figure_id stringlengths 16 17 | paper_id stringlengths 10 10 | paper_link stringlengths 32 32 | caption stringlengths 0 1.8k | text_mentions stringlengths 233 2.55k |
|---|---|---|---|---|
2202.06374_FIG_1 | 2202.06374 | https://arxiv.org/abs/2202.06374 | Dependence of optimal holdout set size on parameters of estimated learning curve ($a,b,c$, with $k_2(n,a,b,c)=a n^{-b}+c$), cost in intervention set $k_1$, and total number of samples $N$. Figures show change in optimal holdout set size $n_*$ while varying one parameter and holding others constant at $(a,b,c)=\left(\fr... | See Supplementary Figure~\ref{supp_fig:partials_nstar} for typical forms of $\nabla_{\Theta} n_\star$. In light of the proportionality assumption in Section~\ref{sec:practicalities}, and the tendency of the accuracy of a risk score with number of training samples (`learning curve') to follow a power-law form \citep{vi... |
2202.06374_FIG_2 | 2202.06374 | https://arxiv.org/abs/2202.06374 | Dependence of minimum total cost on parameters of estimated learning curve ($a,b,c$, with $k_2(n,a,b,c)=a n^{-b}+c$), cost in intervention set $k_1$, and total number of samples $N$. Figures show change in minimal cost $\ell(n_*)$ while varying one parameter and holding others constant at $(a,b,c)=\left(\frac{3}{2},\fr... | In light of the proportionality assumption in Section~\ref{sec:practicalities}, and the tendency of the accuracy of a risk score with number of training samples (`learning curve') to follow a power-law form \citep{viering21}, we recommend considering such a parametric form for $k_2$ (i.e.~$k_2(n; \theta)=a n^{-b} + c$ ... |
2203.16073_FIG_3 | 2203.16073 | https://arxiv.org/abs/2203.16073 | X-MOP: the guidelines for explainable AI purposes in POP | The insights obtained with the use of the research questions and event log analysis are summarized in an XAI-based guidelines map for POP in Figure~\ref{fig:guidelines}. This figure describes the guidelines for Explainable AI purposes in POP. The white boxes are the questions to guide partitioners and researchers in ob... |
2204.09583_FIG_1 | 2204.09583 | https://arxiv.org/abs/2204.09583 | The effect of using different epochs for the feature extractor (phase 1) on classifier retraining's (phase 2) test accuracies. | We present an ablation study of how different early stopping epoch (Figure \ref{fig:epoch-ablation}), average validation accuracy (Figure \ref{fig:accs-ablation} left), and worst-group accuracy (Figure \ref{fig:accs-ablation} right) of the initial ERM trained model affect the group DRO clsasifier retraining phase of C... |
2205.02328_FIG_1 | 2205.02328 | https://arxiv.org/abs/2205.02328 | \textbf{IPD:} The top graph shows the normalized average population reward of MARL experiments with three different cost:benefit ratios when $N=30$ with 95\% confidence intervals. The bottom graph shows incentivized actions by Equation~\ref{eq:final_requrement}, where positive (or zero) is cooperation and negative is d... | The incentives for each team structure in our IPD environment can be visualized in the bottom graph of Figure~\ref{fig:teams_adv}. To capture the behavior of agents after they have converged to a policy, the top graph of Figure \ref{fig:teams_adv} shows the normalized average global reward of the last 25\% of the episo... |
2205.02328_FIG_2 | 2205.02328 | https://arxiv.org/abs/2205.02328 | \textbf{IPD:} The 5/6 team composition showing the percent of cooperation towards teammates and non-teammates when $c = 1$ and $b \in \{ 2, 5 \}$. When benefit is greater, agents develop pro-social policies towards non-teammates despite the incentive to defect. | Figure~\ref{fig:teams_adv_overtime} shows the percent of cooperative actions over time with the 5/6 team structure when $b \in \{2, 5\}$. % Figure~\ref{fig:teams_adv_overtime} shows agent's actions in the 5/6 team structure with two values of benefit, $b \in \{2, 5\}$. Both graphs in Figure~\ref{fig:teams_adv_overtime}... |
2205.02328_FIG_3 | 2205.02328 | https://arxiv.org/abs/2205.02328 | \textbf{Cleanup: }Mean population reward for each team structure with 95\% confidence intervals. 6/1 represents individualistic agents and 1/6 represents when all agents have common interest. Both 2/3 and 3/2 team structures achieve more reward than 1/6 and 6/1. | Figure~\ref{fig:cleanup_rewards} shows the mean population reward for each scenario in Cleanup with 95\% confidence intervals. % This allows the 2 pickers to gather a higher number of apples each than any agent in other scenarios and a good collective reward in Figure~\ref{fig:cleanup_rewards}. The 3/2 team structure t... |
2205.02328_FIG_4 | 2205.02328 | https://arxiv.org/abs/2205.02328 | \textbf{Cleanup: }Inverse Gini index (equality) for each team structure with 95\% confidence intervals. Higher values represent more equality. Both 2/3 and 3/2 team structures have high equality despite the interests of all agents not being aligned. | % However, $T_{1}$ (red) in the 3/2 scenario ends up free-riding on the labor of other teams to support the four apple picker joint policy which leads to less equality in Figure~\ref{fig:cleanup_equality}. Figure~\ref{fig:cleanup_equality} shows our results for equality where higher values represent more equality. |
2205.02328_FIG_5 | 2205.02328 | https://arxiv.org/abs/2205.02328 | \textbf{Cleanup: }Mean number of apples (top) and cleaning beams (bottom) per-episode for each of 6 agents in different team structures. | To achieve the four picker/two cleaner division of labor, one team ($T_{1}$ (red) in Figure~\ref{fig:div_labor}) must free-ride on the labor of the other two teams. For 3/2 to achieve this division of labor, one team ($T_{1}$ (red) in Figure~\ref{fig:div_labor}) must free-ride on the labor of the other two teams. There... |
2208.01455_FIG_2 | 2208.01455 | https://arxiv.org/abs/2208.01455 | TBD. Link to figure: https://drive.google.com/file/d/17vquhVnh0h1q9cpawihRoSM6LcxxYXgR/view?usp=sharing | In this work, our goal is to provide a new minimally invasive surgery environment for endovascular procedures. We aim at facilitating the development of endovascular autonomy through the provision of a standardized environment that confers familiarity to the machine learning community. As such, we propose \textbf{CathS... |
2208.01455_FIG_4 | 2208.01455 | https://arxiv.org/abs/2208.01455 | The forces distributions are visualized through kernel density estimation (left) as well as through the cumulative distribution function (right). The figure depicts the similarities between the two distributions, where the empirical distribution gathered from the simulator shows a heavier tail than the normal distribut... | We sample the same number of samples generated in our simulation and the real experiment in~\cite{kundrat2021mr} for a fair comparison. We use the distributions to generate a cumulative distribution function $F_X(x)=P(X \leq x)$, which can be visualized in Fig.~\ref{fig:force_distributions}. We further assess the norma... |
2208.01455_FIG_5 | 2208.01455 | https://arxiv.org/abs/2208.01455 | The max forces generated by the catheter on the aorta within each of the evaluation episodes during the PPO training are shown. The left figure depicts the results gathered within the Type-I Aortic Arch where the catheter has greater interaction with the aorta, whereas the one on the right shows the results gathered wi... | % We further display the results of the force interaction in Fig.~\ref{fig:boxplot}. The internal observation depicts the greatest amount of force and the greatest interquartile range in all cases except within the Type~II aortic arch when performing the navigation towards the LCCA target. Additionally, it can be visua... |
2208.09418_FIG_4 | 2208.09418 | https://arxiv.org/abs/2208.09418 | GA is applied to test seeds (norm balls) from MNIST and CIFAR10 dataset to find worst case interpretation discrepancy, measure by PCC. First row: fixed population size 1000, and varied iterations, Second row: fixed iterations, and varied population size. ``Gradient$\times$Input" interpretation method is considered. | Additional experiments on hyper-parameter settings in GA and SS are presented in Fig.~\ref{fig:ga_accuracy_additional} and Fig.~\ref{fig:ss_para}. The objective function interpretation discrepancy $\mathfrak{D}$, measured by PCC, is optimised to converge with the increasing number of iterations while the prediction los... |
2203.03673_FIG_1 | 2203.03673 | https://arxiv.org/abs/2203.03673 | Assessing trained graph generators. | \item Regardless of the learning or training procedures (masked in grey), \AgraSSt{} can test a given generator $G$ that is only accessible through its generated samples as shown in Figure.\ref{fig:florentine_sim}. In this setting, we do not need to know how the generator $G$ is obtained and the focus is the assess... |
2204.10453_FIG_2 | 2204.10453 | https://arxiv.org/abs/2204.10453 | \small{Parameter estimate trajectories for the rank-deficient regressor ``Shoot the Gap" example.} | When $\Delta = \Delta_D$, however, both the BLA and LSM methods that succeed under a full-rank regressor fail; each requires the inversion of a non-invertible matrix. As illustrated by Figure \ref{fig.simple_rd}, the proposed controller succeeds in safely reaching the goal by learning a parameter vector, $\hat{\boldsym... |
2205.09787_FIG_1 | 2205.09787 | https://arxiv.org/abs/2205.09787 | Reconstruction Accuracy and MSE when changing the proportion of edges over nodes (see \S\ref{app:edges}). The amount of known edges injected is 10\%, 20\% (as in the paper, see Fig.\ref{fig:alphaplot_noise}) and 50\% (see \S\ref{app:known edges}). | In Fig.~\ref{fig:edges_50percentknown} we vary the proportion of edges injected (10\%, 20\% as in Fig.\ref{fig:alphaplot_noise}, and 50\%). The effect of changing the proportion of edges per node ($e = |E|/|V|$) is presented in Fig.~\ref{fig:edges_50percentknown} (where CASTLE+ vs Injected 20\% is the scenario sho... |
2207.13467_FIG_1 | 2207.13467 | https://arxiv.org/abs/2207.13467 | Sensitivity analysis showing the percentage of TTS improvement compared to the no-control case for a domain of $K_r$ and $C_r$. | The results are reported in Fig.~\ref{fig:RMsensitivity}, where one may observe that the ranges of $K_r$ and $C_r$ that produce positive effects in terms of TTS improvement (i.e., the orange area) is very wide, that is, the estimator is not very sensitive to such parameters choice while we remain within these ranges. |
2208.05181_FIG_3 | 2208.05181 | https://arxiv.org/abs/2208.05181 | Actual numbers of binary variables.\label{fig:val} | In Fig.~\ref{fig:val}, we plot the actual numbers of binary variables required by the conventional QUBO, the proposed HUBO, and the QUBO transformed from HUBO (quadratization), where we fixed $\NCH = \lfloor \NAP / 2 \rfloor$ for illustration. As shown in Fig.~\ref{fig:val}, the number of binary variables for the propo... |
2208.05181_FIG_4 | 2208.05181 | https://arxiv.org/abs/2208.05181 | Expected query complexity in the quantum domain.\label{fig:query} | In Fig.~\ref{fig:query}, we show expected query complexities of both formulations, where the number of channels was $\NCH = \lfloor \NAP / 2 \rfloor$. As shown in Fig.~\ref{fig:query}, the classic exhaustive search exhibited impractical complexities for both conventional and proposed formulations, and is infeasible on ... |
2209.01684_FIG_1 | 2209.01684 | https://arxiv.org/abs/2209.01684 | Attacker's re-identification accuracy (RID-ACC) on the Adult dataset for top-$k$ re-identification using the FK-RI model and the RS+FD[GRR] protocol and by varying the uniform $\epsilon$-LDP privacy metric and the number of surveys. | \noindent\textbf{Results.} Fig.~\ref{fig:reident_rspfd} illustrates the attacker's RID-ACC metric on the Adult dataset for top-$k$ re-identification using the FK-RI model and the RS+FD[GRR] protocol and by varying the uniform $\epsilon$-LDP privacy metric and the number of surveys. \noindent\textbf{Analysis.} From Fig... |
2205.05178_FIG_2 | 2205.05178 | https://arxiv.org/abs/2205.05178 | \label{fig:coverConvergence} $L^{-1} \log \operatorname{Mag} (B_{v_0}(L),t) \rightarrow h(D)$ for $t > 0$, but depends strongly on $t$ even for fairly large $L$. | Continuing Example \ref{ex:cover}, $h(D) \approx 0.2812$ is the logarithm of the so-called \emph{plastic number}, i.e., the unique real solution of $x^3-x-1 = 0$. Numerics suggest that $|V(B_1(L))|$ is given by \cite{oeisA167385}. Assuming this to obtain values for large $L$, we show the convergence of $L^{-1} \log \op... |
2205.05178_FIG_4 | 2205.05178 | https://arxiv.org/abs/2205.05178 | Distribution of correlations for various centralities between two random subgraphs of the digraph in Figure \ref{fig:flare}. * indicates a ball in the digraph with all edges reversed. As $L$ increases, boundary effects cause the log-magnitudes of balls in the universal cover to become (slightly) more correlated to each... | We then computed the (co)weightings at scale 0, the log-magnitudes of balls of radius $\le 3$ at scale $t = 100$ (which is virtually equivalent to $t = \infty$), and various common vertex centrality measures. For each of these quantities and $N$ realizations, we then computed the correlation coefficients on vertices sh... |
2205.05178_FIG_5 | 2205.05178 | https://arxiv.org/abs/2205.05178 | Distribution of correlations for various centralities between two random subgraphs of distinct Erd\H{o}s-Reny\'i digraphs with $n = 100$ vertices and edge probability $q = 4/n$, cf. Figure \ref{fig:centralitiesFlare}. | The strong correlations of log-magnitudes of balls are more robust than those of (co)weightings, as an experiment along the same lines as above but using different realizations of an Erd\H{o}s-Reny\'i digraph ($n = 100$ vertices; edge probability $q = 4/n$) as the ambient digraph for each of $N = 100$ trials shows. We ... |
2208.11702_FIG_1 | 2208.11702 | https://arxiv.org/abs/2208.11702 | A pipeline of a three-stage procedure to generate and validate synthetic data of skin lesions. | The main aims of our study are to explore and evaluate generative modelling techniques for generating synthetic medical data and to inspect its impact on modelling performance. As personal integrity and regulatory issues are central when using personal health data, we chose to test the generation of artificial data by ... |
2209.01769_FIG_1 | 2209.01769 | https://arxiv.org/abs/2209.01769 | \textcolor{black}{Illustration of the analysis transform ($A_1$) with the frame-type adaptation (FA) module. The symbol $'$ indicates that the activation function (i.e. GDN) is omitted. Conv(N, K, S) stands for a convolutional layer with the number of output channels N, kernel size K, and stride S.} | We tackle this issue by introducing a frame-type adaptation (FA) module to every convolutional layer, aiming to share the convolutional layers between B-frames while adapting their features to the B-frame type. Specifically, according to the B-frame type (namely, reference B-frames, non-reference B-frames, or B*-frames... |
2209.01769_FIG_2 | 2209.01769 | https://arxiv.org/abs/2209.01769 | Profiles of the per-frame PSNR's and the per-frame bit rates averaged over GOPs on UVG dataset, with and without frame-type adaptive (FA) coding \textcolor{black}{at high rates ($\lambda_1=2048$). Frame 0, 32 are coded as I-frame, Frame 16 is coded as B*-frame, and the odd-numbered frames are non-reference B-frames whi... | Fig.~\ref{fig:psnr_bitrate_gop_distribution} \textcolor{black}{further} visualizes how \textcolor{black}{the proposed FA module} may impact the bit allocation among B-frames in a GOP and their decoded quality. \textcolor{black}{In this and the following experiments, the distortion weighting is enabled during training. ... |
2209.01769_FIG_3 | 2209.01769 | https://arxiv.org/abs/2209.01769 | \textcolor{black}{Comparison of PSNR-RGB and bit allocation by turning on and off frame-type adaptive coding in the motion and/or inter-frame codecs at high rates ($\lambda_1=2048$). The PSNR-RGB and bit-allocation results are averaged over B-frames of the test sequences in UVG dataset. FA refers to frame-type adaptive... | \textcolor{black}{To gain insights into this observation, Fig.~\ref{fig:fa_psnr_rate} analyzes the bit allocation between motion coding and inter-frame coding for B-frames. Comparing setting C with setting D, we observe that enabling frame-type adaptive coding for the inter-frame codec only is able to reduce effectivel... |
2209.01769_FIG_4 | 2209.01769 | https://arxiv.org/abs/2209.01769 | \textcolor{black}{Rate-distortion comparison of GOP sizes 4, 8, 16 on UVG dataset under intra-period 32.} | We evaluate the effect of GOP size on the performance of B-CANF. A number of GOP sizes, including 4, 8, 16, are tested with intra-period 32. The BD-rates are summarized in Table~\ref{tab:abl_gop} (see the results w/o a separate P-frame codec). The corresponding rate-distortion curves on UVG dataset are presented in Fig... |
2207.01352_FIG_1 | 2207.01352 | https://arxiv.org/abs/2207.01352 | Articles identified with excess attention around the date of the Paris attack in 2015. The color encodes the total views of a given article on a given day (also written within each cell). Articles directly related to the attack are shown in boldface font. | We commence by a thorough examination of the Paris attack in 2015, which is the initial event within our sample. We will broaden our analysis to encompass all attacks later. Focusing on the 12 days surrounding the Paris event, we identify 17 articles that received the most attention in this period, shown in Fig.~\ref{f... |
2207.01352_FIG_5 | 2207.01352 | https://arxiv.org/abs/2207.01352 | Dynamics of decay of collective public attention after terrorist attacks. Three top row panels: the cumulative time series of viewcounts for Paris 2015, Brussels 2016 and London Bridge 2017, for the broad category Enemy. The time of each attack is marked by a vertical line and the period after it is shaded in grey. Sec... | After an attack, how gradually is the public attention restored to routine concerns? Does this depend on the attack and/or the broad category? We look at three examples: Paris 2015, Brussels 2016, and London Bridge 2017. For each of them, we construct a time series that shows the decay of public attention to each broad... |
2207.01352_FIG_7 | 2207.01352 | https://arxiv.org/abs/2207.01352 | The distribution of maximal daily viewcounts (maximal number of views reached on a single day) between 1 January 2015 and 8 August 2019, for all 5,876,878 articles after Step 1. We kept only the articles in the shaded area, i.e., excluded those that were never visited more than 300 times on a single day. Note a slight ... | \paragraph{Step 2: Exclusion of articles with low viewership.} To filter our low view articles, we used a cut-off based on the maximal daily viewcount. We kept only the articles whose daily viewcount at least once reached 300. See Fig.~\ref{fig_maximal_daily} for the distribution of maximum daily views and our cut-off ... |
2207.01352_FIG_8 | 2207.01352 | https://arxiv.org/abs/2207.01352 | Viewcounts of three articles around the time of the November 2015 Paris attack. Attention on the article in red peaks immediately after the attack, and given its content, we reasonably expect that this was a reaction to the attack (true positive). The article in grey exhibits no qualitative change, and indeed, given it... | \paragraph{Step 3: Removal of articles with small Excess Attention after attacks.} Next, for each attack separately, we exclude articles that received no marked increase in attention immediately after the incident (true negatives). As threshold we use $Z=3$: Assuming that \vs are normally distributed, the chance of obs... |
2207.01352_FIG_9 | 2207.01352 | https://arxiv.org/abs/2207.01352 | Examples of four articles identified in Step 5. Top left: \vi time series of \textit{Terrorist incidents linked to ISIL} from 1 March 2017 to 1 September 2017. Attacks that occurred during those 6 months are indicated. Top right: the histogram of $Z$-scores for that article for the entire time span (region $Z>3$ is sha... | See Fig.~\ref{fig-four-example-time-series} for illustration. This step left us with 86 articles. In the same context, recall that in defining Excess Attention, we implicitly assumed that daily viewcounts (in fact, their weekly averages) are normally distributed around the mean, which may not always be the case. But ... |
2207.01352_FIG_10 | 2207.01352 | https://arxiv.org/abs/2207.01352 | The time series of daily viewcounts to all articles in English Wikipedia excluding only stubs and redirects. Top panel: the time series with dates of the attacks shown. Bottom panel: the same time series for 3 months in 2017 during which five attacks occurred (dates shown on $x$-axis are all Mondays). | Do terrorist attacks divert attention from other simultaneously ongoing concerns? Here, we examine the time series of daily viewcounts to all articles cumulatively, excluding only stubs and redirects (see section \ref{Exclusion}, immediately after the exclusion Step 1). The obtained time series is shown in Fig.~\ref{fi... |
2201.04581_FIG_2 | 2201.04581 | https://arxiv.org/abs/2201.04581 | The distribution of the duration of datasets. | Statistics of the duration, shown in Figure \ref{fig:dataset_duration}, reveal several interesting characteristics in quantifying above mentioned datasets. Some lengths of audio in Coswara and COUGHVID samples are less than 1s, which might lead to being unqualified for the training model and errors in the reading input... |
2201.04581_FIG_3 | 2201.04581 | https://arxiv.org/abs/2201.04581 | The number of subjects for each task defined. | Based on the metadata as shown in Table \ref{table:metadata}, the total of 1,310 subjects are separated into: COVID-19 negative and COVID-19 positive subjects, subjects with and without symptoms, and normal and anomaly subjects for task I, II, and III respectively as shown in Figure~\ref{fig:covid_num}. |
2205.09923_FIG_1 | 2205.09923 | https://arxiv.org/abs/2205.09923 | Remote state estimation over unknown channels | %Throughout the first part of this work, we will focus on a remote state estimation scenario as depicted in Fig. \ref{fig:system_model}. Section~\ref{sec:control} will then outline how the results translate to networked systems with closed loop control. where $w_k \sim \mathcal{N}(0,Q)$, $v_k \sim \mathcal{N}(0,R)$, se... |
2205.09923_FIG_3 | 2205.09923 | https://arxiv.org/abs/2205.09923 | Estimation regret over time | In Fig. \ref{fig:regret_comparison_plot} we plot the estimation regret over time for the $\varepsilon$-greedy and Thompson sampling algorithms. Plots for OBS and SBS are qualitatively similar to Thompson sampling and are omitted. The parameters used are the same as those mentioned at the start of Section \ref{sec:numer... |
2207.03890_FIG_4 | 2207.03890 | https://arxiv.org/abs/2207.03890 | High-level overview of our data processing pipeline. The pipeline shows how the encoding is created for the given NetFlow input data using our encoding algorithm. | Figure~\ref{fig:data_processing_pipeline} provides a high-level overview of our data processing pipeline. For the clustering of the vectors, we have used the KMeans Clustering implementation that is provided by Scikit-learn machine-learning package~\cite{sklearn_kmeans}. We opted to use KMeans clustering to cluster the... |
2207.03890_FIG_5 | 2207.03890 | https://arxiv.org/abs/2207.03890 | Architecture of the Kubernetes Cluster that was used to generate the ANONYMISED dataset | To evaluate our approach, we use three different datasets to learn models for anomaly detection. For the first dataset, we created a new dataset for our project~\cite{assureMOSS_project}. This dataset (AssureMOSS) contains both benign and malicious NetFlows gathered from a Kubernetes cluster running multiple applicatio... |
2202.05687_FIG_2 | 2202.05687 | https://arxiv.org/abs/2202.05687 | Probability of finding at least $k$ orthogonal adversarial directions for $\epsilon=12/255$ for increasing ensemble sizes. Probabilities expectedly decrease as ensemble size increases. These results are for ensemble configuration $S4$. | %For configuration $S4$, Figure~\ref{fig: gaas}, plots the probability of finding at least $k$ maximal orthogonal adversarial directions, as ensemble size increases. We observe that probability of finding these adversarial directions indeed decreases as ensemble size, i.e., number of orthogonal constraints increases. W... |
2202.05687_FIG_3 | 2202.05687 | https://arxiv.org/abs/2202.05687 | Probability distribution of the adversarial subspace dimension for three settings based on fraction of frequency features. | % Figure~\ref{fig:gaas_plot} plots the proportion of images (out of $1000$) with at least $k$ adversarial dimensions. We observe that both the $50\%$ and $25\%$ settings have fewer images with high dimension of the adversarial subspace. These results support the analysis done in Section~\ref{sec: gaas}. Further, the re... |
2202.05687_FIG_4 | 2202.05687 | https://arxiv.org/abs/2202.05687 | Distribution of adversarial saliency magnitudes for frequency components. | %Recall that each model in our ensembles uses a subset of frequencies. However, as per~\cite{ilyas2019adversarial}'s robust features model, these frequencies can likely be categorized as either robust, or non-robust. We would like each model to receive an equal number of robust frequencies. In Section~\ref{sec:redundan... |
2203.08550_FIG_2 | 2203.08550 | https://arxiv.org/abs/2203.08550 | The principle drawing of the light paths from the top, front, back, left and right directions. | The one-camera-multiple-mirror imaging device consists of 1 camera and 7 mirrors placed around the observed water tank. Fig. \ref{IROSFigMirrorDrawing} describes its imaging principle. The image information of the soft hand will be reflected by the mirrors and captured by one camera. As shown in Fig. \ref{IROSFigMirror... |
2203.08550_FIG_6 | 2203.08550 | https://arxiv.org/abs/2203.08550 | The three key angles of the reconstructed result based on the bidirectional grasping video. (a) The bending angle of soft fingers in the main grasping axis. (b) The deflection angle of soft fingers in the side-deflection axis. (c) The bending angle of soft palm actuators. (P1: Palm actuator 1, P2: Palm actuator 2, P3: ... | Using the 3D reconstructed algorithm, the experimental video about bidirectional grasping was calculated frame by frame. We skipped the image in the frames when the researcher is operating the bottle to be grasped, or the bottle is greatly shelter the detection features. To eliminate the data fluctuation, the data in ... |
2203.08550_FIG_7 | 2203.08550 | https://arxiv.org/abs/2203.08550 | The finger contribution comparison of the bending angle for human and robot hands. (a) Single-direction grasp, (b) Bidirectional grasp. | Based on the collected data of the fingers of the soft robot and human hands, the proportion of the bending angle were calculated to analyze the contribution of each fingers. As shown in Figure \ref{IROSFigBiomimeticRatioComparsion}, for the single-direction grasp, the thumb, index and middle fingers act as the main ro... |
2203.11028_FIG_2 | 2203.11028 | https://arxiv.org/abs/2203.11028 | INT4-packing for performing four parallel multiplications on a single DSP. Here, \$ denotes the (extended) sign bits and $\delta$ denotes the padding bits. The offsets of the individual inputs determine the offsets of the individual results as described in Eqn.~\eqref{eq:int4_packing}. | The operand $a_0$ is mapped to the B-Port (see Fig.~\ref{fig:DSP48E2}) with an offset of 0. $a_1$ is also mapped to the B-Port but with an offset of 11. This is a hardware-efficient way of implementing $a_{1} \cdot 2^{11} + a_{0} \cdot 2^{0}$. Input $w_{0}$ is mapped to the preadder port A with an offset of zero. Since... |
2203.12437_FIG_1 | 2203.12437 | https://arxiv.org/abs/2203.12437 | Convolution with address events. A 2D input fmap is stored as a queue of address events (the AEQ). To perform convolution, the AEQ is processed sequentially. Two address events allow convolution to be performed in two successive steps, whereas sliding window based convolution in this example would require $4 \times 4 =... | The goal is to maximize the utilization of PEs while minimizing the number of PEs. To achieve this, the individual fmaps are stored in a compressed format. This compressed format is a queue of all spikes in a fmap whereby the spikes are not represented by a logical 1 but by their Address Event Representation (AER). The... |
2203.12437_FIG_4 | 2203.12437 | https://arxiv.org/abs/2203.12437 | Interlacing scheme applied to the AEQ. The spikes represented as address events $(i,j)$ and are sorted into the respective column $s \in (0,...,8)$, displayed in red. Note that the binary fmap containing the spikes is only shown as a easy to understand visualization of the AEQ's contents. | The AEQ has to store the address events in queues. This is done in an interlacing fashion (see Fig. \ref{fig:AEQ_interlaced}). The AEQ is not only a data structure but also features two independent circuits: one for writing and one for reading the queue columns. The thresholding unit writes the address events and the ... |
2203.12437_FIG_5 | 2203.12437 | https://arxiv.org/abs/2203.12437 | Pipelined architecture of the convolution unit. The convolution unit is setup in four stages. The core is operating on 9 membrane potentials (green) and 9 kernel weights (orange) in parallel. In S1, the read addresses of all 9 membrane potentials is calculated. These read addresses get passed along the pipeline and bec... | are executing their respective task in parallel (see Fig. \ref{fig:Pipelined_ConvCore}). This ensures that all parts Each of the 9 PEs is essentially an adder, performing the updates of the membrane potentials is connected to one of the 9 memory columns of MemPot. Each PE receives on of the nine kernel elements $K[0]$ ... |
2203.14717_FIG_2 | 2203.14717 | https://arxiv.org/abs/2203.14717 | Fuzzy Neural Network architecture. | In the proposed method a fuzzy neural network is presented to determine the assignment degree of each core based on its current state. Figure \ref{fig:fnn} shows the architecture of this fuzzy neural network. This fuzzy neural network consists of five layers: 1- input layer, 2- fuzzification layer, 3- fuzzy rules layer... |
2203.14717_FIG_3 | 2203.14717 | https://arxiv.org/abs/2203.14717 | Showing the uniformly partitioning of input variables. The domain of each input variable x is partitioned to 5 fuzzy sets: Very Low level (VL), Low level (L), Medium level (M), High level (H), and Very High level (VH). The input variable is normalized to range [0,1] by dividing its value to its maximum value. | The proposed structure is a control unit that should cover the whole input space properly. It is necessary to have one rule for each combination of different values of different variables. Therefore, it is not expected to determine fuzzy sets and fuzzy rules based on data-driven clustering approaches. Moreover, in the ... |
2203.14717_FIG_5 | 2203.14717 | https://arxiv.org/abs/2203.14717 | The extracted four-dimensional Pareto front. Three dimensions including temperature ($\theta$), power ($P$), and GSFR ($\Lambda$) are shown on axes and the fourth dimension, the execution time, is encoded in the colors. | \item \textbf{Joint Optimization Ability:} We have trained the model based on NSGA-II on a training dataset consisting of several synthetic and real-life application graphs with various sizes. It includes ten application graphs with sizes in range of 6 to 85 nodes. In order to show the joint optimization capability of ... |
2203.14717_FIG_6 | 2203.14717 | https://arxiv.org/abs/2203.14717 | The most frequently fired fuzzy rules extracted from our proposed learning procedure. Columns are the input/output variables and the rows indicate the fired rules' indexes. The colors from blue to red indicates the lower values to the higher ones | \item \textbf{Extracted Fuzzy Rules:} We have selected one solution from those extracted as the Pareto front based on the process explained in section \ref{sec:learning} and Algorithm \ref{alg:learning}. According to the uniform partitioning of four input variables to five fuzzy sets, we have 625 extracted rules. Conse... |
2203.14717_FIG_7 | 2203.14717 | https://arxiv.org/abs/2203.14717 | Statistical comparison of the proposed method with an FIS based on expert knowledge \cite{Ekhtiyari2019} according to different criteria. | The second comparison is performed to a fuzzy-based offline scheduling approach, that utilized a static fuzzy inference system to jointly optimize utilization, power consumption and temperature of heterogeneous MPSoCs~\cite{Ekhtiyari2019}. In this method, the degree of processors during task scheduling and mapping are ... |
2205.06900_FIG_2 | 2205.06900 | https://arxiv.org/abs/2205.06900 | Average execution time for MM-BD versus the number of initializations, averaged over backdoor attack ensembles chessboard-S, 1-pixel-S, and BadNet-M. | % In Fig. \ref{fig:num_init}, we show the effect of the number of random initializations on the performance of MM-BD. In particular, we focus on the detection accuracy and execution time as the number of initializations is varied. Here, we apply MM-BD with $\{1, 10, 20, 30\}$ initializations to backdoor attack ensemble... |
2205.09515_FIG_2 | 2205.09515 | https://arxiv.org/abs/2205.09515 | ADVI and MCMC posterior confidence bands for the true B-spline curve $\mu = f_{\bfbeta}(x)$ simulated in the first replica. Only one curve is shown for the point estimates of $\mu$ for ease of visualization, since both methods produced virtually indistinguishable curves. | The prior hyperparameters specified for Bayesian inference under both MCMC and Variational Bayes (VB) approach via ADVI were fixed as $a_{\alpha} = a_{\lambda} = a_{\phi} = b_{\alpha} = b_{\lambda} = b_{\phi} = 1,$ therefore representing vague prior knowledge about $\alpha, \ \lambda$ and $\phi$. Figure \ref{fig:conf_... |
2205.09515_FIG_3 | 2205.09515 | https://arxiv.org/abs/2205.09515 | Posterior estimates for the true B-spline curve $\mu = f_{\bfbeta}(x)$ simulated in the first replica obtained by ADVI and Bayesian P-splines (MCMC). The classical point estimate produced by the R function smooth.spline is also shown. | Next, Figure \ref{fig:comaprison_vb_pslines_smooth_splines} compares the ADVI with bridge penalization on the spline basis coefficients as in equation \eqref{eq:bridge_classical}, to its analogous Bayesian p-splines formulation by \cite{lang2004} (MCMC) and to the off-the-shelf smooth.spline R function. The objective i... |
2205.09515_FIG_5 | 2205.09515 | https://arxiv.org/abs/2205.09515 | Marginal posterior distributions for $\phi$, $\lambda$ and $\alpha$ based on the first replicate of the simulated data. Red dashed curves represent VB and black curves represent MCMC marginal posterior approximations. | Figure \ref{fig:marginals_vb_mcmc} shows that both methods produce very similar posterior marginal distributions for the spline coefficients and that the marginals capture the true values of the coefficients. Figure \ref{fig:lambda_alpha_phi_vb_mcmc} shows MCMC and ADVI posterior marginals for $\lambda$, $\phi$ and $\a... |
2205.09515_FIG_6 | 2205.09515 | https://arxiv.org/abs/2205.09515 | ADVI and MCMC bivariate joint posteriors for the first 10 B-spline coefficients. | Figure \ref{fig:marginals_vb_mcmc} shows that both methods produce very similar posterior marginal distributions for the spline coefficients and that the marginals capture the true values of the coefficients. Figure \ref{fig:lambda_alpha_phi_vb_mcmc} shows MCMC and ADVI posterior marginals for $\lambda$, $\phi$ and $\a... |
2205.09515_FIG_7 | 2205.09515 | https://arxiv.org/abs/2205.09515 | ADVI and MCMC bivariate joint posteriors for $\beta_1, \beta_2, \phi, \lambda, \alpha$. | Figure \ref{fig:marginals_vb_mcmc} shows that both methods produce very similar posterior marginal distributions for the spline coefficients and that the marginals capture the true values of the coefficients. Figure \ref{fig:lambda_alpha_phi_vb_mcmc} shows MCMC and ADVI posterior marginals for $\lambda$, $\phi$ and $\a... |
2205.09515_FIG_8 | 2205.09515 | https://arxiv.org/abs/2205.09515 | Point estimates for all parameters under MCMC and VB for each one of the 100 replicas. The black lines represent the identity function. The scatterplots for the remaining coefficients can be seen in Figure \ref{fig:point_estimates_all_replicas_APP} in the appendix. | Furthermore, similar conclusions can be drawn from the other 99 replicates. Posterior point estimates produced by the proposed ADVI and by the baseline MCMC on all 100 replicas are compared in Figure \ref{fig:point_estimates_all_replicas} and Figure \ref{fig:point_estimates_all_replicas_APP} in the appendix. Point esti... |
2205.09515_FIG_9 | 2205.09515 | https://arxiv.org/abs/2205.09515 | Simulated mean of $y$ as a function of covariates $x_1$ and $x_2$ according to section \ref{sec:scenario3}. Posterior point estimate for the average of $y$ under the proposed MCMC approach. | Figure \ref{fig:persp} shows that the underlying simulated surface for the mean of the response variable $y$ as a function of $x_1$ and $x_2$ is accurately estimated by the MCMC and VB approaches. It is worth to mention that the adopted VB is much faster than MCMC. |
2205.09515_FIG_10 | 2205.09515 | https://arxiv.org/abs/2205.09515 | Comparison of MCMC and GAM when fitting the simulated data from section \ref{sec:scenario3}. | When comparing the model formulation under MCMC, VB, and the point estimates from GAM, Figure \ref{fig:gam_vc_mcmc_multi_covariates} shows that they yield virtually indistinguishable estimates for the non-parametric effects. While GAM penalizes the second-order differences of the basis functions' coefficients, the bri... |
2205.09515_FIG_12 | 2205.09515 | https://arxiv.org/abs/2205.09515 | VB 95\% credibility bands and posterior mean for the average response. Only the first 1000 observations of Energy Charge data are shown. | The proposed ADVI inference approach took approximately 831 seconds (less than 14min) to complete 2000 iterations using 100 Monte Carlo samples to estimate the gradient of the ELBO. On the other hand, MCMC takes 69239 seconds (19.23 hours) to run the same 2000 iterations. Posterior estimates for the beginning and end o... |
2205.09515_FIG_15 | 2205.09515 | https://arxiv.org/abs/2205.09515 | Point estimates for all parameters under MCMC and VB for each one of the 100 replicas. The black lines represent the identity function. | Furthermore, similar conclusions can be drawn from the other 99 replicates. Posterior point estimates produced by the proposed ADVI and by the baseline MCMC on all 100 replicas are compared in Figure \ref{fig:point_estimates_all_replicas} and Figure \ref{fig:point_estimates_all_replicas_APP} in the appendix. Point esti... |
2205.09515_FIG_16 | 2205.09515 | https://arxiv.org/abs/2205.09515 | p-values for the Kolmogorov-Smirnov test for comparison of each marginal posterior predictive distributions under ADVI and MCMC. In red, we highlight P(type 1 error) = 5\%. The expression P(Rej. $H_0$) represents the proportion of times we would reject the null hypothesis of equal distributions under ADVI and MCMC if w... | To further address the quality of the variational approximations to the true posterior distribution, Figure \ref{fig:hist_pvalues} shows the p-values obtained when comparing (via Kolmogorov-Smirnov tests) marginal posterior predictive distributions estimated by VB and MCMC. The shapes of the histograms are approximatel... |
2206.07171_FIG_2 | 2206.07171 | https://arxiv.org/abs/2206.07171 | Categorization of the 38 papers reviewed in this survey. The papers are first categorized on the learning paradigm (fully vs. semi/un/self-supervised) and on the segmentation type (semantic vs. instance). Each quadrant shows the distributions of applications (2D vs. 3D) and DL backbones (U-Net vs. FCN vs. Other) of the... | Fig.~\ref{fig:SearchResultSummary} summarizes this collection of 38 papers in terms of learning technique (fully supervised or not), segmentation type (semantic or instance), application (2D or 3D) and the underlying modeling backbone. Before reviewing these papers, we discuss the key EM datasets and describe the evolu... |
2206.15306_FIG_1 | 2206.15306 | https://arxiv.org/abs/2206.15306 | {\bf Tabular transfer learning pipeline with MetaMIMIC.} We pre-train deep tabular neural networks on abundant upstream patient data with 11 diagnosis targets via multi-label classification. Then, we fine-tune the pre-trained models on limited downstream data with similar features to predict the new target diagnosis. | In total, we have 60 combinations of upstream and downstream datasets for our transfer learning experiments. We pre-train our models as multi-label classifiers on the upstream datasets with 11 targets and then transfer the feature extractors onto the downstream binary diagnosis tasks, Figure \ref{fig: mimic_teaser} pre... |
2206.15306_FIG_3 | 2206.15306 | https://arxiv.org/abs/2206.15306 | {\bf Comparison of supervised and self-supervised pre-training strategies for FT-Transformer.} The left panel illustrates end-to-end fine-tuning with linear head and the right plot illustrates end-to-end fine-tuning with MLP head, the two most effective strategies for FT-Transformer. Within each panel, columns represen... | While self-supervised learning makes for transferable feature extractors in other domains, our experiments show that supervised pre-training is consistently better than the recent SSL pre-training methods we try that are designed for tabular data. In Figure \ref{fig:overall-rank-mainbody-ssl}, we compare supervised pre... |
2206.15306_FIG_12 | 2206.15306 | https://arxiv.org/abs/2206.15306 | {\bf Comparison of different pre-training strategies for FT-Transformer.} The left panel illustrates end-to-end fine-tuning with linear head and the right plot illustrates end-to-end fine-tuning with MLP head. Within each panel, columns represent pre-training strategies and rows correspond to the number of available do... | We extend our comparison of supervised pre-training and contrastive pre-training strategies by including two additional corruption hyperparameters $\lambda=0.7, 0.8$ for contrastive pre-training used in experiments in Section \ref{sec:mlm} (which used $\lambda=0.9$). We also include SCARF contrastive pretraining strate... |
2208.00953_FIG_3 | 2208.00953 | https://arxiv.org/abs/2208.00953 | Random selected 16 samples. The dataset was curated from multiple international COVID-19 X-ray testing centers during several time periods. The dataset is made up of an unbalanced percentage of the four classes in which we have 48.2 $\%$ normal X-ray images, 28.4 $\%$ cases with lung opacity, 17.1 $ \%$ of COVID-19 pat... | Unlike the brain cancer MRI dataset which comes with segmentation masks from experts in the field, the COVID-19 X-ray dataset~\cite{chowdhury2020can} used in this work has no ground truth segmentation masks. This was chosen as an edge-case analysis due to the fact that a vast majority of datasets do not have segmentati... |
2208.00953_FIG_4 | 2208.00953 | https://arxiv.org/abs/2208.00953 | A t-SNE~\cite{van2008visualizing} two-dimensional projection of the unrolled pixel space representation of MRI slices where the colors purple, green, and yellow represent the three classes of Meningioma, Glioma, and Pituitary tumor respectively. However, given that the data is generated under differing physical and st... | With the brain MRI dataset, the DenseNet121 model shows the best overall test performance reaching 98.10\%. While the hybrid InceptionResNetV2 outperformed the other models on the COVID-19 X-ray dataset with an accuracy of $89.0\%$. The test results indicate the high confidence and stability of model prediction. This ... |
2208.00953_FIG_5 | 2208.00953 | https://arxiv.org/abs/2208.00953 | A similar 2D t-SNE visualization of the InceptionResNetV2 latent representations for the chest X-ray dataset. This dataset has a rich statistical structure across all classes, however, it is also imbalanced like many medical datasets. (Right) A plot of latent embeddings prior to training, the dataset is biased towards ... | With the brain MRI dataset, the DenseNet121 model shows the best overall test performance reaching 98.10\%. While the hybrid InceptionResNetV2 outperformed the other models on the COVID-19 X-ray dataset with an accuracy of $89.0\%$. The test results indicate the high confidence and stability of model prediction. This ... |
2208.11176_FIG_1 | 2208.11176 | https://arxiv.org/abs/2208.11176 | Number of papers related to training with noisy labels that use each type of data augmentation. Sixty-one papers from the last years were evaluated for this analysis. | It is a common practice in modern image classification models to use data augmentation to improve the diversity of the training set and increase model generalization capabilities. Several training strategies have been proposed in the last decade to address the noisy label problem ~\cite{algan2021image}. The main strate... |
2202.00574_FIG_6 | 2202.00574 | https://arxiv.org/abs/2202.00574 | \textbf{Varying simulated sample size.} We show the accuracy for varying sizes of the simulated data set. In all cases the simulated pairs were augmented until 50,000 total pairs are created. 25,000 pairs in the training set corresponds to the analysis shown in the main text in Fig~\ref{fig:performance}. | We investigate the role of the number of simulated pairs when training with only simulated data. We simulate 10, 100, 1,000, or 10,000 pairs and augment them until we have 50,000 pairs. The accuracy for classifiers trained on those data sets in comparison with the sample size used in the main text (25,000 sampled pairs... |
2202.00574_FIG_8 | 2202.00574 | https://arxiv.org/abs/2202.00574 | \textbf{Benchmarking results on simulated data with LeNet architecture. }Description as in Fig.~\ref{fig:performance_lenet} but for testing on simulated data. | The results are similar to those found testing on experimental data but with notable changes when comparing the models with different sizes. Fig.~\ref{fig:performance_lenet_on_sim} shows the performance of the smaller LeNet models and Fig.~\ref{fig:performance_resnet_on_sim} of the larger ResNet models. With an ensembl... |
2204.01248_FIG_1 | 2204.01248 | https://arxiv.org/abs/2204.01248 | Block diagram of proposed composite differentiable rendering approach | Therefore, in this work we present a proof-of-concept and starting point for differentiable rendering in SAR imagery. Our proposed approach is composed of a Feature Rasterizer followed by a Neural Shader, depicted in the block diagram of Figure~\ref{fig:compositeDiffRend}. The Feature Rasterizer produces image-like... |
2207.10651_FIG_2 | 2207.10651 | https://arxiv.org/abs/2207.10651 | PDF of $C_L$ obtained from $5000$ Monte Carlo simulations for $a_{\infty} \sim N(4.00 ^{\circ},0.40 ^{\circ})$ and $M_{\infty} \sim N(0.75,0.16)$. | In the first scenario, uncertainty is introduced only in the free-stream conditions, that follow a normal distribution with $a_{\infty} \sim N(4.0 ^{\circ},0.40 ^{\circ})$ and $M_{\infty} \sim N(0.75,0.16)$. We run Monte-Carlo simulations with $5000$ samples to quantify the effect of these uncertainties on the QoI. T... |
2207.10651_FIG_4 | 2207.10651 | https://arxiv.org/abs/2207.10651 | PDF of $C_L$ for the NACA12 airfoil produced by $5000$ Monte-Carlo simulations. The $m = 40$ stochastic inputs are $M_{\infty} = N(0.75,0.16)$, $a_{\infty} = N(4^{\circ},0.4^{\circ})$ and $a_i \sim N(1,0.01)$ with $i = 1,\dots,38$. | The PDF of the lift coefficient $C_L$, obtained from $5000$ Monte Carlo simulations, is shown in figure \ref{NACA12_Monte_Carlo}. The PDF remains bi-modal as in the first scenario, but has different statistical properties. The performance of the airfoil is significantly affected by the geometric uncertainties. For exa... |
2207.11448_FIG_1 | 2207.11448 | https://arxiv.org/abs/2207.11448 | Application of DbM to 2D airfoils. Column 1 shows the baseline shapes. Column 2 depicts the elements of the collocation vectors of the baseline shapes plotted as a function of the index $i$ of the collocation vector. Column 3 shows the weighted elements of the collocation vector plotted as a function of the index $i$ o... | \noindent Here $\vec{S}_n(\vec{x})$ is the $y$-coordinate collocation vector of the $n^{\text{th}}$ baseline shape, collocated at $\vec{x} = \left[ x_0, \cdots , x_F \right]$ where the $i^\text{th}$ $x$-coordinate $x_i = \left| 1-2i/F \right|$ and $F$ is the number of collocation points. The first half of the elements ... |
2207.11448_FIG_2 | 2207.11448 | https://arxiv.org/abs/2207.11448 | Percentage of airfoils reconstructed within the mean absolute error tolerance of 0.5 \% using DbM, compared to those from DbM only with interpolation (DbM-I) and three airfoil shape parameterization methods (see Table~\ref{table:comparison_methods}). | Applying DbM for the reconstruction of the UIUC database, we found that 1,618 of the 1,620 airfoils of the entire UIUC database, were reconstructed with a MAE error <~1\%. Even for the two airfoils with the highest error, the DbM reconstruction still resulted in a MAE error of less than 1.5 \%. Figure~\ref{fig:comparis... |
2207.11448_FIG_4 | 2207.11448 | https://arxiv.org/abs/2207.11448 | Mean weight distributions of the Pareto-optimal airfoil shapes with respect to twenty-five baseline airfoil shapes. | The mean weight distributions with respect to 25 original baseline shapes are shown in Figure~\ref{fig:weightdist}. Overall, the weight distributions of the three clusters conform to the weight distribution of the total mean. It turned out that baseline shape \#13 (model name: AS6097) was commonly the most significant ... |
2207.11448_FIG_5 | 2207.11448 | https://arxiv.org/abs/2207.11448 | Schematic diagram of airfoil performance curves. | can be a proper choice to evaluate the off-design performance \parencite{li_method_2013}. Figure~\ref{fig:prelim} depicts a schematic diagram of how $CLD$ and $\Delta \alpha$ are determined on airfoil performance curves. Our optimization methodology does not rely on one specific airfoil evaluation tool. To compare our... |
2208.01996_FIG_1 | 2208.01996 | https://arxiv.org/abs/2208.01996 | Comparison of ERM, Entropy Minimization (EM)-based Test-Time Adaptation (TTA) method, and the proposed AdaODM. Here we take for instance a two-class classification task with three source domains. Solid and dashed circles represent two classes. Colors represent various domains. (a) shows the vanilla ERM which performs w... | Moreover, losses like entropy minimization can only enforce target features away from decision boundaries whose optimization direction might be less reliable (see Fig. \ref{fig:main}(b)). Such a process can also be seen as achieving domain-invariant feature space encompassed by domain-specific class decision boundarie... |
2203.00407_FIG_1 | 2203.00407 | https://arxiv.org/abs/2203.00407 | Data adjustments according to the methodology of reporting the Covid19 numbers in Romania. The rows describe the data (from top to bottom) of recovered, infected and dead. The left column represents the raw data and the right column represents the adjusted data as we described above. Notice the scale and the spike... | We found a statistically significant difference between the averages of the reported cases according to the day of the week ($p < 0.05$). However, by law, the methodology of reporting from the health centers allows reporting cases within an interval of two weeks. In order to mitigate the above deficiencies, we replac... |
2204.05885_FIG_1 | 2204.05885 | https://arxiv.org/abs/2204.05885 | $\textsl{YouTube}$ network---\textsc{HBDM-Re} runtime complexity in miliseconds (ms) as a function of increasing sample sizes of network nodes (in terms of $N\log N$) until the sample set becomes the node set of the total graph. The y-axis showcases the average runtime over $100$ iterations of the forward pass while th... | We assess the runtime complexity of the \textsc{HBDM-Re} framework in terms of increasing network sizes. In Fig. \ref{fig:compl_yes_}, we consider the \textsl{YouTube} network and show the runtime complexity in milliseconds (ms) as a function of increasing sample sizes of network nodes (in terms of $N\log N$) until the... |
2209.00128_FIG_1 | 2209.00128 | https://arxiv.org/abs/2209.00128 | AP50 of the YOLOv5 models fine-tuned on the various balanced hybrid sets constructed by \textit{Arch-Syn-FT-B} and the different subsets of \textit{Arch-Mann-FT37}. | The results are shown in Fig.~\ref{fig:ft_mann_syn_b},~\ref{fig:ft_mann_syn_b_diff}~and~\ref{fig:ft_mann_syn_b_nano}. Comparing Fig.~\ref{fig:ft_mann_syn_diff} with Fig.~\ref{fig:ft_mann_syn_b_diff}, we can observe that the negative effect of fine-tuning larger models (i.e., \textit{YOLOv5s6} and \textit{YOLOv5m6}) on ... |
2209.00128_FIG_2 | 2209.00128 | https://arxiv.org/abs/2209.00128 | AP50 improvement of the YOLOv5 models after adding \textit{Arch-Syn-FT-B} to the original fine-tuning datasets based on the different subsets of \textit{Arch-Mann-FT37}. | The results are shown in Fig.~\ref{fig:ft_mann_syn_b},~\ref{fig:ft_mann_syn_b_diff}~and~\ref{fig:ft_mann_syn_b_nano}. Comparing Fig.~\ref{fig:ft_mann_syn_diff} with Fig.~\ref{fig:ft_mann_syn_b_diff}, we can observe that the negative effect of fine-tuning larger models (i.e., \textit{YOLOv5s6} and \textit{YOLOv5m6}) on ... |
2209.00128_FIG_3 | 2209.00128 | https://arxiv.org/abs/2209.00128 | AP50 comparison of \textit{YOLOv5n6} fine-tuned on the different hybrid sets of \textit{Arch-Syn-FT}, \textit{Arch-Syn-FT-B}, and the different subsets of \textit{Arch-Mann-FT37}. | The results are shown in Fig.~\ref{fig:ft_mann_syn_b},~\ref{fig:ft_mann_syn_b_diff}~and~\ref{fig:ft_mann_syn_b_nano}. Comparing Fig.~\ref{fig:ft_mann_syn_diff} with Fig.~\ref{fig:ft_mann_syn_b_diff}, we can observe that the negative effect of fine-tuning larger models (i.e., \textit{YOLOv5s6} and \textit{YOLOv5m6}) on ... |
2209.00128_FIG_4 | 2209.00128 | https://arxiv.org/abs/2209.00128 | AP50 of \textit{YOLOv5n6} fine-tuned on the various balanced hybrid sets constructed by the three leave-one-pose-out synthetic fine-tuning datasets (i.e., \textit{Arch-Syn-FT-NoSt-B}, \textit{Arch-Syn-FT-NoKn-B} and \textit{Arch-Syn-FT-NoPr-B}) and the different subsets of \textit{Arch-Mann-FT37}. | The results are presented in Fig.~\ref{fig:ft_mann_syn_b_nano_pose}. In general, the detection accuracy of \textit{stand} and \textit{kneel} did not obviously change when we removed any one of the poses from the synthetic fine-tuning dataset. However, the performance of \textit{prone} degraded drastically when we exclu... |
2209.00128_FIG_9 | 2209.00128 | https://arxiv.org/abs/2209.00128 | AP50 of the pre-trained YOLOv5 models evaluated on \textit{Arch-Mann-Eval} and \textit{Arch-Real-Eval}. | \input{tables/comp_yolov5}\noindent{\bf The Performance of Pre-trained Models.} To begin with, we evaluated the three pre-trained YOLOv5 models on \textit{Arch-Mann-Eval} and \textit{Arch-Real-Eval}. The models were pre-trained on MS-COCO, a representative ground-based dataset. The results are shown in Fig.~\ref{fig:pr... |
2209.00128_FIG_10 | 2209.00128 | https://arxiv.org/abs/2209.00128 | AP50 of the YOLOv5 models fine-tuned on the different subsets of \textit{Arch-Mann-FT37}. Each model was evaluated on \textit{Arch-Mann-Eval} (top) and \textit{Arch-Real-Eval} (bottom). | The results are presented in Fig.~\ref{fig:ft_mann}. We would like to highlight the importance of having an evaluation dataset with different characteristics from the fine-tuning dataset. As we fine-tuned the models on data from \textit{Arch-Mann-FT37}, we could improve their detection accuracy on a similar evaluation ... |
2202.04936_FIG_4 | 2202.04936 | https://arxiv.org/abs/2202.04936 | Mask matrix visualization for \textbf{PubMed} with the first $3000$ nodes and $500$ features at threshold=$0.005$. | Figure~\ref{fig:gae_citeseer}-Figure~\ref{Fig:gae_pubmed} visualize the sparse mask matrix of the five datasets with anomaly injection. We are interested in the recall of the mask matrix that exposes the quality of the mask matrix approximation. As a result, we print the conditional mask matrix that marks the positions... |
2204.01558_FIG_1 | 2204.01558 | https://arxiv.org/abs/2204.01558 | Overview of the Con$^{2}$DA framework. Our model takes weak and strong augmented samples and passes them through a feature extractor. The resulting vectors are normalized to the unit hypersphere. In the supervised case, the standard cross-entropy is applied for weak and strong augmented data. When labels are not availa... | In this work we adopted the normalized feature representation proposed by \cite{saito_2019}, and we extent recent advances on \textbf{Con}sistency regularization and \textbf{Con}trastive learning for the SS\textbf{DA} problem. We present \textbf{Con$^{2}$DA}, a simple framework that takes input images and generates pai... |
2202.00916_FIG_2 | 2202.00916 | https://arxiv.org/abs/2202.00916 | Performance improvement of decision-focused v.s. two-stage method with varying number of trajectories. | As shown in Section~\ref{sec:results} and Figure~\ref{fig:trajectory}, we notice a smaller improvement between decision-focused learning and two-stage approach when there is sufficient data available in the training set. This is because the data is sufficient enough to train a predictive model with small predictive los... |
2206.04596_FIG_3 | 2206.04596 | https://arxiv.org/abs/2206.04596 | Communication flowchart for a $2\times2$ module. The control of three actuators of each robot in a four-robot module is accomplished by three motor drivers. Colored arrows show the distributed control framework between drivers, actuators, and ADCs. | To efficiently control the entire array of 64 robots, communication factors like latency, noise, and amplitude of signal need to have stable optimal values. Instead of using TTL communication using wires, which results in exhaustive cable management and noise, we use off-the-shelf ESP-01 WiFi modules operating at 115,2... |
2206.04596_FIG_4 | 2206.04596 | https://arxiv.org/abs/2206.04596 | Top-view and side-view representation of the ellipsoid trajectories of the two-beat gaits. (a) represents 2 odd rows from the array for execution of the Dexterous Gripping Primitive and (c) represents the entire delta array executing the Planar Rotation Primitive. Both follow two-beat gait patterns shown in (b) and (d)... | \item Facing Up - Objects are placed on top of the array and manipulated from underneath or on the sides (Fig. \ref{fig: upsideUp}). The delta array policies are designed as two-beat finger-gaiting strategies that repeatedly cause the deltas to make and break contact with the objects being manipulated. The planar t... |
2206.04596_FIG_7 | 2206.04596 | https://arxiv.org/abs/2206.04596 | Using REPS to generate weights of the basis functions of a DMP to generate a 2D trajectory (curved lines from (0,0) till end-point predicted by REPS) for the delta robots. Faint trajectories represent initial exploration trajectories with low/negative rewards, darker trajectories represent learned exploitation with hig... | We implemented the learning of dexterous manipulation strategies in Section \ref{dexmanip} for the downward facing configuration. This was done only in the facing-down configuration due to the ease of resetting the environment as compared to the upright configuration.The robot could learn to grasp and tilt objects plac... |
2207.00721_FIG_4 | 2207.00721 | https://arxiv.org/abs/2207.00721 | Convergence of Gaussian curves over training epochs for all the parameters of a robot in 1 experiment. | Fig. \ref{fig:figure5} Shows the general trend that the Gaussian curves follow in due course of training a parameter using REPS. Updates over all the parameters in a single experiment for a robot are shown in Fig. \ref{fig:figure6}. The stochasticity of the algorithm depends on the physical characteristics of the robot... |
End of preview. Expand in Data Studio
SciFigData
This dataset contains figures (as .png files) extracted from scientific articles.
The metadata file contains the caption and referencing text passages for all tables.
To access the raw .png files, the tar files must be extracted.
Purpose
This dataset is part of the VQA Benchmark developed for my master's thesis. It is designed to evaluate Multimodal LLMs on VQA tasks focused on scientific content.
- Downloads last month
- 58