many online companies sell advertisement space second-price auctions reserve. paper, develop probabilistic method learn profitable strategy set reserve price. use historical auction data features fit predictor best reserve price. problem delicate - structure auction reserve price set high much worse reserve price set low. address develop objective variables, new framework combining probabilistic modeling optimal decision-making. objective variables ""hallucinated observations"" transform revenue maximization task regularized maximum likelihood estimation problem, solve em algorithm. framework enables variety prediction mechanisms set reserve price. examples, study objective variable methods regression, kernelized regression, neural networks simulated real data. methods outperform previous approaches terms scalability profit.",18
"training adversarial discriminators cross-channel abnormal event detection crowds. abnormal crowd behaviour detection attracts large interest due importance video surveillance scenarios. however, ambiguity lack sufficient ""abnormal"" ground truth data makes end-to-end training large deep networks hard domain. paper propose use generative adversarial nets (gans), trained generate ""normal"" distribution data. adversarial gan training, discriminator ""d"" used supervisor generator network ""g"" vice versa. testing time use ""d"" solve discriminative task (abnormality detection), ""d"" trained without need manually-annotated abnormal data. moreover, order prevent ""g"" learn trivial identity function, use cross-channel approach, forcing ""g"" transform raw-pixel data motion information vice versa. quantitative results standard benchmarks show method outperforms previous state-of-the-art methods frame-level pixel-level evaluation.",3
"social participation ontology: community documentation, enhancements use examples. participatory democracy advances virtually governments especially south america exhibits mixed culture social predisposition. article presents ""social participation ontology"" (ops brazilian name \emph{ontologia de participa\c{c}\~ao social}) implemented compliance web ontology language standard (owl) fostering social participation, specially virtual platforms. entities links ops defined based extensive collaboration specialists. shown ops instrumental information retrieval contents portal, terms actors (at various levels) well mechanisms activities. significantly, ops linked owl ontologies upper ontology via foaf bfo higher upper ontologies, yields sound organization access knowledge data. order illustrate usefulness ops, present results ontological expansion integration ontologies data. ongoing work involves adoption ops official brazilian federal portal social participation ngo s, linkage ontologies social participation.",3
"scale-based gaussian coverings: combining intra inter mixture models image segmentation. ""covering"" mean gaussian mixture model fit observed data. approximations bayes factor availed judge model fit data within given gaussian mixture model. families gaussian mixture models, propose r\'enyi quadratic entropy excellent tractable model comparison framework. exemplify using segmentation mri image volume, based (1) direct gaussian mixture model applied marginal distribution function, (2) gaussian model fit k-means applied 4d multivalued image volume furnished wavelet transform. visual preference one model another immediate. r\'enyi quadratic entropy allows us show clearly one modelings superior other.",3
"generating visual representations zero-shot classification. paper addresses task learning image clas-sifier categories defined semantic descriptions (e.g. visual attributes) others defined exemplar images well. task often referred zero-shot classification task (zsc). previous methods rely learning common embedding space allowing compare visual features unknown categories semantic descriptions. paper argues approaches limited i) efficient discrimi-native classifiers can't used ii) classification tasks seen unseen categories (generalized zero-shot classification gzsc) can't addressed efficiently. contrast , paper suggests address zsc gzsc i) learning conditional generator using seen classes ii) generate artificial training examples categories without exemplars. zsc turned standard supervised learning problem. experiments 4 generative models 5 datasets experimentally validate approach, giving state-of-the-art results zsc gzsc.",3
"sky detection log illumination refinement pde-based hazy image contrast enhancement. report presents results sky detection technique used improve performance previously developed partial differential equation (pde)-based hazy image enhancement algorithm. additionally, proposed alternative method utilizes function log illumination refinement improve de-hazing results avoiding over-enhancement sky homogeneous regions. algorithms tested several benchmark calibration images compared several standard algorithms literature. results indicate algorithms yield mostly consistent results surpasses several algorithms terms colour contrast enhancement addition improved edge visibility.",3
"importance sampling bayesian networks: influence-based approximation strategy importance functions. one main problems importance sampling bayesian networks representation importance function, ideally close possible posterior joint distribution. typically, represent importance function factorization, i.e., product conditional probability tables (cpts). given diagnostic evidence, explicit forms cpts networks. first derive exact form cpts optimal importance function. since calculation hard, usually use approximations. review several popular strategies point limitations. based analysis influence evidence, propose method approximating exact form importance function explicitly modeling important additional dependence relations introduced evidence. experimental results show new approximation strategy offers immediate improvement quality importance function.",3
"bernstein function: unifying framework nonconvex penalization sparse estimation. paper study nonconvex penalization using bernstein functions. since bernstein function concave nonsmooth origin, induce class nonconvex functions high-dimensional sparse estimation problems. derive threshold function based bernstein penalty give mathematical properties sparsity modeling. show coordinate descent algorithm especially appropriate penalized regression problems bernstein penalty. additionally, prove bernstein function defined concave conjugate $\varphi$-divergence develop conjugate maximization algorithm finding sparse solution. finally, particularly exemplify family bernstein nonconvex penalties based generalized gamma measure conduct empirical analysis family.",18
"investigating practical linear temporal difference learning. off-policy reinforcement learning many applications including: learning demonstration, learning multiple goal seeking policies parallel, representing predictive knowledge. recently proliferation new policy-evaluation algorithms fill longstanding algorithmic void reinforcement learning: combining robustness off-policy sampling, function approximation, linear complexity, temporal difference (td) updates. paper contains two main contributions. first, derive two new hybrid td policy-evaluation algorithms, fill gap collection algorithms. second, perform empirical comparison elicit new linear td methods preferred different situations, make concrete suggestions practical use.",3
"bayesian methodology estimating uncertainty decisions safety-critical systems. uncertainty decisions safety-critical engineering applications estimated basis bayesian markov chain monte carlo (mcmc) technique averaging decision models. use decision tree (dt) models assists experts interpret causal relations find factors uncertainty. bayesian averaging also allows experts estimate uncertainty accurately priori information favored structure dts available. expert select single dt model, typically maximum posteriori model, interpretation purposes. unfortunately, priori information favored structure dts always available. reason, suggest new prior dts bayesian mcmc technique. also suggest new procedure selecting single dt describe application scenario. experiments short-term conflict alert data technique outperforms existing bayesian techniques predictive accuracy selected single dts.",3
"nonnegative matrix factorization. nonnegative matrix factorization (nmf) become widely used tool analysis high-dimensional data automatically extracts sparse meaningful features set nonnegative data vectors. first illustrate property nmf three applications, image processing, text mining hyperspectral imaging --this why. address problem solving nmf, np-hard general. review standard nmf algorithms, also present recent subclass nmf problems, referred near-separable nmf, solved efficiently (that is, polynomial time), even presence noise --this how. finally, briefly describe problems mathematics computer science closely related nmf via nonnegative rank.",18
"guarantees greedy maximization non-submodular functions applications. investigate performance standard greedy algorithm cardinality constrained maximization non-submodular nondecreasing set functions. strong theoretical guarantees performance greedy maximizing submodular functions, guarantees non-submodular ones. however, greedy enjoys strong empirical performance many important non-submodular functions, e.g., bayesian a-optimality objective experimental design. prove theoretical guarantees supporting empirical performance. guarantees characterized combination (generalized) curvature $\alpha$ submodularity ratio $\gamma$. particular, prove greedy enjoys tight approximation guarantee $\frac{1}{\alpha}(1- e^{-\gamma\alpha})$ cardinality constrained maximization. addition, bound submodularity ratio curvature several important real-world objectives, including bayesian a-optimality objective, determinantal function square submatrix certain linear programs combinatorial constraints. experimentally validate theoretical findings synthetic real-world applications.",3
"glop: global local penalty capturing predictive heterogeneity. faced supervised learning problem, hope rich enough data build model predicts future instances well. however, practice, problems exhibit predictive heterogeneity: instances might relatively easy predict, others might predictive outliers model trained entire dataset perform well. identifying help focus future data collection. present glop, global local penalty, framework capturing predictive heterogeneity identifying predictive outliers. glop based penalized regression multitask learning, improves learning leveraging training signal information related tasks. give two optimization algorithms glop, one space-efficient, another giving full regularization path. also characterize uniqueness terms data tuning parameters, present empirical results synthetic data two health research problems.",18
"linear maximum margin classifier learning uncertain data. paper, propose maximum margin classifier deals uncertainty data input. specifically, reformulate svm framework training example modeled multi-dimensional gaussian distribution described mean vector covariance matrix -- latter modeling uncertainty. address classification problem define cost function expected value classical svm cost data samples drawn multi-dimensional gaussian distributions form set training examples. formulation approximates classical svm formulation training examples isotropic gaussians variance tending zero. arrive convex optimization problem, solve efficiently primal form using stochastic gradient descent approach. resulting classifier, name svm gaussian sample uncertainty (svm-gsu), tested synthetic data five publicly available popular datasets; namely, mnist, wdbc, deap, tv news channel commercial detection, trecvid med datasets. experimental results verify effectiveness proposed method.",3
"virtual cnn branching: efficient feature ensemble person re-identification. paper introduce ensemble method convolutional neural network (cnn), called ""virtual branching,"" implemented nearly additional parameters computation top standard cnns. propose method context person re-identification (re-id). cnn model consists shared bottom layers, followed ""virtual"" branches, neurons block regular convolutional fully-connected layers partitioned multiple sets. virtual branch trained different data specialize different aspects, e.g., specific body region pose orientation. way, robust ensemble representations obtained human body misalignment, deformations, variations viewing angles, nearly additional cost. proposed method achieves competitive performance multiple person re-id benchmark datasets, including market-1501, cuhk03, dukemtmc-reid.",3
"learning latent variable gaussian graphical models. gaussian graphical models (ggm) widely used many high-dimensional applications ranging biological financial data recommender systems. sparsity ggm plays central role statistically computationally. unfortunately, real-world data often fit well sparse graphical models. paper, focus family latent variable gaussian graphical models (lvggm), model conditionally sparse given latent variables, marginally non-sparse. lvggm, inverse covariance matrix low-rank plus sparse structure, learned regularized maximum likelihood framework. derive novel parameter estimation error bounds lvggm mild conditions high-dimensional setting. results complement existing theory structural learning, open new possibilities using lvggm statistical inference.",18
"information directed sampling bandits heteroscedastic noise. stochastic bandit problem, goal maximize unknown function via sequence noisy function evaluations. typically, observation noise assumed independent evaluation point satisfies tail bound taken uniformly domain. work, consider setting heteroscedastic noise, is, explicitly allow noise distribution depend evaluation point. show leads new trade-offs information regret, taken account existing approaches like upper confidence bound algorithms (ucb) thompson sampling. address shortcomings, introduce frequentist regret framework, similar bayesian analysis russo van roy (2014). prove new high-probability regret bound general, possibly randomized policies, depending quantity call regret-information ratio. bound, define frequentist version information directed sampling (ids) minimize surrogate regret-information ratio possible action sampling distributions. order construct surrogate function, generalize known concentration inequalities least squares regression separable hilbert spaces case heteroscedastic noise. allows us formulate several variants ids linear reproducing kernel hilbert space response functions, yielding family novel algorithms bayesian optimization. also provide frequentist regret bounds, homoscedastic case comparable existing bounds ucb, much better noise heteroscedastic. finally, empirically demonstrate linear setting, methods outperform ucb thompson sampling, even noise homoscedastic.",18
"spudd: stochastic planning using decision diagrams. markov decisions processes (mdps) becoming increasing popular models decision theoretic planning. traditional dynamic programming methods perform well problems small state spaces, structured methods needed large problems. propose examine value iteration algorithm mdps uses algebraic decision diagrams(adds) represent value functions policies. mdp represented using bayesian networks adds dynamic programming applied directly adds. demonstrate method large mdps (up 63 million states) show significant gains compared tree-structured representations (with thirty-fold reduction number nodes required represent optimal value functions).",3
"prediction scores window classifier behavior. multi-class classifiers make prediction test sample scoring classes selecting one highest score. analyzing prediction scores useful understand classifier behavior assess reliability. present interactive visualization facilitates per-class analysis scores. system, called classilist, enables relating scores classification correctness underlying samples features. illustrate analysis reveals varying behavior different classifiers. classilist available use online, along source code, video tutorials, plugins r, rapidminer, knime https://katehara.github.io/classilist-site/.",18
"dance dance convolution. dance dance revolution (ddr) popular rhythm-based video game. players perform steps dance platform synchronization music directed on-screen step charts. many step charts available standardized packs, players may grow tired existing charts, wish dance song chart exists. introduce task learning choreograph. given raw audio track, goal produce new step chart. task decomposes naturally two subtasks: deciding place steps deciding steps select. step placement task, combine recurrent convolutional neural networks ingest spectrograms low-level audio features predict steps, conditioned chart difficulty. step selection, present conditional lstm generative model substantially outperforms n-gram fixed-window approaches.",3
"sentiment analysis twitter data: survey techniques. advancement web technology growth, huge volume data present web internet users lot data generated too. internet become platform online learning, exchanging ideas sharing opinions. social networking sites like twitter, facebook, google+ rapidly gaining popularity allow people share express views topics,have discussion different communities, post messages across world. lot work field sentiment analysis twitter data. survey focuses mainly sentiment analysis twitter data helpful analyze information tweets opinions highly unstructured, heterogeneous either positive negative, neutral cases. paper, provide survey comparative analyses existing techniques opinion mining like machine learning lexicon-based approaches, together evaluation metrics. using various machine learning algorithms like naive bayes, max entropy, support vector machine, provide research twitter data streams.general challenges applications sentiment analysis twitter also discussed paper.",3
"full analytic expression overlap reduction function gravitational wave background pulsar timing arrays. pulsar timing array (pta) expected detect gravitational wave background (gwb) nanohertz band within next decade. provides opportunity test gravity theory cosmology. typical data analysis method detect gwb cross-correlation analysis. overlap reduction function (orf) plays important role correlation data analysis gwb. present approach dealing intricate integration orf use short-wave approximation drop tricky terms. paper, provide full analytic expression orf pta without approximation possible polarizations allowed modifications general relativity. compared numerical simulation short-wave approximation, results efficient widely applicable. especially scalar-longitudinal mode short-wave approximation available, analytical expression particularly significant.",6
"unbiased bayesian inference population markov jump processes via random truncations. consider continuous time markovian processes populations individual agents interact stochastically according kinetic rules. despite increasing prominence models fields ranging biology smart cities, bayesian inference systems remains challenging, continuous time, discrete state systems potentially infinite state-space. propose novel efficient algorithm joint state / parameter posterior sampling population markov jump processes. introduce class pseudo-marginal sampling algorithms based random truncation method enables principled treatment infinite state spaces. extensive evaluation number benchmark models shows approach achieves considerable savings compared state art methods, retaining accuracy fast convergence. also present results synthetic biology data set showing potential practical usefulness work.",18
"atomnet: deep convolutional neural network bioactivity prediction structure-based drug discovery. deep convolutional neural networks comprise subclass deep neural networks (dnn) constrained architecture leverages spatial temporal structure domain model. convolutional networks achieve best predictive performance areas speech image recognition hierarchically composing simple local features complex models. although dnns used drug discovery qsar ligand-based bioactivity predictions, none models benefited powerful convolutional architecture. paper introduces atomnet, first structure-based, deep convolutional neural network designed predict bioactivity small molecules drug discovery applications. demonstrate apply convolutional concepts feature locality hierarchical composition modeling bioactivity chemical interactions. contrast existing dnn techniques, show atomnet's application local convolutional filters structural target information successfully predicts new active molecules targets previously known modulators. finally, show atomnet outperforms previous docking approaches diverse set benchmarks large margin, achieving auc greater 0.9 57.8% targets dude benchmark.",3
"automated epilepsy diagnosis using interictal scalp eeg. approximately 50 million people worldwide suffer epilepsy. traditional diagnosis epilepsy relies tedious visual screening highly trained clinicians lengthy eeg recording contains presence seizure (ictal) activities. nowadays, many automatic systems recognize seizure-related eeg signals help diagnosis. however, costly inconvenient obtain long-term eeg data seizure activities, especially areas short medical resources. demonstrate paper use interictal scalp eeg data, much easier collect ictal data, automatically diagnose whether person epileptic. automated eeg recognition system, extract three classes features eeg data build probabilistic neural networks (pnns) fed features. optimize feature extraction parameters combine pnns voting mechanism. result, system achieves impressive 94.07% accuracy, close reported human recognition accuracy experienced medical professionals.",3
"fair deep learning prediction healthcare applications confounder filtering. rapid development deep learning methods permitted fast accurate medical decision making complex structured data, like ct images mri. however, problems still exist applications may lead imperfect predictions. previous observations shown that, confounding factors, handled inappropriately, lead biased prediction results towards major properties data distribution. words, naively applying deep learning methods applications lead unfair prediction results minority group defined characteristics including age, gender, even hospital collects data, etc. paper, extending previous successes correcting confounders, propose stable method, namely confounder filtering, effectively reduce influence confounding factors, leading better generalizability trained discriminative deep neural networks, therefore, fairer prediction results. experimental results indicate confounder filtering method able improve performance different neural networks including cnn, lstm, arbitrary architecture, different data types including ct-scan, mri, eeg brain wave data, well different confounding factors including age, gender, physical factors medical devices etc",18
"robust kronecker-decomposable component analysis low-rank modeling. dictionary learning component analysis part one well-studied active research fields, intersection signal image processing, computer vision, statistical machine learning. dictionary learning, current methods choice arguably k-svd variants, learn dictionary (i.e., decomposition) sparse coding via singular value decomposition. robust component analysis, leading methods derive principal component pursuit (pcp), recovers low-rank matrix sparse corruptions unknown magnitude support. however, k-svd sensitive presence noise outliers training set. additionally, pcp provide dictionary respects structure data (e.g., images), requires expensive svd computations solved convex relaxation. paper, introduce new robust decomposition images combining ideas sparse dictionary learning pcp. propose novel kronecker-decomposable component analysis robust gross corruption, used low-rank modeling, leverages separability solve significantly smaller problems. design efficient learning algorithm drawing links restricted form tensor factorization. effectiveness proposed approach demonstrated real-world applications, namely background subtraction image denoising, performing thorough comparison current state art.",18
"taming primary key violations query large inconsistent data. consistent query answering database violates primary key constraints classical hard problem database research traditionally dealt logic programming. however, applicability existing logic-based solutions restricted data sets moderate size. paper presents novel decomposition pruning strategy reduces, polynomial time, problem computing consistent answer conjunctive query database subject primary key constraints collection smaller problems sort solved independently. new strategy naturally modeled implemented using answer set programming (asp). experiment run benchmarks database world prove effectiveness efficiency asp-based approach also large data sets. appear theory practice logic programming (tplp), proceedings iclp 2015.",3
"orientations cycles supersingular isogeny graphs. paper concerns several theoretical aspects oriented supersingular l-isogeny volcanoes relationship closed walks supersingular l-isogeny graph. main result bijection rims union oriented supersingular l-isogeny volcanoes $\bar{\mathbb{f}}_p$ (up conjugation orientations), isogeny cycles (non-backtracking closed walks powers smaller walks) supersingular l-isogeny graph modulo p. exact proof statement bijection made intricate special behaviours arising extra automorphisms ramification p certain quadratic orders. use bijection count isogeny cycles given length supersingular l-isogeny graph exactly sum class numbers, also give explicit upper bound estimating class numbers.",10
"intensity video guided 4d fusion improved highly dynamic 3d reconstruction. availability high-speed 3d video sensors greatly facilitated 3d shape acquisition dynamic deformable objects, high frame rate 3d reconstruction always degraded spatial noise temporal fluctuations. paper presents simple yet powerful intensity video guided multi-frame 4d fusion pipeline. temporal tracking intensity image points (of moving deforming objects) allows registration corresponding 3d data points, whose 3d noise fluctuations reduced spatio-temporal multi-frame 4d fusion. conducted simulated noise tests real experiments four 3d objects using 1000 fps 3d video sensor. results demonstrate proposed algorithm effective reducing 3d noise robust intensity noise. outperforms existing algorithms good scalability stationary dynamic objects.",3
"quantum density matrix theory laser without adiabatic elimination population inversion: transition lasing class-b limit. despite enormous technological interest micro nanolasers, surprisingly, class-b quantum density-matrix model available date, capable accurately describing coherence photon correlations within unified theory. class-b lasers $-$applicable solid-state lasers room temperature$-$, macroscopic polarization decay rate larger cavity damping rate which, turn, exceeds upper level population decay rate. carry density-matrix theoretical approach generic class-b lasers, provide closed equations photonic atomic reduced density matrix fock basis photons. relatively simple model numerically integrated straightforward way, exhibits expected phenomena, one-atom photon antibunching, well-known s-shaped input-output laser emission super-poissonian autocorrelation many atoms ($1\leq g^{(2)}(0)\leq 2$), photons (large spontaneous emission factors, $\beta\sim1$) thermodynamic limit ($n\gg1$ $\beta\sim 0$). based analysis $g^{(2)}(\tau)$, conclude super-poissonian fluctuations clearly related relaxation oscillations photon number. predict strong damping relaxation oscillations atom number small $n\sim 10$. model enables study few-photon bifurcations non-classical photon correlations class-b laser devices, also leveraging quantum descriptions coherently coupled nanolaser arrays.",17
"stein variational message passing continuous graphical models. propose novel distributed inference algorithm continuous graphical models, extending stein variational gradient descent (svgd) leverage markov dependency structure distribution interest. approach combines svgd set structured local kernel functions defined markov blanket node, alleviates curse high dimensionality simultaneously yields distributed algorithm decentralized inference tasks. justify method theoretical analysis show use local kernels viewed new type localized approximation matches target distribution conditional distributions node markov blanket. empirical results show method outperforms variety baselines including standard mcmc particle message passing methods.",18
"deep learning architecture temporal sleep stage classification using multivariate multimodal time series. sleep stage classification constitutes important preliminary exam diagnosis sleep disorders. traditionally performed sleep expert assigns 30s signal sleep stage, based visual inspection signals electroencephalograms (eeg), electrooculograms (eog), electrocardiograms (ecg) electromyograms (emg). introduce first deep learning approach sleep stage classification learns end-to-end without computing spectrograms extracting hand-crafted features, exploits multivariate multimodal polysomnography (psg) signals (eeg, emg eog), exploit temporal context 30s window data. modality first layer learns linear spatial filters exploit array sensors increase signal-to-noise ratio, last layer feeds learnt representation softmax classifier. model compared alternative automatic approaches based convolutional networks decisions trees. results obtained 61 publicly available psg records 20 eeg channels demonstrate network architecture yields state-of-the-art performance. study reveals number insights spatio-temporal distribution signal interest: good trade-off optimal classification performance measured balanced accuracy use 6 eeg 2 eog (left right) 3 emg chin channels. also exploiting one minute data data segment offers strongest improvement limited number channels available. sleep experts, system exploits multivariate multimodal nature psg signals order deliver state-of-the-art classification performance small computational cost.",18
"learning square loss: localization offset rademacher complexity. consider regression square loss general classes functions without boundedness assumption. introduce notion offset rademacher complexity provides transparent way study localization expectation high probability. (possibly non-convex) class, excess loss two-step estimator shown upper bounded offset complexity novel geometric inequality. convex case, estimator reduces empirical risk minimizer. method recovers results \citep{raksritsy15} bounded case also providing guarantees without boundedness assumption.",18
"sar image segmentation using vector quantization technique entropy images. development application various remote sensing platforms result production huge amounts satellite image data. therefore, increasing need effective querying browsing image databases. order take advantage make good use satellite images data, must able extract meaningful information imagery. hence proposed new algorithm sar image segmentation. paper propose segmentation using vector quantization technique entropy image. initially, obtain entropy image second step use kekre's fast codebook generation (kfcg) algorithm segmentation entropy image. thereafter, codebook size 128 generated entropy image. code vectors clustered 8 clusters using kfcg algorithm converted 8 images. 8 images displayed result. approach lead segmentation segmentation. compared results well known gray level co-occurrence matrix. proposed algorithm gives better segmentation less complexity.",3
"language-depedent i-vectors lre15. standard recipe spoken language recognition apply gaussian back-end i-vectors. ignores uncertainty i-vector extraction, could important especially short utterances. recent paper cumani, plchot fer proposes solution propagate uncertainty backend. propose alternative method propagating uncertainty.",18
"restricting weak-generative capacity synchronous tree-adjoining grammars. formalism synchronous tree-adjoining grammars, variant standard tree-adjoining grammars (tag), intended allow use tags language transduction addition language specification. previous work, definition transduction relation defined synchronous tag given appeal iterative rewriting process. rewriting definition derivation problematic greatly extends expressivity formalism makes design parsing algorithms difficult impossible. introduce simple, natural definition synchronous tree-adjoining derivation, based isomorphisms standard tree-adjoining derivations, avoids expressivity implementability problems original rewriting definition. decrease expressivity, would otherwise make method unusable, offset incorporation alternative definition standard tree-adjoining derivation, previously proposed completely separate reasons, thereby making practical entertain using natural definition synchronous derivation. nonetheless, remaining problematic cases call yet flexibility definition; isomorphism requirement may relaxed. remains future research tune exact requirements allowable mappings.",1
"people around me: face discovery egocentric photo-streams. given unconstrained stream images captured wearable photo-camera (2fpm), propose unsupervised bottom-up approach automatic clustering appearing faces individual identities present data. problem challenging since images acquired real world conditions; hence visible appearance people images undergoes intensive variations. proposed pipeline consists first arranging photo-stream events, later, localizing appearance multiple people them, finally, grouping various appearances person across different events. experimental results performed dataset acquired wearing photo-camera one month, demonstrate effectiveness proposed approach considered purpose.",3
"gaia data release 3: gaia andromeda photometric survey. context. part gaia data release 3 (gaia dr3), epoch photometry released 1.2 million sources centred m31. taster gaia data release 4 epoch photometry released. aims. paper content gaia andromeda photometric survey described, including statistics assess quality data. known issues photometry also outlined. methods. methods given improve interpretation photometry, particular, method error renormalization. also, use correlations three photometric passbands allows clearer identification variables affected false detections caused systematic effects. results. gaia andromeda photometric survey presents unique opportunity look gaia epoch photometry preselected due variability. allows investigations carried applied rest sky using mean source results. additionally scientific studies variability carried m31 milky way general.",0
"optimal size, freshness time-frame voice search vocabulary. paper, investigate optimize vocabulary voice search language model. metric optimize out-of-vocabulary (oov) rate since strong indicator user experience. departure usual way measuring oov rates, web search logs allow us compute per-session oov rate thus estimate percentage users experience given oov rate. conservative text normalization, find voice search vocabulary consisting 2 2.5 million words extracted 1 week search query data result aggregate oov rate 1%; size, oov rate also experienced 90% users. number words included vocabulary stable indicator oov rate. altering freshness vocabulary duration time window training data gathered significantly change oov rate. surprisingly, significantly larger vocabulary (approximately 10 million words) required guarantee oov rates 1% 95% users.",3
"many vote operations needed manipulate voting system?. paper, propose framework study general class strategic behavior voting, call vote operations. prove following theorem: fix number alternatives, generate $n$ votes i.i.d. according distribution $\pi$, let $n$ go infinity, $\epsilon >0$, probability least $1-\epsilon$, minimum number operations needed strategic individual achieve goal falls one following four categories: (1) 0, (2) $\theta(\sqrt n)$, (3) $\theta(n)$, (4) $\infty$. theorem holds set vote operations, individual vote distribution $\pi$, integer generalized scoring rule, includes (but limited to) almost commonly studied voting rules, e.g., approval voting, positional scoring rules (including borda, plurality, veto), plurality runoff, bucklin, copeland, maximin, stv, ranked pairs. also show many well-studied types strategic behavior fall framework, including (but limited to) constructive/destructive manipulation, bribery, control adding/deleting votes, margin victory, minimum manipulation coalition size. therefore, main theorem naturally applies problems.",3
"indian spontaneous expression database emotion recognition. automatic recognition spontaneous facial expressions major challenge field affective computing. head rotation, face pose, illumination variation, occlusion etc. attributes increase complexity recognition spontaneous expressions practical applications. effective recognition expressions depends significantly quality database used. well-known facial expression databases consist posed expressions. however, currently huge demand spontaneous expression databases pragmatic implementation facial expression recognition algorithms. paper, propose establish new facial expression database containing spontaneous expressions male female participants indian origin. database consists 428 segmented video clips spontaneous facial expressions 50 participants. experiment, emotions induced among participants using emotional videos simultaneously self-ratings collected experienced emotion. facial expression clips annotated carefully four trained decoders, validated nature stimuli used self-report emotions. extensive analysis carried database using several machine learning algorithms results provided future reference. spontaneous database help development validation algorithms recognition spontaneous expressions.",3
"improving spectral clustering using asymptotic value normalised cut. spectral clustering popular versatile clustering method based relaxation normalised graph cut objective. despite popularity, however, single agreed upon method tuning important scaling parameter, determining automatically number clusters extract. popular heuristics exist, corresponding theoretical results scarce. paper investigate asymptotic value normalised cut increasing sample assumed arise underlying probability distribution, based result provide recommendations improving spectral clustering methodology. corresponding algorithm proposed strong empirical performance.",18
"clustering co-occurrence maximal frequent patterns streams. one way getting better view data using frequent patterns. paper frequent patterns subsets occur minimal number times stream itemsets. however, discovery frequent patterns streams always problematic. streams potentially endless principle impossible say pattern often occurring not. furthermore number patterns huge good overview structure stream lost quickly. proposed approach use clustering facilitate analysis structure stream. clustering co-occurrence patterns give user improved view structure stream. patterns might occur much together form combined pattern. way patterns clustering largest frequent patterns: maximal frequent patterns. approach decide patterns occur often together based method clustering distance pairs known. number maximal frequent patterns much smaller combined clustering methods patterns provide good view structure stream.",3
"detecting interrogative utterances recurrent neural networks. paper, explore different neural network architectures predict speaker given utterance asking question making statement. com- pare outcomes regularization methods popularly used train deep neural networks study different context functions affect classification performance. also compare efficacy gated activation functions favorably used recurrent neural networks study combine multimodal inputs. evaluate models two multimodal datasets: msr-skype callhome.",3
"evolution representation simple cognitive networks. representations internal models environment provide guidance behaving agent, even absence sensory information. clear representations developed whether necessary even essential intelligent behavior. argue ability represent relevant features environment expected consequence adaptive process, give formal definition representation based information theory, quantify measure r. measure r changes time, evolve two types networks---an artificial neural network network hidden markov gates---to solve categorization task using genetic algorithm. find capacity represent increases evolutionary adaptation, agents form representations environment lifetime. ability allows agents act sensorial inputs context acquired representations enables complex context-dependent behavior. examine concepts (features environment) networks representing, representations logically encoded networks, form agent behaves solve task. conclude r able quantify representations within cognitive system, predictive agent's long-term adaptive success.",15
"computational model short-cut rule 2d shape decomposition. propose new 2d shape decomposition method based short-cut rule. short-cut rule originates cognition research, states human visual system prefers partition object parts using shortest possible cuts. propose implement computational model short-cut rule apply problem shape decomposition. model proposed generates set cut hypotheses passing points silhouette represent negative minima curvature. show part-cut hypotheses eliminated analysis local properties each. finally, remaining hypotheses evaluated ascending length order, guarantees pair conflicting cuts shortest accepted. demonstrate that, compared state-of-the-art shape decomposition methods, proposed approach achieves decomposition results better correspond human intuition revealed psychological experiments.",3
"new approach building interindustry input--output table. present new approach estimating interdependence industries economy applying data science solutions. exploiting interfirm buyer--seller network data, show problem estimating interdependence industries similar problem uncovering latent block structure network science literature. estimate underlying structure greater accuracy, propose extension sparse block model incorporates node textual information unbounded number industries interactions among them. latter task accomplished extending well-known chinese restaurant process two dimensions. inference based collapsed gibbs sampling, model evaluated synthetic real-world datasets. show proposed model improves predictive accuracy successfully provides satisfactory solution motivated problem. also discuss issues affect future performance approach.",18
"quantum learning: optimal classification qubit states. pattern recognition central topic learning theory numerous applications voice text recognition, image analysis, computer diagnosis. statistical set-up classification following: given i.i.d. training set $(x_{1},y_{1}),... (x_{n},y_{n})$ $x_{i}$ represents feature $y_{i}\in \{0,1\}$ label attached feature. underlying joint distribution $(x,y)$ unknown, learn training set aim devising low error classifiers $f:x\to y$ used predict label new incoming features. solve quantum analogue problem, namely classification two arbitrary unknown qubit states. given number `training' copies states, would like `learn' performing measurement training set. outcome used design mesurements classification future systems unknown labels. find asymptotically optimal classification strategy show typically, performs strictly better plug-in strategy based state estimation. figure merit excess risk difference probability error probability error optimal measurement states known, helstrom measurement. show excess risk rate $n^{-1}$ compute exact constant rate.",17
"black box modelling hvac system : improving performances neural networks. paper deals neural networks modelling hvac systems. order increase neural networks performances, method based sensitivity analysis applied. technique also used compute relevance input. avoid prediction errors dry coil conditions, metamodel capacity derived neural networks. regression coefficients polynomial forms identified use spectral analysis. methods based sensitivity spectral analysis lead optimized neural network model, regard architecture predictions.",3
"getting feasible variable estimates infeasible ones: mrf local polytope study. paper proposes method construction approximate feasible primal solutions dual ones large-scale optimization problems possessing certain separability properties. whereas infeasible primal estimates typically produced (sub-)gradients dual function, often easy project primal feasible set, since projection complexity comparable complexity initial problem. propose alternative efficient method obtain feasibility show properties influencing convergence optimum similar properties euclidean projection. apply method local polytope relaxation inference problems markov random fields demonstrate superiority existing methods.",3
"learning equivalence classes bayesian networks structures. approaches learning bayesian networks data typically combine scoring function heuristic search procedure. given bayesian network structure, many scoring functions derived literature return score entire equivalence class structure belongs. using scoring function, appropriate heuristic search algorithm search equivalence classes bayesian networks opposed individual structures. present general formulation search space states search correspond equivalence classes structures. using space, one number heuristic search algorithms easily applied. compare greedy search performance proposed search space greedy search performance search space states correspond individual bayesian network structures.",3
"cluster based symbolic representation skewed text categorization. work, problem associated imbalanced text corpora addressed. method converting imbalanced text corpus balanced one presented. presented method employs clustering algorithm conversion. initially avoid curse dimensionality, effective representation scheme based term class relevancy measure adapted, drastically reduces dimension number classes corpus. subsequently, samples larger sized classes grouped number subclasses smaller sizes make entire corpus balanced. subclass given single symbolic vector representation use interval valued features. symbolic representation addition compact helps reducing space requirement also classification time. proposed model empirically demonstrated superiority bench marking datasets viz., reuters 21578 tdt2. further, compared several existing contemporary models including model based support vector machine. comparative analysis indicates proposed model outperforms existing models.",3
"bridge type classification: supervised learning modified nbi dataset. key phase bridge design process selection structural system. due budget time constraints, engineers typically rely engineering judgment prior experience selecting structural system, often considering limited range design alternatives. objective study explore suitability supervised machine learning preliminary design aid provides guidance engineers regards statistically optimal bridge type choose, ultimately improving likelihood optimized design, design standardization, reduced maintenance costs. order devise supervised learning system, data 600,000 bridges national bridge inventory database analyzed. key attributes determining bridge structure type identified three feature selection techniques. potentially useful attributes like seismic intensity historic data cost materials (steel concrete) added us geological survey (usgs) database engineering news record. decision tree, bayes network support vector machines used predicting bridge design type. due state-to-state variations material availability, material costs, design codes, supervised learning models based complete data set yield favorable results. supervised learning models trained tested using 10-fold cross validation data state. inclusion seismic data improved model performance noticeably. data resampled reduce bias models towards common design types, supervised learning models thus constructed showed improvements performance. average recall precision state models 88.6% 88.0% using decision trees, 84.0% 83.7% using bayesian networks, 80.8% 75.6% using svm.",18
"machine learning approach rf transmitter identification. development widespread use wireless devices recent years (mobile phones, internet things, wi-fi), electromagnetic spectrum become extremely crowded. order counter security threats posed rogue unknown transmitters, important identify rf transmitters data content transmissions based intrinsic physical characteristics transmitters. rf waveforms represent particular challenge extremely high data rates involved potentially large number transmitters present given location. factors outline need rapid fingerprinting identification methods go beyond traditional hand-engineered approaches. study, investigate use machine learning (ml) strategies classification identification problems, use wavelets reduce amount data required. four different ml strategies evaluated: deep neural nets (dnn), convolutional neural nets (cnn), support vector machines (svm), multi-stage training (mst) using accelerated levenberg-marquardt (a-lm) updates. a-lm mst method preconditioned wavelets far accurate, achieving 100% classification accuracy transmitters, tested using data originating 12 different transmitters. discuss strategies extension mst much larger number transmitters.",5
"detecting vanishing points using global image context non-manhattan world. propose novel method detecting horizontal vanishing points zenith vanishing point man-made environments. dominant trend existing methods first find candidate vanishing points, remove outliers enforcing mutual orthogonality. method reverses process: propose set horizon line candidates score based vanishing points contains. key element approach use global image context, extracted deep convolutional network, constrain set candidates consideration. method make manhattan-world assumption operate effectively scenes single horizontal vanishing point. evaluate approach three benchmark datasets achieve state-of-the-art performance each. addition, approach significantly faster previous best method.",3
"first order algorithms variational image processing. variational methods imaging nowadays developing towards quite universal flexible tool, allowing highly successful approaches tasks like denoising, deblurring, inpainting, segmentation, super-resolution, disparity, optical flow estimation. overall structure approaches form ${\cal d}(ku) + \alpha {\cal r} (u) \rightarrow \min_u$ ; functional ${\cal d}$ data fidelity term also depending input data $f$ measuring deviation $ku$ ${\cal r}$ regularization functional. moreover $k$ (often linear) forward operator modeling dependence data underlying image, $\alpha$ positive regularization parameter. ${\cal d}$ often smooth (strictly) convex, current practice almost exclusively uses nonsmooth regularization functionals. majority successful techniques using nonsmooth convex functionals like total variation generalizations thereof $\ell_1$-norms coefficients arising scalar products frame system. efficient solution variational problems imaging demands appropriate algorithms. taking account specific structure sum two different terms minimized, splitting algorithms quite canonical choice. consequently field revived interest techniques like operator splittings augmented lagrangians. shall provide overview methods currently developed recent results well computational studies providing comparison different methods also illustrating success applications.",10
"deep double sparsity encoder: learning sparsify features also parameters. paper emphasizes significance jointly exploit problem structure parameter structure, context deep modeling. specific interesting example, describe deep double sparsity encoder (ddse), inspired double sparsity model dictionary learning. ddse simultaneously sparsities output features learned model parameters, one unified framework. addition intuitive model interpretation, ddse also possesses compact model size low complexity. extensive simulations compare ddse several carefully-designed baselines, verify consistently superior performance ddse. apply ddse novel application domain brain encoding, promising preliminary results achieved.",3
"mining complex hydrobiological data galois lattices. used galois lattices mining hydrobiological data. data macrophytes, macroscopic plants living water bodies. plants characterized several biological traits, several modalities. aim cluster plants according common traits modalities find relations traits. galois lattices efficient methods aim, apply binary data. article, detail approaches used transform complex hydrobiological data binary data compare first results obtained thanks galois lattices.",3
invertibility robustness phaseless reconstruction. paper concerned question reconstructing vector finite-dimensional real hilbert space magnitudes coefficients vector redundant linear map known. analyze various lipschitz bounds nonlinear analysis map establish theoretical performance bounds reconstruction algorithm. show robust stable reconstruction requires additional redundancy critical threshold.,10
"v2v-posenet: voxel-to-voxel prediction network accurate 3d hand human pose estimation single depth map. existing deep learning-based methods 3d hand human pose estimation single depth map based common framework takes 2d depth map directly regresses 3d coordinates keypoints, hand human body joints, via 2d convolutional neural networks (cnns). first weakness approach presence perspective distortion 2d depth map. depth map intrinsically 3d data, many previous methods treat depth maps 2d images distort shape actual object projection 3d 2d space. compels network perform perspective distortion-invariant estimation. second weakness conventional approach directly regressing 3d coordinates 2d image highly non-linear mapping, causes difficulty learning procedure. overcome weaknesses, firstly cast 3d hand human pose estimation problem single depth map voxel-to-voxel prediction uses 3d voxelized grid estimates per-voxel likelihood keypoint. design model 3d cnn provides accurate estimates running real-time. system outperforms previous methods almost publicly available 3d hand human pose estimation datasets placed first hands 2017 frame-based 3d hand pose estimation challenge.",3
"relativistic qft bohmian perspective: proof concept. since bohmian mechanics explicitly nonlocal, widely believed hard, impossible, make bohmian mechanics compatible relativistic quantum field theory (qft). explain, simple terms, hard construct bohmian theory lacks lorentz covariance, makes measurable predictions relativistic qft. one construct bohmian theory makes measurable predictions qft one lorentz frame, standard relativistic qft guarantees predictions lorentz invariant. first explain general terms, describe simple bohmian model makes measurable predictions standard model elementary particles, give hints towards fundamental theory beyond standard model. finally, present short story telling views fundamental physics general, bohmian mechanics particular, evolved time.",17
"guided cost learning: deep inverse optimal control via policy optimization. reinforcement learning acquire complex behaviors high-level specifications. however, defining cost function optimized effectively encodes correct task challenging practice. explore inverse optimal control (ioc) used learn behaviors demonstrations, applications torque control high-dimensional robotic systems. method addresses two key challenges inverse optimal control: first, need informative features effective regularization impose structure cost, second, difficulty learning cost function unknown dynamics high-dimensional continuous systems. address former challenge, present algorithm capable learning arbitrary nonlinear cost functions, neural networks, without meticulous feature engineering. address latter challenge, formulate efficient sample-based approximation maxent ioc. evaluate method series simulated tasks real-world robotic manipulation problems, demonstrating substantial improvement prior methods terms task complexity sample efficiency.",3
"infochemical core. vocalizations less often gestures object linguistic research decades. however, development general theory communication human language particular case requires clear understanding organization communication means. infochemicals chemical compounds carry information employed small organisms cannot emit acoustic signals optimal frequency achieve successful communication. distribution infochemicals across species investigated ranked degree number species associated (because produce sensitive it). quality fit different functions dependency degree rank evaluated penalty number parameters function. surprisingly, double zipf (a zipf distribution two regimes different exponent each) model yielding best fit although function largest number parameters. suggests world wide repertoire infochemicals contains chemical nucleus shared many species reminiscent core vocabularies found human language dictionaries large corpora.",15
"multiscale shrinkage lévy processes. new shrinkage-based construction developed compressible vector $\boldsymbol{x}\in\mathbb{r}^n$, cases components $\xv$ naturally associated tree structure. important examples $\xv$ corresponds coefficients wavelet block-dct representation data. method consider detail, numerical results presented, based increments gamma process. however, demonstrate general framework appropriate many types shrinkage priors, within l\'{e}vy process family, gamma process special case. bayesian inference carried approximating posterior samples mcmc algorithm, well constructing heuristic variational approximation posterior. also consider expectation-maximization (em) map (point) solution. state-of-the-art results manifested compressive sensing denoising applications, latter spiky (non-gaussian) noise.",18
"non-gaussian scale space filtering 2 2 matrix linear filters. construction scale space convolution filter studied extensively past. proven convolution kernel satisfies scale space requirements gaussian type. paper, consider matrix convolution filters introduced [1] building kernel scale space, shows construct non-gaussian scale space $2\times 2$ matrix filters. paper derives sufficient conditions matrix filters scale space kernel, present numerical demonstrations.",3
"classification log-polar-visual eigenfaces using multilayer perceptron. paper present simple novel approach tackle challenges scaling rotation face images face recognition. proposed approach registers training testing visual face images log-polar transformation, capable handle complicacies introduced scaling rotation. log-polar images projected eigenspace finally classified using improved multi-layer perceptron. experiments used orl face database object tracking classification beyond visible spectrum (otcbvs) database visual face images. experimental results show proposed approach significantly improves recognition performances visual log-polar-visual face images. case orl face database, recognition rate visual face images 89.5% increased 97.5% log-polar-visual face images whereas otcbvs face database recognition rate visual images 87.84% 96.36% log-polar-visual face images.",3
"context models sequences covers. present class models that, via simple construction, enables exact, incremental, non-parametric, polynomial-time, bayesian inference conditional measures. approach relies upon creating sequence covers conditioning variable maintaining different model set within cover. inference remains tractable specifying probabilistic model terms random walk within sequence covers. demonstrate approach problems conditional density estimation, which, knowledge first closed-form, non-parametric bayesian approach problem.",18
"deep transfer learning: new deep learning glitch classification method advanced ligo. exquisite sensitivity advanced ligo detectors enabled detection multiple gravitational wave signals. sophisticated design detectors mitigates effect types noise. however, advanced ligo data streams contaminated numerous artifacts known glitches: non-gaussian noise transients complex morphologies. given high rate occurrence, glitches lead false coincident detections, obscure even mimic gravitational wave signals. therefore, successfully characterizing removing glitches advanced ligo data utmost importance. here, present first application deep transfer learning glitch classification, showing knowledge deep learning algorithms trained real-world object recognition transferred classifying glitches time-series based spectrogram images. using gravity spy dataset, containing hand-labeled, multi-duration spectrograms obtained real ligo data, demonstrate method enables optimal use deep convolutional neural networks classification given small training datasets, significantly reduces time training networks, achieves state-of-the-art accuracy 98.8%, perfect precision-recall 8 22 classes. furthermore, new types glitches classified accurately given labeled examples technique. trained via transfer learning, show convolutional neural networks truncated used excellent feature extractors unsupervised clustering methods identify new classes based morphology, without labeled examples. therefore, provides new framework dynamic glitch classification gravitational wave detectors, expected encounter new types noise undergo gradual improvements attain design sensitivity.",6
"good mistakes? theoretical analysis cegis. counterexample-guided inductive synthesis cegis used synthesize programs candidate space programs. technique guaranteed terminate synthesize correct program space candidate programs finite. technique may may terminate correct program candidate space programs infinite. paper, perform theoretical analysis counterexample-guided inductive synthesis technique. investigate whether set candidate spaces correct program synthesized using cegis depends counterexamples used inductive synthesis, is, whether good mistakes would increase synthesis power. investigate whether use minimal counterexamples instead arbitrary counterexamples expands set candidate spaces programs inductive synthesis successfully synthesize correct program. consider two kinds counterexamples: minimal counterexamples history bounded counterexamples. history bounded counterexample used iteration cegis bounded examples used previous iterations inductive synthesis. examine relative change power inductive synthesis cases. show synthesis technique using minimal counterexamples mincegis synthesis power cegis synthesis technique using history bounded counterexamples hcegis different power cegis, none dominates other.",3
quantum image classification using principal component analysis. present novel quantum algorithm classification images. algorithm constructed using principal component analysis von neuman quantum measurements. order apply algorithm present new quantum representation grayscale images.,17
"extending fo(id) knowledge producing definitions: preliminary results. previous research relation asp classical logic identified least two different ways former extends latter. first, asp program typically contain sets rules naturally interpreted inductive definitions, language fo(id) shown inductive definitions elegantly added classical logic modular way. second, course also well-known epistemic component asp, mainly emphasized early papers stable model semantics. investigate whether kind knowledge also, similarly modular way, added classical logic, language ordered epistemic logic presented recent work. however, logic views epistemic component entirely separate inductive definition component, thus ignoring possible interplay two. paper, present language extends inductive definition construct found fo(id) epistemic component, making interplay possible. eventual goal work discover whether really appropriate view epistemic component inductive definition component asp two separate extensions classical logic, whether also something importance combination two.",3
"note lack symmetry graphical lasso. graphical lasso (glasso) widely-used fast algorithm estimating sparse inverse covariance matrices. glasso solves l1 penalized maximum likelihood problem available r library cran. output glasso, regularized covariance matrix estimate sparse inverse covariance matrix estimate, identify graphical model also serve intermediate inputs multivariate procedures pca, lda, manova, others. glasso indeed produces covariance matrix estimate solves l1 penalized optimization problem dual sense; however, method producing inverse covariance matrix estimator optimization inexact may produce asymmetric estimates. problem exacerbated amount l1 regularization applied small, turn likely occur true underlying inverse covariance matrix sparse. lack symmetry potentially consequences. first, implies covariance inverse covariance estimates numerical inverses one another, second, asymmetry possibly lead negative complex eigenvalues,rendering many multivariate procedures may depend inverse covariance estimator unusable. demonstrate problem, explain causes, propose possible remedies.",18
generating segment durations text-to-speech system: hybrid rule-based/neural network approach. combination neural network rule firing information rule-based system used generate segment durations text-to-speech system. system shows slight improvement performance neural network system without rule firing information. synthesized speech using segment durations accepted listeners quality speech generated using segment durations extracted natural speech.,3
"alternative em gaussian mixture models: batch stochastic riemannian optimization. consider maximum likelihood estimation gaussian mixture models (gmms). task almost invariably solved (in theory practice) via expectation maximization (em) algorithm. em owes success various factors, ability fulfill positive definiteness constraints closed form key importance. propose alternative em appealing rich riemannian geometry positive definite matrices, using cast gmm parameter estimation riemannian optimization problem. surprisingly, out-of-the-box riemannian formulation completely fails proves much inferior em. motivates us take closer look problem geometry, derive better formulation much amenable riemannian optimization. develop (riemannian) batch stochastic gradient algorithms outperform em, often substantially. provide non-asymptotic convergence analysis stochastic method, also first (to knowledge) global analysis riemannian stochastic gradient. numerous empirical results included demonstrate effectiveness methods.",18
"long-range fractal correlations literary corpora. paper analyse fractal structure long human-language records mapping large samples texts onto time series. particular mapping set work inspired linguistic basis sense retains {\em word} fundamental unit communication. results confirm beyond short-range correlations resulting syntactic rules acting sentence level, long-range structures emerge large written language samples give rise long-range correlations use words.",2
"joint estimation sparse multivariate regression conditional graphical models. multivariate regression model natural generalization classical univari- ate regression model fitting multiple responses. paper, propose high- dimensional multivariate conditional regression model constructing sparse estimates multivariate regression coefficient matrix accounts dependency struc- ture among multiple responses. proposed method decomposes multivariate regression problem series penalized conditional log-likelihood response conditioned covariates responses. allows simultaneous estimation sparse regression coefficient matrix sparse inverse covariance matrix. asymptotic selection consistency normality established diverging dimension covariates number responses. effectiveness pro- posed method also demonstrated variety simulated examples well application glioblastoma multiforme cancer data.",18
"generative modeling approach limited channel ecg classification. processing temporal sequences central variety applications health care, particular multi-channel electrocardiogram (ecg) highly prevalent diagnostic modality relies robust sequence modeling. recurrent neural networks (rnns) led significant advances automated diagnosis time-series data, perform poorly models trained using limited set channels. crucial limitation existing solutions rely solely discriminative models, tend generalize poorly scenarios. order combat limitation, develop generative modeling approach limited channel ecg classification. approach first uses seq2seq model implicitly generate missing channel information, uses latent representation perform actual supervisory task. decoupling enables use unsupervised data also provides highly robust metric spaces subsequent discriminative learning. experiments physionet dataset clearly evidence effectiveness approach standard rnns disease prediction.",18
"satellite galaxies groups cielo project i. gas removal galaxies re-distribution intragroup medium. study impact environment galaxies fall orbit potential well local group (lg) analogue, following high cadence. analysis performed eight disc satellite galaxies cielo suite hydrodynamical simulations. galaxies stellar masses within range $[10^{8.1} - 10^{9.56}] m_{\sun} $h$^{-1}$. measure tidal torques, ram pressure specific star formation rates (ssfr) function time, correlate amount gas lost satellites along orbits. stronger removal episodes occur disc plane oriented perpendicular direction motion. one peripassage required significantly modify orientations discs respect orbital plane. gas removed interaction central galaxies may also found opposite direction motion, depending orbital configuration. satellites totally quenched galaxies reach first peripassage, continue forming $10\%$ final stellar mass event. fraction removed gas found product joint action tidal torque ram pressure, also trigger new star formation activity subsequent supernova feedback.",0
"convergence nucleation rate first-order phase transitions. paper investigates importance radiative corrections first-order phase transitions, particular focus bubble-nucleation rate. calculations done strict power-counting, observables consistently calculated every order. ensures physical quantities gauge renormalization-scale invariant. furthermore, avoid large logarithms high-temperatures, effective three-dimensional theory used. effective theory automatically incorporates higher-order thermal masses. results paper indicate sub-leading corrections rate large. partly radiative corrections enhanced large bubbles. illustrate calculations, three models considered: real-scalar model, radiative-barrier model, model effective dimension $6$ operator. relevant observables calculated model, reliability perturbation theory discussed.",8
"parameterized complexity results plan reuse. planning notoriously difficult computational problem high worst-case complexity. researchers investing significant efforts develop heuristics restrictions make planning practically feasible. case-based planning heuristic approach one tries reuse previous experience solving similar problems order avoid planning effort. plan reuse may offer interesting alternative plan generation settings. provide theoretical results identify situations plan reuse provably tractable. perform analysis framework parameterized complexity, supports rigorous worst-case complexity analysis takes structural properties input account terms parameters. central notion parameterized complexity fixed-parameter tractability extends classical notion polynomial-time tractability utilizing effect structural properties problem input. draw detailed map parameterized complexity landscape several variants problems arise context case-based planning. particular, consider problem reusing existing plan, imposing various restrictions terms parameters, number steps added existing plan turn solution planning instance hand.",3
"pragmatic qft measurement problem need heisenberg-like cut qft. despite quantum theory's remarkable success predicting (statistical) results experiments, many philosophers worry nonetheless lacks crucial connection theory experiment. worries root quantum measurement problem. identify two kinds worries: 1) pragmatic: unclear model experiments extract theoretical predictions, 2) realist: realist narrative experiment underlying theoretical predictions. worries deserve attention, pragmatic worries far worse consequences left unanswered. moreover, argue, upon reflection, satisfactory explanation almost quantum theory's experimental successes unavoidably involves modeling quantum fields point. thus, without pragmatic theory-to-experiment link qft, risk losing right claim evidential support large parts quantum theory. hence, focus pragmatic qft measurement problem. but, makes modeling measurements qft hard? discuss, attempts naively transplant non-relativistic quantum measurement theory qft deeply unphysical unsatisfying. thus need new (or least refined) measurement theory qft. however, argue, aiming directly new measurement theory incautious way proceed apt lead us astray. paper proposes alternate way forward: ought first better understand non-relativistic quantum measurement theory rooted notions measurement chains heisenberg cuts. ought generalize notions transplant qft. transplant carried paper. analysis suggests need pragmatic qft-cut analogous need pragmatic heisenberg cut non-relativistic contexts.",14
"minimax-optimal semi-supervised regression unknown manifolds. consider semi-supervised regression predictor variables drawn unknown manifold. simple two step approach problem to: (i) estimate manifold geodesic distance pair points using labeled unlabeled instances; (ii) apply k nearest neighbor regressor based distance estimates. prove given sufficiently many unlabeled points, simple method geodesic knn regression achieves optimal finite-sample minimax bound mean squared error, manifold known. furthermore, show approach efficiently implemented, requiring o(k n log n) operations estimate regression function n labeled unlabeled points. illustrate approach two datasets manifold structure: indoor localization using wifi fingerprints facial pose estimation. cases, geodesic knn accurate much faster popular laplacian eigenvector regressor.",18
"path integral control reproducing kernel hilbert space embedding. present embedding stochastic optimal control problems, called path integral form, reproducing kernel hilbert spaces. using consistent, sample based estimates embedding leads model free, non-parametric approach calculation approximate solution control problem. formulation admits decomposition problem invariant task dependent component. consequently, make much efficient use sample data compared previous sample based approaches domain, e.g., allowing sample re-use across tasks. numerical examples test problems, illustrate sample efficiency, provided.",3
"ethics robotics. three laws robotics first appeared together isaac asimov's story 'runaround' mentioned form previous works asimov. three laws commonly known three laws robotics earliest forms depiction needs ethics robotics. simplistic language isaac asimov able explain rules robot must confine order maintain societal sanctity. however, even though outdated still represent innate fears beginning resurface present day 21st century. society advent new revolution; revolution led advances computer science, artificial intelligence & nanotechnology. advances phenomenal surpassed predicted moore's law. advancements comes fear future may mercy androids. humans today scared we, ourselves, might create something cannot control. may end creating something learn much faster anyone us can, also evolve faster theory evolution allowed us to. greatest fear might lose jobs intelligent beings, beings might end replacing us top cycle. public hysteria heightened number cultural works depict annihilation human race robots. right frankenstein i, robot mass media also depicted issues. paper effort understand need ethics robotics simply termed roboethics. achieved study artificial beings thought put behind them. end paper, however, concluded need ethical robots ever need ethical roboticists.",3
"users prefer guetzli jpeg same-sized libjpeg. report pairwise comparisons human raters jpeg images libjpeg new guetzli encoder. although files size-matched, 75% ratings favor guetzli. implies butteraugli psychovisual image similarity metric guides guetzli reasonably close human perception high quality levels. provide access raw ratings source images analysis study.",3
"credible persuasion. propose new notion credibility bayesian persuasion problems. disclosure policy credible sender cannot profit tampering messages keeping message distribution unchanged. show credibility disclosure policy equivalent cyclical monotonicity condition induced distribution states actions. also characterize credibility restricts sender's ability persuade different payoff structures. particular, sender's payoff state-independent, disclosure policies credible. apply results market lemons, show useful information credibly disclosed seller, even though seller commit disclosure policy would perfectly reveal private information maximize profit.",4
"line: large-scale information network embedding. paper studies problem embedding large information networks low-dimensional vector spaces, useful many tasks visualization, node classification, link prediction. existing graph embedding methods scale real world information networks usually contain millions nodes. paper, propose novel network embedding method called ""line,"" suitable arbitrary types information networks: undirected, directed, and/or weighted. method optimizes carefully designed objective function preserves local global network structures. edge-sampling algorithm proposed addresses limitation classical stochastic gradient descent improves effectiveness efficiency inference. empirical experiments prove effectiveness line variety real-world information networks, including language networks, social networks, citation networks. algorithm efficient, able learn embedding network millions vertices billions edges hours typical single machine. source code line available online.",3
"cumulative culture spontaneously emerges artificial navigators social memory-guided. previously thought uniquely human, cumulative cultural evolution continues found non-human animals. occurs adaptive innovation individual repeatedly passed onto consecutive generations social learning. example, pigeons fly alone stable pairs show relatively rigid sub-optimal routes, gradually improve route efficiency generations pairs experienced members swapped naive ones. raises question minimally required cognitive architecture cumulative cultural evolution emerge. here, aimed answer question artificial agents employ three main functions: goal-direction, social proximity, route memory. optima efficiency generational efficiency improvement, agents replicated cumulative culture observed pigeons. optimum, paths determined primarily memory, lesser extent social proximity goal-direction. need social proximity, naive agent stayed close experienced counterpart followed memorised path. however, unhindered route memory, naive agent's heading likely err towards goal. subtly biased pairs' routes, resulting efficiency improvement thus regression goal. resulting incremental improvements generations meet core criteria current frameworks cumulative cultural evolution, suggesting rudimentary cumulative optimisation evolutionary mechanism emerges even simple systems prefer social proximity memory capacity.",15
"protvec: continuous distributed representation biological sequences. introduce new representation feature extraction method biological sequences. named bio-vectors (biovec) refer biological sequences general protein-vectors (protvec) proteins (amino-acid sequences) gene-vectors (genevec) gene sequences, representation widely used applications deep learning proteomics genomics. present paper, focus protein-vectors utilized wide array bioinformatics investigations family classification, protein visualization, structure prediction, disordered protein identification, protein-protein interaction prediction. method, adopt artificial neural network approaches represent protein sequence single dense n-dimensional vector. evaluate method, apply classification 324,018 protein sequences obtained swiss-prot belonging 7,027 protein families, average family classification accuracy 93%+-0.06% obtained, outperforming existing family classification methods. addition, use protvec representation predict disordered proteins structured proteins. two databases disordered sequences used: disprot database well database featuring disordered regions nucleoporins rich phenylalanine-glycine repeats (fg-nups). using support vector machine classifiers, fg-nup sequences distinguished structured protein sequences found protein data bank (pdb) 99.8% accuracy, unstructured disprot sequences differentiated structured disprot sequences 100.0% accuracy. results indicate providing sequence data various proteins model, accurate information protein structure determined.",15
"target-driven visual navigation indoor scenes using deep reinforcement learning. two less addressed issues deep reinforcement learning (1) lack generalization capability new target goals, (2) data inefficiency i.e., model requires several (and often costly) episodes trial error converge, makes impractical applied real-world scenarios. paper, address two issues apply model task target-driven visual navigation. address first issue, propose actor-critic model whose policy function goal well current state, allows better generalize. address second issue, propose ai2-thor framework, provides environment high-quality 3d scenes physics engine. framework enables agents take actions interact objects. hence, collect huge number training samples efficiently. show proposed method (1) converges faster state-of-the-art deep reinforcement learning methods, (2) generalizes across targets across scenes, (3) generalizes real robot scenario small amount fine-tuning (although model trained simulation), (4) end-to-end trainable need feature engineering, feature matching frames 3d reconstruction environment. supplementary video accessed following link: https://youtu.be/smbxmdiorvs.",3
"nefi: network extraction images. networks network-like structures amongst central building blocks many technological biological systems. given mathematical graph representation network, methods graph theory enable precise investigation properties. software analysis graphs widely available applied graphs describing large scale networks social networks, protein-interaction networks, etc. applications, graph acquisition, i.e., extraction mathematical graph network, relatively simple. however, many network-like structures, e.g. leaf venations, slime molds mud cracks, data collection relies images graph extraction requires domain-specific solutions even manual. introduce network extraction images, nefi, software tool automatically extracts accurate graphs images wide range networks originating various domains. previous work graph extraction images, theoretical results fully accessible expert audience ready-to-use implementations non-experts rarely available insufficiently documented. nefi provides novel platform allowing practitioners many disciplines easily extract graph representations images supplying flexible tools image processing, computer vision graph theory bundled convenient package. thus, nefi constitutes scalable alternative tedious error-prone manual graph extraction special purpose tools. anticipate nefi enable collection larger datasets reducing time spent graph extraction. analysis new datasets may open possibility gain new insights structure function various types networks. nefi open source available http://nefi.mpi-inf.mpg.de.",3
"existence epipolar matrices. paper considers foundational question existence fundamental (resp. essential) matrix given $m$ point correspondences two views. present complete answer existence fundamental matrices value $m$. using examples disprove widely held beliefs fundamental matrices always exist whenever $m \leq 7$. time, prove exist unconditionally $m \leq 5$. mild genericity condition, show essential matrix always exists $m \leq 4$. also characterize six seven point configurations two views matrices satisfying epipolar constraint rank one.",3
"geometric dirichlet means algorithm topic inference. propose geometric algorithm topic learning inference built convex geometry topics arising latent dirichlet allocation (lda) model nonparametric extensions. end study optimization geometric loss function, surrogate lda's likelihood. method involves fast optimization based weighted clustering procedure augmented geometric corrections, overcomes computational statistical inefficiencies encountered techniques based gibbs sampling variational inference, achieving accuracy comparable gibbs sampler. topic estimates produced method shown statistically consistent conditions. algorithm evaluated extensive experiments simulated real data.",18
"exploring significance using perceptually relevant image decolorization method scene classification. color image contains luminance chrominance components representing intensity color information respectively. objective work presented paper show significance incorporating chrominance information task scene classification. improved color-to-grayscale image conversion algorithm effectively incorporating chrominance information proposed using color-to-gay structure similarity index (c2g-ssim) singular value decomposition (svd) improve perceptual quality converted grayscale images. experimental result analysis based image quality assessment image decolorization called c2g-ssim success rate (cadik color250 datasets) shows proposed image decolorization technique performs better 8 existing benchmark algorithms image decolorization. second part paper, effectiveness incorporating chrominance component scene classification task demonstrated using deep belief network (dbn) based image classification system developed using dense scale invariant feature transform (sift) features. levels chrominance information incorporated proposed image decolorization technique confirmed improvement overall scene classification accuracy . also, overall scene classification performance improved combination models obtained using proposed conventional decolorization methods.",3
"automata networks model alignment least effort vocabulary formation. artificial communities agents develop language scaling relations close zipf law? preliminary answer question, propose automata networks model formation vocabulary population individuals, two principle opposite strategies: alignment least effort principle. within previous account emergence linguistic conventions (specially, naming game), focus modeling speaker hearer efforts actions vocabularies study impact actions formation shared language. numerical simulations essentially based energy function, measures amount local agreement vocabularies. results suggests one dimensional lattices best strategy formation shared languages one minimizes efforts speakers communicative tasks.",3
"driver action prediction using deep (bidirectional) recurrent neural network. advanced driver assistance systems (adas) significantly improved effective driver action prediction (dap). predicting driver actions early accurately help mitigate effects potentially unsafe driving behaviors avoid possible accidents. paper, formulate driver action prediction timeseries anomaly prediction problem. anomaly (driver actions interest) detection might trivial context, finding patterns consistently precede anomaly requires searching extracting features across multi-modal sensory inputs. present driver action prediction system, including real-time data acquisition, processing learning framework predicting future impending driver action. proposed system incorporates camera-based knowledge driving environment driver themselves, addition traditional vehicle dynamics. uses deep bidirectional recurrent neural network (dbrnn) learn correlation sensory inputs impending driver behavior achieving accurate high horizon action prediction. proposed system performs better existing systems driver action prediction tasks accurately predict key driver actions including acceleration, braking, lane change turning durations 5sec action executed driver.",18
"using moeas outperform stock benchmarks presence typical investment constraints. portfolio managers typically constrained turnover limits, minimum maximum stock positions, cardinality, target market capitalization sometimes need hew style (such growth value). addition, portfolio managers often use multifactor stock models choose stocks based upon respective fundamental data. use multiobjective evolutionary algorithms (moeas) satisfy real-world constraints. portfolios generated consistently outperform typical performance benchmarks statistically significant asset selection.",16
"dewitt wave function hořava-lifshitz cosmology tensor perturbation. present well-tempered dewitt wave function, vanishes classical big-bang singularity, ho\v{r}ava-lifshitz (hl) cosmology tensor perturbation, analytically numerically. general relativity, dewitt wave function ill-behaved tensor perturbation taken account. essentially amplitude perturbation diverges singularity perturbative expansion completely breaks down. hand, hl gravity known higher dimensional operators required perturbative renormalizability render tensor perturbation scale-invariant regular way singularity. paper analytically show $d+1$ dimensional hl gravity dewitt wave function tensor perturbation indeed well-defined around classical big-bang singularity. also, numerically demonstrate well-behaved dewitt wave function tensor perturbation big-bang finite size universe.",6
"good good enough? re-evaluating bar energy disaggregation. since early 1980s, research community developed ever sophisticated algorithms problem energy disaggregation, despite decades research, still dearth applications demonstrated value. work, explore question highly pertinent research community: good energy disaggregation need order infer characteristics household? present novel techniques use unsupervised energy disaggregation predict household occupancy static properties household size home number occupants. results show basic disaggregation approaches performs 30% better occupancy estimation using aggregate power data alone, 10% better estimating static household characteristics. results show even rudimentary energy disaggregation techniques sufficient improved inference household characteristics. conclude, re-evaluate bar set community energy disaggregation accuracy try answer question ""how good good enough?""",3
"comparison word embeddings english cross-lingual chinese word sense disambiguation. word embeddings ubiquitous forms word representation natural language processing. applications word embeddings monolingual word sense disambiguation (wsd) english, comparisons done. paper attempts bridge gap examining popular embeddings task monolingual english wsd. simplified method leads comparable state-of-the-art performance without expensive retraining. cross-lingual wsd - word senses word source language e come separate target translation language f - also assist language learning; example, providing translations target vocabulary learners. thus also applied word embeddings novel task cross-lingual wsd chinese provide public dataset benchmarking. also experimented using word embeddings lstm networks found surprisingly basic lstm network work well. discuss ramifications outcome.",3
"convolutional approach reflection symmetry. present convolutional approach reflection symmetry detection 2d. model, built products complex-valued wavelet convolutions, simplifies previous edge-based pairwise methods. parameter-centered, opposed feature-centered, certain computational advantages object sizes known priori, demonstrated ellipse detection application. method outperforms best-performing algorithm cvpr 2013 symmetry detection competition database single-symmetry case. code new database 2d symmetry detection available.",3
"well generative adversarial networks learn densities: nonparametric view. study paper rate convergence learning densities generative adversarial networks (gan) framework, borrowing insights nonparametric statistics. introduce improved gan estimator achieves faster rate, simultaneously leveraging level smoothness target density evaluation metric, theory remedies mode collapse problem reported literature. minimax lower bound constructed show dimension large, exponent rate new gan estimator near optimal. one view results answering quantitative way well gan learns wide range densities different smoothness properties, hierarchy evaluation metrics. byproduct, also obtain improved generalization bounds gan deeper relu discriminator network.",18
"lexicographic choice functions. investigate generalisation coherent choice functions considered seidenfeld et al. (2010), sticking convexity axiom imposing archimedeanity condition. define choice functions vector spaces options, allows us incorporate special cases seidenfeld et al.'s (2010) choice functions horse lotteries sets desirable gambles (quaeghebeur, 2014), investigate connections. show choice functions based sets desirable options (gambles) satisfy seidenfeld's convexity axiom particular types sets desirable options, one-to-one relationship lexicographic probabilities. call lexicographic choice functions. finally, prove choice functions used determine conservative convex choice function associated given binary relation.",3
"shortcut-stacked sentence encoders multi-domain inference. present simple sequential sentence encoder multi-domain natural language inference. encoder based stacked bidirectional lstm-rnns shortcut connections fine-tuning word embeddings. overall supervised model uses encoder encode two input sentences two vectors, uses classifier vector combination label relationship two sentences entailment, contradiction, neural. shortcut-stacked sentence encoders achieve strong improvements existing encoders matched mismatched multi-domain natural language inference (top non-ensemble single-model result emnlp repeval 2017 shared task (nangia et al., 2017)). moreover, achieve new state-of-the-art encoding result original snli dataset (bowman et al., 2015).",3
"introduction conditional random fields. often wish predict large number variables depend well observed variables. structured prediction methods essentially combination classification graphical modeling, combining ability graphical models compactly model multivariate data ability classification methods perform prediction using large sets input features. tutorial describes conditional random fields, popular probabilistic method structured prediction. crfs seen wide application natural language processing, computer vision, bioinformatics. describe methods inference parameter estimation crfs, including practical issues implementing large scale crfs. assume previous knowledge graphical modeling, tutorial intended useful practitioners wide variety fields.",18
"learning hybrid architecture sequence regression annotation. learning hidden markov model (hmm), sequen- tial observations often complemented real-valued summary response variables generated path hid- den states. settings arise numerous domains, includ- ing many applications biology, like motif discovery genome annotation. paper, present flexible frame- work jointly modeling latent sequence features functional mapping relates summary response variables hidden state sequence. algorithm com- patible rich set mapping functions. results show availability additional continuous response vari- ables simultaneously improve annotation se- quential observations yield good prediction performance synthetic data real-world datasets.",18
"integrated tools engineering ontologies. article presents overview current specialized ontology engineering tools, well texts' annotation tools based ontologies. main functions features tools, advantages disadvantages discussed. systematic comparative analysis means engineering ontologies presented.",3
"multi-spectral image panchromatic sharpening, outcome process quality assessment protocol. multispectral (ms) image panchromatic (pan) sharpening algorithms proposed remote sensing community ever increasing number variety. aim sharpen coarse spatial resolution ms image fine spatial resolution pan image acquired simultaneously spaceborne airborne earth observation (eo) optical imaging sensor pair. unfortunately, date, standard evaluation procedure ms image pan sharpening outcome process community agreed upon, contrast quality assurance framework earth observation (qa4eo) guidelines proposed intergovernmental group earth observations (geo). general, process easier measure, outcome important. original contribution present study fourfold. first, existing procedures quantitative quality assessment (q2a) (sole) pan sharpened ms product critically reviewed. conceptual implementation drawbacks highlighted overcome quality improvement. second, novel (to best authors' knowledge, first) protocol q2a ms image pan sharpening product process designed, implemented validated independent means. third, within protocol, innovative categorization spectral spatial image quality indicators metrics presented. fourth, according new taxonomy, original third order isotropic multi scale gray level co occurrence matrix (tims glcm) calculator tims glcm texture feature extractor proposed replace popular second order glcms.",3
"constant regret, generalized mixability, mirror descent. consider setting prediction expert advice; learner makes predictions aggregating group experts. setting, right choice loss function ""mixing"" algorithm, possible learner achieve constant regret regardless number prediction rounds. example, constant regret achieved \emph{mixable} losses using \emph{aggregating algorithm} (aa). \emph{generalized aggregating algorithm} (gaa) name family algorithms parameterized convex functions simplices (entropies), reduce aa using \emph{shannon entropy}. given entropy $\phi$, losses constant regret possible using gaa called $\phi$-mixable. losses $\phi$-mixable previously left open question. fully characterize $\phi$-mixability, answer open questions posed \cite{reid2015}. also elaborate tight link gaa \emph{mirror descent algorithm} minimizes weighted loss experts.",3
"development validation novel dementia alzheimer's type (dat) score based metabolism fdg-pet imaging. fluorodeoxyglucose positron emission tomography (fdg-pet) imaging based 3d topographic brain glucose metabolism patterns normal controls (nc) individuals dementia alzheimer's type (dat) used train novel multi-scale ensemble classification model. ensemble model outputs fdg-pet dat score (fpds) 0 1 denoting probability subject clinically diagnosed dat based metabolism profile. novel 7 group image stratification scheme devised groups images based associated clinical diagnosis also past future trajectories clinical diagnoses, yielding continuous representation different stages dat spectrum mimics real-world clinical setting. potential using fpds dat biomarker validated large number fdg-pet images (n=2984) obtained alzheimer's disease neuroimaging initiative (adni) database taken across proposed stratification, good classification auc (area curve) 0.78 achieved distinguishing images belonging subjects dat trajectory images taken subjects progressing dat diagnosis. further, fpds biomarker achieved state-of-the-art performance mild cognitive impairment (mci) dat conversion prediction task auc 0.81, 0.80, 0.77 2, 3, 5 years conversion windows respectively.",3
"constraint answer set solver ezcsp integration schemas matter. researchers answer set programming constraint programming spent significant efforts development hybrid languages solving algorithms combining strengths traditionally separate fields. efforts resulted new research area: constraint answer set programming. constraint answer set programming languages systems proved successful providing declarative, yet efficient solutions problems involving hybrid reasoning tasks. one main contributions paper first comprehensive account constraint answer set language solver ezcsp, mainstream representative research area used various successful applications. also develop extension transition systems proposed nieuwenhuis et al. 2006 capture boolean satisfiability solvers. use extension describe ezcsp algorithm prove formal claims it. design algorithmic details behind ezcsp clearly demonstrate development hybrid systems kind challenging. many questions arise one faces various design choices attempt maximize system's benefits. one key decisions developer hybrid solver makes settling particular integration schema within implementation. thus, another important contribution paper thorough case study based ezcsp, focused various integration schemas provides. consideration theory practice logic programming (tplp).",3
"towards large-scale ultrahigh dimensional feature selection via feature generation. many real-world applications text mining, desirable select relevant features variables improve generalization ability, provide better interpretation prediction models. {in paper, novel adaptive feature scaling (afs) scheme proposed introducing feature scaling {vector $\d \in [0, 1]^m$} alleviate bias problem brought scaling bias diverse features.} reformulating resultant afs model semi-infinite programming problem, novel feature generating method presented identify relevant features classification problems. contrast traditional feature selection methods, new formulation advantage solving extremely high-dimensional large-scale problems. exact solution worst-case analysis identification relevant features, proposed feature generating scheme converges globally. importantly, proposed scheme facilitates group selection without special structures. comprehensive experiments wide range synthetic real-world datasets demonstrate proposed method {achieves} better competitive performance compared existing methods (group) feature selection terms generalization performance training efficiency. c++ matlab implementations algorithm available \emph{http://c2inet.sce.ntu.edu.sg/mingkui/robust-fgm.rar}.",3
"capsulegan: generative adversarial capsule network. present generative adversarial capsule network (capsulegan), framework uses capsule networks (capsnets) instead standard convolutional neural networks (cnns) discriminators within generative adversarial network (gan) setting, modeling image data. provide guidelines designing capsnet discriminators updated gan objective function, incorporates capsnet margin loss, training capsulegan models. show capsulegan outperforms convolutional-gan modeling image data distribution mnist cifar-10 datasets, evaluated generative adversarial metric semi-supervised image classification.",18
"fixed-rank matrix factorizations riemannian low-rank optimization. motivated problem learning linear regression model whose parameter large fixed-rank non-symmetric matrix, consider optimization smooth cost function defined set fixed-rank matrices. adopt geometric framework optimization riemannian quotient manifolds. study underlying geometries several well-known fixed-rank matrix factorizations exploit riemannian quotient geometry search space design class gradient descent trust-region algorithms. proposed algorithms generalize previous results fixed-rank symmetric positive semidefinite matrices, apply broad range applications, scale high-dimensional problems confer geometric basis recent contributions learning fixed-rank non-symmetric matrices. make connections existing algorithms context low-rank matrix completion discuss relative usefulness proposed framework. numerical experiments suggest proposed algorithms compete state-of-the-art manifold optimization offers effective versatile framework design machine learning algorithms learn fixed-rank matrix.",3
"characterization monotone influence measures data classification. work focus following question: important i-th feature determining outcome given datapoint? identify family influence measures; functions that, given datapoint x, assign value phi_i(x) every feature i, roughly corresponds i's importance determining outcome x. family uniquely derived set axioms: desirable properties reasonable influence measure satisfy. departing prior work influence measures, assume knowledge - access - underlying classifier labelling dataset. words, influence measures based dataset alone, make queries classifier. requirement naturally limits scope explanations provide, show effective real datasets.",3
"midgar: detection people computer vision internet things scenarios improve security smart cities, smart towns, smart homes. could use computer vision internet things using pictures sensors? principal hypothesis want resolve. currently, order create safety areas, cities, homes, people use ip cameras. nevertheless, system needs people watch camera images, watch recording something occurred, watch camera notifies movement. disadvantages. furthermore, many smart cities smart homes around world. thought using idea internet things add way automating use ip cameras. case, propose analysis pictures computer vision detect people analysed pictures. analysis, able obtain pictures contain people handle pictures sensors two possible states. notwithstanding, computer vision complicated field. needed second hypothesis: could work computer vision internet things good accuracy automate semi-automate kind events? demonstration hypotheses required testing computer vision module check possibilities use module possible real environment good accuracy. proposal, possible solution, analysis entire sequence instead isolated pictures using pictures sensors internet things.",3
"general multiblock method structured variable selection. regularised canonical correlation analysis recently extended two sets variables multiblock method regularised generalised canonical correlation analysis (rgcca). further, sparse gcca (sgcca) proposed address issue variable selection. however, technical reasons, variable selection offered sgcca restricted covariance link blocks (i.e., $\tau=1$). one main contributions paper go beyond covariance link propose extension sgcca full rgcca model (i.e., $\tau\in[0, 1]$). addition, propose extension sgcca exploits structural relationships variables within blocks. specifically, propose algorithm allows structured sparsity-inducing penalties included rgcca optimisation problem. proposed multiblock method illustrated real three-block high-grade glioma data set, aim predict location brain tumours, simulated data set, aim illustrate method's ability reconstruct true underlying weight vectors.",18
"exploring text virality social networks. paper aims shed light concept virality - especially social networks - provide new insights structure. argue that: (a) virality phenomenon strictly connected nature content spread, rather influencers spread it, (b) virality phenomenon many facets, i.e. generic term several different effects persuasive communication comprised partially overlap. give ground claims, provide initial experiments machine learning framework show various aspects virality independently predicted according content features.",3
"deep generative models relational data side information. present probabilistic framework overlapping community discovery link prediction relational data, given graph. proposed framework has: (1) deep architecture enables us infer multiple layers latent features/communities node, providing superior link prediction performance complex networks better interpretability latent features; (2) regression model allows directly conditioning node latent features side information available form node attributes. framework handles (1) (2) via clean, unified model, enjoys full local conjugacy via data augmentation, facilitates efficient inference via closed form gibbs sampling. moreover, inference cost scales number edges attractive massive sparse networks. framework also easily extendable model weighted networks count-valued edges. compare various state-of-the-art methods report results, quantitative qualitative, several benchmark data sets.",18
"look same? deciphering chinese, japanese koreans fine-grained deep learning. study extend chinese, japanese korean faces classified facial attributes offer important cues. first, propose novel way obtaining large numbers facial images nationality labels. train state-of-the-art neural networks labeled images. able achieve accuracy 75.03% classification task, chances 33.33% human accuracy 38.89% . further, train multiple facial attribute classifiers identify distinctive features group. find chinese, japanese koreans exhibit substantial differences certain attributes, bangs, smiling, bushy eyebrows. along way, uncover several gender-related cross-country patterns well. work, complements existing apis microsoft cognitive services face++, could find potential applications tourism, e-commerce, social media marketing, criminal justice even counter-terrorism.",3
"deep learning real time crime forecasting. accurate real time crime prediction fundamental issue public safety, remains challenging problem scientific community. crime occurrences depend many complex factors. compared many predictable events, crime sparse. different spatio-temporal scales, crime distributions display dramatically different patterns. distributions low regularity space time. work, adapt state-of-the-art deep learning spatio-temporal predictor, st-resnet [zhang et al, aaai, 2017], collectively predict crime distribution los angeles area. models two staged. first, preprocess raw crime data. includes regularization space time enhance predictable signals. second, adapt hierarchical structures residual convolutional units train multi-factor crime prediction models. experiments half year period los angeles reveal highly accurate predictive power models.",10
"data analysis bayesian networks: bootstrap approach. recent years significant progress algorithms methods inducing bayesian networks data. however, complex data analysis problems, need go beyond satisfied inducing networks high scores. need provide confidence measures features networks: existence edge two nodes warranted? markov blanket given node robust? say something ordering variables? able address questions, even amount data enough induce high scoring network. paper propose efron's bootstrap computationally efficient approach answering questions. addition, propose use confidence measures induce better structures data, detect presence latent variables.",3
"large scale distributed acoustic modeling back-off n-grams. paper revives older approach acoustic modeling borrows n-gram language modeling attempt scale amount training data model size (as measured number parameters model), approximately 100 times larger current sizes used automatic speech recognition. data-rich setting, expand phonetic context significantly beyond triphones, well increase number gaussian mixture components context-dependent states allow it. experimented contexts span seven context-independent phones, 620 mixture components per state. dealing unseen phonetic contexts accomplished using familiar back-off technique used language modeling due implementation simplicity. back-off acoustic model estimated, stored served using mapreduce distributed computing infrastructure. speech recognition experiments carried n-best list rescoring framework google voice search. training big models large amounts data proves effective way increase accuracy state-of-the-art automatic speech recognition system. use 87,000 hours training data (speech along transcription) obtained filtering utterances voice search logs automatic speech recognition confidence. models ranging size 20--40 million gaussians estimated using maximum likelihood training. achieve relative reductions word-error-rate 11% 6% combined first-pass models trained using maximum likelihood, boosted maximum mutual information, respectively. increasing context size beyond five phones (quinphones) help.",3
"local procrustes manifold embedding: measure embedding quality embedding algorithms. present procrustes measure, novel measure based procrustes rotation enables quantitative comparison output manifold-based embedding algorithms (such lle (roweis saul, 2000) isomap (tenenbaum et al, 2000)). measure also serves natural tool choosing dimension-reduction parameters. also present two novel dimension-reduction techniques attempt minimize suggested measure, compare results techniques results existing algorithms. finally, suggest simple iterative method used improve output existing algorithms.",18
"intelligent parameter tuning optimization-based iterative ct reconstruction via deep reinforcement learning. number image-processing problems formulated optimization problems. objective function typically contains several terms specifically designed different purposes. parameters front terms used control relative weights among them. critical importance tune parameters, quality solution depends values. tuning parameter relatively straightforward task human, one intelligently determine direction parameter adjustment based solution quality. yet manual parameter tuning tedious many cases, becomes impractical number parameters exist problem. aiming solving problem, paper proposes approach employs deep reinforcement learning train system automatically adjust parameters human-like manner. demonstrate idea example problem optimization-based iterative ct reconstruction pixel-wise total-variation regularization term. set parameter tuning policy network (ptpn), maps ct image patch output specifies direction amplitude parameter patch center adjusted. train ptpn via end-to-end reinforcement learning procedure. demonstrate guidance trained ptpn parameter tuning pixel, reconstructed ct images attain quality similar better reconstructed manually tuned parameters.",14
"slow learners fast. online learning algorithms impressive convergence properties comes risk minimization convex games large problems. however, inherently sequential design prevents taking advantage modern multi-core architectures. paper prove online learning delayed updates converges well, thereby facilitating parallel online learning.",10
"dictionary-based tensor canonical polyadic decomposition. ensure interpretability extracted sources tensor decomposition, introduce paper dictionary-based tensor canonical polyadic decomposition enforces one factor belong exactly known dictionary. new formulation sparse coding proposed enables high dimensional tensors dictionary-based canonical polyadic decomposition. benefits using dictionary tensor decomposition models explored terms parameter identifiability estimation accuracy. performances proposed algorithms evaluated decomposition simulated data unmixing hyperspectral images.",18
"monomial reduction knot polynomials. natural numbers $n$ prime numbers $p$, find knot $k$ whose skein polynomial $p_k(a,z)$ evaluated $z=n$ trivial reduction modulo $p$. interesting consequence construction polynomials $p_k(a,n)$ (mod~$p$) bounded $a$-span realised knots bounded braid index. application, classify polynomials form $p_k(a,1)$ (mod $2$) $a$-span $\leq 10$.",10
"discrete evolutionary model chess players' ratings. elo system rating chess players, also used games sports, adopted world chess federation four decades ago. although without controversy, accepted generally reliable provides method assessing players' strengths ranking official tournaments. generally accepted distribution players' rating data approximately normal but, date, stochastic model distribution might arisen proposed. propose evolutionary stochastic model, models arrival players rating pool, games play other, results games affect ratings. using continuous approximation discrete model, derive distribution players' ratings time $t$ normal distribution, variance increases time logarithmic function $t$. validate model using published rating data 2007 2010, showing parameters obtained data recovered simulations stochastic model. distribution players' ratings approximately normal shown small negative skew. show modify evolutionary stochastic model take skewness account, validate modified model using published official rating data.",14