Datasets:

dim

/

scitldr

like 0

Domain adaptation is critical for success in new, unseen environments. Adversarial adaptation models applied in feature spaces discover domain invariant representations, but are difficult to visualize and sometimes fail to capture pixel-level and low-level domain shifts. Recent work has shown that generative adversarial networks combined with cycle-consistency constraints are surprisingly effective at mapping images between domains, even without the use of aligned image pairs. We propose a novel discriminatively-trained Cycle-Consistent Adversarial Domain Adaptation model. CyCADA adapts representations at both the pixel-level and feature-level, enforces cycle-consistency while leveraging a task loss, and does not require aligned pairs. Our model can be applied in a variety of visual recognition and prediction settings. We show new state-of-the-art results across multiple adaptation tasks, including digit classification and semantic segmentation of road scenes demonstrating transfer from synthetic to real world domains.

An unsupervised domain adaptation approach which adapts at both the pixel and feature levels

Stemming is the process of removing affixes( i.e. prefixes, infixes and suffixes) that improve the accuracy and performance of information retrieval systems.This paper presents the reduction of Amharic words to corresponding stem where with the intention that it preserves semantic information. The proposed approach efficiently removes affixes from an Amharic word. The process of removing such affixes (prefixes, infixes and suffixes) from a word to its base form is called stemming. While many stemmers exist for dominant languages such as English, under resourced languages such as Amharic which lacks such powerful tool support. In this paper, we design a light Amharic stemmer relying on the rules that receives an Amharic word and then it finds a match to the beginning of a word to the possible prefixes and to its ending with the possible suffixes and finally it checks whether it has infix. The final result is the stem if there is any prefix, infix or/and suffix, otherwise it remains in one of the earlier states. The technique does not rely on any additional resource (e.g. dictionary) to verify the generated stem. The performance of the generated stemmer is evaluated using manually annotated Amharic words. The result is compared with current state-of-the-art stemmer for Amharic showing an increase of 7% in stemmer correctness.

Amharic Light Stemmer is designed for improving performance of Amharic Sentiment Classification.

Place and grid-cells are known to aid navigation in animals and humans. Together with concept cells, they allow humans to form an internal representation of the external world, namely the concept space. We investigate the presence of such a space in deep neural networks by plotting the activation profile of its hidden layer neurons. Although place cell and concept-cell like properties are found, grid-cell like firing patterns are absent thereby indicating a lack of path integration or feature transformation functionality in trained networks. Overall, we present a plausible inadequacy in current deep learning practices that restrict deep networks from performing analogical reasoning and memory retrieval tasks.

We investigated if simple deep networks possess grid cell-like artificial neurons while memory retrieval in the learned concept space.

We develop a comprehensive description of the active inference framework, as proposed by Friston (2010), under a machine-learning compliant perspective. Stemming from a biological inspiration and the auto-encoding principles, a sketch of a cognitive architecture is proposed that should provide ways to implement estimation-oriented control policies. Computer simulations illustrate the effectiveness of the approach through a foveated inspection of the input data. The pros and cons of the control policy are analyzed in detail, showing interesting promises in terms of processing compression. Though optimizing future posterior entropy over the actions set is shown enough to attain locally optimal action selection, offline calculation using class-specific saliency maps is shown better for it saves processing costs through saccades pathways pre-processing, with a negligible effect on the recognition/compression rates.

Pros and cons of saccade-based computer vision under a predictive coding perspective

Graphs possess exotic features like variable size and absence of natural ordering of the nodes that make them difficult to analyze and compare. To circumvent this problem and learn on graphs, graph feature representation is required. Main difficulties with feature extraction lie in the trade-off between expressiveness, consistency and efficiency, i.e. the capacity to extract features that represent the structural information of the graph while being deformation-consistent and isomorphism-invariant. While state-of-the-art methods enhance expressiveness with powerful graph neural-networks, we propose to leverage natural spectral properties of graphs to study a simple graph feature: the graph Laplacian spectrum (GLS). We analyze the representational power of this object that satisfies both isomorphism-invariance, expressiveness and deformation-consistency. In particular, we propose a theoretical analysis based on graph perturbation to understand what kind of comparison between graphs we do when comparing GLS. To do so, we derive bounds for the distance between GLS that are related to the divergence to isomorphism, a standard computationally expensive graph divergence. Finally, we experiment GLS as graph representation through consistency tests and classification tasks, and show that it is a strong graph feature representation baseline.

We study theoretically the consistency the Laplacian spectrum and use it as whole-graph embeddding

Adversarial training, a method for learning robust deep networks, is typically assumed to be more expensive than traditional training due to the necessity of constructing adversarial examples via a first-order method like projected gradient decent (PGD). In this paper, we make the surprising discovery that it is possible to train empirically robust models using a much weaker and cheaper adversary, an approach that was previously believed to be ineffective, rendering the method no more costly than standard training in practice. Specifically, we show that adversarial training with the fast gradient sign method (FGSM), when combined with random initialization, is as effective as PGD-based training but has significantly lower cost. Furthermore we show that FGSM adversarial training can be further accelerated by using standard techniques for efficient training of deep networks, allowing us to learn a robust CIFAR10 classifier with 45% robust accuracy at epsilon=8/255 in 6 minutes, and a robust ImageNet classifier with 43% robust accuracy at epsilon=2/255 in 12 hours, in comparison to past work based on ``free'' adversarial training which took 10 and 50 hours to reach the same respective thresholds.

FGSM-based adversarial training, with randomization, works just as well as PGD-based adversarial training: we can use this to train a robust classifier in 6 minutes on CIFAR10, and 12 hours on ImageNet, on a single machine.

In seeking for sparse and efficient neural network models, many previous works investigated on enforcing L1 or L0 regularizers to encourage weight sparsity during training. The L0 regularizer measures the parameter sparsity directly and is invariant to the scaling of parameter values. But it cannot provide useful gradients and therefore requires complex optimization techniques. The L1 regularizer is almost everywhere differentiable and can be easily optimized with gradient descent. Yet it is not scale-invariant and causes the same shrinking rate to all parameters, which is inefficient in increasing sparsity. Inspired by the Hoyer measure (the ratio between L1 and L2 norms) used in traditional compressed sensing problems, we present DeepHoyer, a set of sparsity-inducing regularizers that are both differentiable almost everywhere and scale-invariant. Our experiments show that enforcing DeepHoyer regularizers can produce even sparser neural network models than previous works, under the same accuracy level. We also show that DeepHoyer can be applied to both element-wise and structural pruning.

We propose almost everywhere differentiable and scale invariant regularizers for DNN pruning, which can lead to supremum sparsity through standard SGD training.

Self-supervision, in which a target task is improved without external supervision, has primarily been explored in settings that assume the availability of additional data. However, in many cases, particularly in healthcare, one may not have access to additional data (labeled or otherwise). In such settings, we hypothesize that self-supervision based solely on the structure of the data at-hand can help. We explore a novel self-supervision framework for time-series data, in which multiple auxiliary tasks (e.g., forecasting) are included to improve overall performance on a sequence-level target task without additional training data. We call this approach limited self-supervision, as we limit ourselves to only the data at-hand. We demonstrate the utility of limited self-supervision on three sequence-level classification tasks, two pertaining to real clinical data and one using synthetic data. Within this framework, we introduce novel forms of self-supervision and demonstrate their utility in improving performance on the target task. Our results indicate that limited self-supervision leads to a consistent improvement over a supervised baseline, across a range of domains. In particular, for the task of identifying atrial fibrillation from small amounts of electrocardiogram data, we observe a nearly 13% improvement in the area under the receiver operating characteristics curve (AUC-ROC) relative to the baseline (AUC-ROC=0.55 vs. AUC-ROC=0.62). Limited self-supervision applied to sequential data can aid in learning intermediate representations, making it particularly applicable in settings where data collection is difficult.

We show that extra unlabeled data is not required for self-supervised auxiliary tasks to be useful for time series classification, and present new and effective auxiliary tasks.

Are neural networks biased toward simple functions? Does depth always help learn more complex features? Is training the last layer of a network as good as training all layers? These questions seem unrelated at face value, but in this work we give all of them a common treatment from the spectral perspective. We will study the spectra of the *Conjugate Kernel, CK,* (also called the *Neural Network-Gaussian Process Kernel*), and the *Neural Tangent Kernel, NTK*. Roughly, the CK and the NTK tell us respectively ``"what a network looks like at initialization" and "``what a network looks like during and after training." Their spectra then encode valuable information about the initial distribution and the training and generalization properties of neural networks. By analyzing the eigenvalues, we lend novel insights into the questions put forth at the beginning, and we verify these insights by extensive experiments of neural networks. We believe the computational tools we develop here for analyzing the spectra of CK and NTK serve as a solid foundation for future studies of deep neural networks. We have open-sourced the code for it and for generating the plots in this paper at github.com/jxVmnLgedVwv6mNcGCBy/NNspectra.

Eigenvalues of Conjugate (aka NNGP) and Neural Tangent Kernel can be computed in closed form over the Boolean cube and reveal the effects of hyperparameters on neural network inductive bias, training, and generalization.

To communicate, to ground hypotheses, to analyse data, neuroscientists often refer to divisions of the brain. Here we consider atlases used to parcellate the brain when studying brain function. We discuss the meaning and the validity of these parcellations, from a conceptual point of view as well as by running various analytical tasks on popular functional brain parcellations.

All functional brain parcellations are wrong, but some are useful

High-dimensional sparse reward tasks present major challenges for reinforcement learning agents. In this work we use imitation learning to address two of these challenges: how to learn a useful representation of the world e.g. from pixels, and how to explore efficiently given the rarity of a reward signal? We show that adversarial imitation can work well even in this high dimensional observation space. Surprisingly the adversary itself, acting as the learned reward function, can be tiny, comprising as few as 128 parameters, and can be easily trained using the most basic GAN formulation. Our approach removes limitations present in most contemporary imitation approaches: requiring no demonstrator actions (only video), no special initial conditions or warm starts, and no explicit tracking of any single demo. The proposed agent can solve a challenging robot manipulation task of block stacking from only video demonstrations and sparse reward, in which the non-imitating agents fail to learn completely. Furthermore, our agent learns much faster than competing approaches that depend on hand-crafted, staged dense reward functions, and also better compared to standard GAIL baselines. Finally, we develop a new adversarial goal recognizer that in some cases allows the agent to learn stacking without any task reward, purely from imitation.

Imitation from pixels, with sparse or no reward, using off-policy RL and a tiny adversarially-learned reward function.

In this paper we show strategies to easily identify fake samples generated with the Generative Adversarial Network framework. One strategy is based on the statistical analysis and comparison of raw pixel values and features extracted from them. The other strategy learns formal specifications from the real data and shows that fake samples violate the specifications of the real data. We show that fake samples produced with GANs have a universal signature that can be used to identify fake samples. We provide results on MNIST, CIFAR10, music and speech data.

We show strategies to easily identify fake samples generated with the Generative Adversarial Network framework.

Efforts to reduce the numerical precision of computations in deep learning training have yielded systems that aggressively quantize weights and activations, yet employ wide high-precision accumulators for partial sums in inner-product operations to preserve the quality of convergence. The absence of any framework to analyze the precision requirements of partial sum accumulations results in conservative design choices. This imposes an upper-bound on the reduction of complexity of multiply-accumulate units. We present a statistical approach to analyze the impact of reduced accumulation precision on deep learning training. Observing that a bad choice for accumulation precision results in loss of information that manifests itself as a reduction in variance in an ensemble of partial sums, we derive a set of equations that relate this variance to the length of accumulation and the minimum number of bits needed for accumulation. We apply our analysis to three benchmark networks: CIFAR-10 ResNet 32, ImageNet ResNet 18 and ImageNet AlexNet. In each case, with accumulation precision set in accordance with our proposed equations, the networks successfully converge to the single precision floating-point baseline. We also show that reducing accumulation precision further degrades the quality of the trained network, proving that our equations produce tight bounds. Overall this analysis enables precise tailoring of computation hardware to the application, yielding area- and power-optimal systems.

We present an analytical framework to determine accumulation bit-width requirements in all three deep learning training GEMMs and verify the validity and tightness of our method via benchmarking experiments.

Unsupervised domain adaptation is a promising avenue to enhance the performance of deep neural networks on a target domain, using labels only from a source domain. However, the two predominant methods, domain discrepancy reduction learning and semi-supervised learning, are not readily applicable when source and target domains do not share a common label space. This paper addresses the above scenario by learning a representation space that retains discriminative power on both the (labeled) source and (unlabeled) target domains while keeping representations for the two domains well-separated. Inspired by a theoretical analysis, we first reformulate the disjoint classification task, where the source and target domains correspond to non-overlapping class labels, to a verification one. To handle both within and cross domain verifications, we propose a Feature Transfer Network (FTN) to separate the target feature space from the original source space while aligned with a transformed source space. Moreover, we present a non-parametric multi-class entropy minimization loss to further boost the discriminative power of FTNs on the target domain. In experiments, we first illustrate how FTN works in a controlled setting of adapting from MNIST-M to MNIST with disjoint digit classes between the two domains and then demonstrate the effectiveness of FTNs through state-of-the-art performances on a cross-ethnicity face recognition problem.

A new theory of unsupervised domain adaptation for distance metric learning and its application to face recognition across diverse ethnicity variations.

In this paper, we consider the problem of training neural networks (NN). To promote a NN with specific structures, we explicitly take into consideration the nonsmooth regularization (such as L1-norm) and constraints (such as interval constraint). This is formulated as a constrained nonsmooth nonconvex optimization problem, and we propose a convergent proximal-type stochastic gradient descent (Prox-SGD) algorithm. We show that under properly selected learning rates, momentum eventually resembles the unknown real gradient and thus is crucial in analyzing the convergence. We establish that with probability 1, every limit point of the sequence generated by the proposed Prox-SGD is a stationary point. Then the Prox-SGD is tailored to train a sparse neural network and a binary neural network, and the theoretical analysis is also supported by extensive numerical tests.

We propose a convergent proximal-type stochastic gradient descent algorithm for constrained nonsmooth nonconvex optimization problems

The loss of a few neurons in a brain rarely results in any visible loss of function. However, the insight into what “few” means in this context is unclear. How many random neuron failures will it take to lead to a visible loss of function? In this paper, we address the fundamental question of the impact of the crash of a random subset of neurons on the overall computation of a neural network and the error in the output it produces. We study fault tolerance of neural networks subject to small random neuron/weight crash failures in a probabilistic setting. We give provable guarantees on the robustness of the network to these crashes. Our main contribution is a bound on the error in the output of a network under small random Bernoulli crashes proved by using a Taylor expansion in the continuous limit, where close-by neurons at a layer are similar. The failure mode we adopt in our model is characteristic of neuromorphic hardware, a promising technology to speed up artificial neural networks, as well as of biological networks. We show that our theoretical bounds can be used to compare the fault tolerance of different architectures and to design a regularizer improving the fault tolerance of a given architecture. We design an algorithm achieving fault tolerance using a reasonable number of neurons. In addition to the theoretical proof, we also provide experimental validation of our results and suggest a connection to the generalization capacity problem.

We give a bound for NNs on the output error in case of random weight failures using a Taylor expansion in the continuous limit where nearby neurons are similar

Truly intelligent agents need to capture the interplay of all their senses to build a rich physical understanding of their world. In robotics, we have seen tremendous progress in using visual and tactile perception; however we have often ignored a key sense: sound. This is primarily due to lack of data that captures the interplay of action and sound. In this work, we perform the first large-scale study of the interactions between sound and robotic action. To do this, we create the largest available sound-action-vision dataset with 15,000 interactions on 60 objects using our robotic platform Tilt-Bot. By tilting objects and allowing them to crash into the walls of a robotic tray, we collect rich four-channel audio information. Using this data, we explore the synergies between sound and action, and present three key insights. First, sound is indicative of fine-grained object class information, e.g., sound can differentiate a metal screwdriver from a metal wrench. Second, sound also contains information about the causal effects of an action, i.e. given the sound produced, we can predict what action was applied on the object. Finally, object representations derived from audio embeddings are indicative of implicit physical properties. We demonstrate that on previously unseen objects, audio embeddings generated through interactions can predict forward models 24% better than passive visual embeddings.

We explore and study the synergies between sound and action.

Hierarchical label structures widely exist in many machine learning tasks, ranging from those with explicit label hierarchies such as image classification to the ones that have latent label hierarchies such as semantic segmentation. Unfortunately, state-of-the-art methods often utilize cross-entropy loss which in-explicitly assumes the independence among class labels. Motivated by the fact that class members from the same hierarchy need to be similar to each others, we design a new training diagram called Hierarchical Complement Objective Training (HCOT). In HCOT, in addition to maximizing the probability of the ground truth class, we also neutralize the probabilities of rest of the classes in a hierarchical fashion, making the model take advantage of the label hierarchy explicitly. We conduct our method on both image classification and semantic segmentation. Results show that HCOT outperforms state-of-the-art models in CIFAR100, Imagenet, and PASCAL-context. Our experiments also demonstrate that HCOT can be applied on tasks with latent label hierarchies, which is a common characteristic in many machine learning tasks.

We propose Hierarchical Complement Objective Training, a novel training paradigm to effectively leverage category hierarchy in the labeling space on both image classification and semantic segmentation.

There is a growing interest in automated neural architecture search (NAS). To improve the efficiency of NAS, previous approaches adopt weight sharing method to force all models share the same set of weights. However, it has been observed that a model performing better with shared weights does not necessarily perform better when trained alone. In this paper, we analyse existing weight sharing one-shot NAS approaches from a Bayesian point of view and identify the posterior fading problem, which compromises the effectiveness of shared weights. To alleviate this problem, we present a practical approach to guide the parameter posterior towards its true distribution. Moreover, a hard latency constraint is introduced during the search so that the desired latency can be achieved. The resulted method, namely Posterior Convergent NAS (PC-NAS), achieves state-of-the-art performance under standard GPU latency constraint on ImageNet. In our small search space, our model PC-NAS-S attains76.8% top-1 accuracy, 2.1% higher than MobileNetV2 (1.4x) with the same latency. When adopted to our large search space, PC-NAS-L achieves 78.1% top-1 accuracy within 11ms. The discovered architecture also transfers well to other computer vision applications such as object detection and person re-identification.

Our paper identifies the issue of existing weight sharing approach in neural architecture search and propose a practical method, achieving strong results.

Noisy labels are very common in real-world training data, which lead to poor generalization on test data because of overfitting to the noisy labels. In this paper, we claim that such overfitting can be avoided by "early stopping" training a deep neural network before the noisy labels are severely memorized. Then, we resume training the early stopped network using a "maximal safe set," which maintains a collection of almost certainly true-labeled samples at each epoch since the early stop point. Putting them all together, our novel two-phase training method, called Prestopping, realizes noise-free training under any type of label noise for practical use. Extensive experiments using four image benchmark data sets verify that our method significantly outperforms four state-of-the-art methods in test error by 0.4–8.2 percent points under existence of real-world noise.

We propose a novel two-phase training approach based on "early stopping" for robust training on noisy labels.

Learning when to communicate and doing that effectively is essential in multi-agent tasks. Recent works show that continuous communication allows efficient training with back-propagation in multi-agent scenarios, but have been restricted to fully-cooperative tasks. In this paper, we present Individualized Controlled Continuous Communication Model (IC3Net) which has better training efficiency than simple continuous communication model, and can be applied to semi-cooperative and competitive settings along with the cooperative settings. IC3Net controls continuous communication with a gating mechanism and uses individualized rewards foreach agent to gain better performance and scalability while fixing credit assignment issues. Using variety of tasks including StarCraft BroodWars explore and combat scenarios, we show that our network yields improved performance and convergence rates than the baselines as the scale increases. Our results convey that IC3Net agents learn when to communicate based on the scenario and profitability.

We introduce IC3Net, a single network which can be used to train agents in cooperative, competitive and mixed scenarios. We also show that agents can learn when to communicate using our model.

Neural sequence-to-sequence models are a recently proposed family of approaches used in abstractive summarization of text documents, useful for producing condensed versions of source text narratives without being restricted to using only words from the original text. Despite the advances in abstractive summarization, custom generation of summaries (e.g. towards a user's preference) remains unexplored. In this paper, we present CATS, an abstractive neural summarization model, that summarizes content in a sequence-to-sequence fashion but also introduces a new mechanism to control the underlying latent topic distribution of the produced summaries. Our experimental results on the well-known CNN/DailyMail dataset show that our model achieves state-of-the-art performance.

We present the first neural abstractive summarization model capable of customization of generated summaries.

We propose a software framework based on ideas of the Learning-Compression algorithm , that allows one to compress any neural network by different compression mechanisms (pruning, quantization, low-rank, etc.). By design, the learning of the neural net (handled by SGD) is decoupled from the compression of its parameters (handled by a signal compression function), so that the framework can be easily extended to handle different combinations of neural net and compression type. In addition, it has other advantages, such as easy integration with deep learning frameworks, efficient training time, competitive practical performance in the loss-compression tradeoff, and reasonable convergence guarantees. Our toolkit is written in Python and Pytorch and we plan to make it available by the workshop time, and eventually open it for contributions from the community.

We propose a software framework based on ideas of the Learning-Compression algorithm , that allows one to compress any neural network by different compression mechanisms (pruning, quantization, low-rank, etc.).

This work seeks the possibility of generating the human face from voice solely based on the audio-visual data without any human-labeled annotations. To this end, we propose a multi-modal learning framework that links the inference stage and generation stage. First, the inference networks are trained to match the speaker identity between the two different modalities. Then the pre-trained inference networks cooperate with the generation network by giving conditional information about the voice.

This paper proposes a method of end-to-end multi-modal generation of human face from speech based on a self-supervised learning framework.

We present a simple neural model that given a formula and a property tries to answer the question whether the formula has the given property, for example whether a propositional formula is always true. The structure of the formula is captured by a feedforward neural network recursively built for the given formula in a top-down manner. The results of this network are then processed by two recurrent neural networks. One of the interesting aspects of our model is how propositional atoms are treated. For example, the model is insensitive to their names, it only matters whether they are the same or distinct.

A top-down approach how to recursively represent propositional formulae by neural networks is presented.

Despite significant advances in the field of deep Reinforcement Learning (RL), today's algorithms still fail to learn human-level policies consistently over a set of diverse tasks such as Atari 2600 games. We identify three key challenges that any algorithm needs to master in order to perform well on all games: processing diverse reward distributions, reasoning over long time horizons, and exploring efficiently. In this paper, we propose an algorithm that addresses each of these challenges and is able to learn human-level policies on nearly all Atari games. A new transformed Bellman operator allows our algorithm to process rewards of varying densities and scales; an auxiliary temporal consistency loss allows us to train stably using a discount factor of 0.999 (instead of 0.99) extending the effective planning horizon by an order of magnitude; and we ease the exploration problem by using human demonstrations that guide the agent towards rewarding states. When tested on a set of 42 Atari games, our algorithm exceeds the performance of an average human on 40 games using a common set of hyper parameters.

Ape-X DQfD = Distributed (many actors + one learner + prioritized replay) DQN with demonstrations optimizing the unclipped 0.999-discounted return on Atari.

The knowledge that humans hold about a problem often extends far beyond a set of training data and output labels. While the success of deep learning mostly relies on supervised training, important properties cannot be inferred efficiently from end-to-end annotations alone, for example causal relations or domain-specific invariances. We present a general technique to supplement supervised training with prior knowledge expressed as relations between training instances. We illustrate the method on the task of visual question answering to exploit various auxiliary annotations, including relations of equivalence and of logical entailment between questions. Existing methods to use these annotations, including auxiliary losses and data augmentation, cannot guarantee the strict inclusion of these relations into the model since they require a careful balancing against the end-to-end objective. Our method uses these relations to shape the embedding space of the model, and treats them as strict constraints on its learned representations. %The resulting model encodes relations that better generalize across instances. In the context of VQA, this approach brings significant improvements in accuracy and robustness, in particular over the common practice of incorporating the constraints as a soft regularizer. We also show that incorporating this type of prior knowledge with our method brings consistent improvements, independently from the amount of supervised data used. It demonstrates the value of an additional training signal that is otherwise difficult to extract from end-to-end annotations alone.

Training method to enforce strict constraints on learned embeddings during supervised training. Applied to visual question answering.

Artificial neural networks revolutionized many areas of computer science in recent years since they provide solutions to a number of previously unsolved problems. On the other hand, for many problems, classic algorithms exist, which typically exceed the accuracy and stability of neural networks. To combine these two concepts, we present a new kind of neural networks—algorithmic neural networks (AlgoNets). These networks integrate smooth versions of classic algorithms into the topology of neural networks. Our novel reconstructive adversarial network (RAN) enables solving inverse problems without or with only weak supervision.

Solving inverse problems by using smooth approximations of the forward algorithms to train the inverse models.

Pointwise localization allows more precise localization and accurate interpretability, compared to bounding box, in applications where objects are highly unstructured such as in medical domain. In this work, we focus on weakly supervised localization (WSL) where a model is trained to classify an image and localize regions of interest at pixel-level using only global image annotation. Typical convolutional attentions maps are prune to high false positive regions. To alleviate this issue, we propose a new deep learning method for WSL, composed of a localizer and a classifier, where the localizer is constrained to determine relevant and irrelevant regions using conditional entropy (CE) with the aim to reduce false positive regions. Experimental results on a public medical dataset and two natural datasets, using Dice index, show that, compared to state of the art WSL methods, our proposal can provide significant improvements in terms of image-level classification and pixel-level localization (low false positive) with robustness to overfitting. A public reproducible PyTorch implementation is provided.

A deep learning method for weakly-supervised pointwise localization that learns using image-level label only. It relies on conditional entropy to localize relevant and irrelevant regions aiming to minimize false positive regions.

Model-based reinforcement learning has been empirically demonstrated as a successful strategy to improve sample efficiency. Particularly, Dyna architecture, as an elegant model-based architecture integrating learning and planning, provides huge flexibility of using a model. One of the most important components in Dyna is called search-control, which refers to the process of generating state or state-action pairs from which we query the model to acquire simulated experiences. Search-control is critical to improve learning efficiency. In this work, we propose a simple and novel search-control strategy by searching high frequency region on value function. Our main intuition is built on Shannon sampling theorem from signal processing, which indicates that a high frequency signal requires more samples to reconstruct. We empirically show that a high frequency function is more difficult to approximate. This suggests a search-control strategy: we should use states in high frequency region of the value function to query the model to acquire more samples. We develop a simple strategy to locally measure the frequency of a function by gradient norm, and provide theoretical justification for this approach. We then apply our strategy to search-control in Dyna, and conduct experiments to show its property and effectiveness on benchmark domains.

Acquire states from high frequency region for search-control in Dyna.

We propose a new architecture for distributed image compression from a group of distributed data sources. The work is motivated by practical needs of data-driven codec design, low power consumption, robustness, and data privacy. The proposed architecture, which we refer to as Distributed Recurrent Autoencoder for Scalable Image Compression (DRASIC), is able to train distributed encoders and one joint decoder on correlated data sources. Its compression capability is much better than the method of training codecs separately. Meanwhile, for 10 distributed sources, our distributed system remarkably performs within 2 dB peak signal-to-noise ratio (PSNR) of that of a single codec trained with all data sources. We experiment distributed sources with different correlations and show how our methodology well matches the Slepian-Wolf Theorem in Distributed Source Coding (DSC). Our method is also shown to be robust to the lack of presence of encoded data from a number of distributed sources. Moreover, it is scalable in the sense that codes can be decoded simultaneously at more than one compression quality level. To the best of our knowledge, this is the first data-driven DSC framework for general distributed code design with deep learning.

We introduce a data-driven Distributed Source Coding framework based on Distributed Recurrent Autoencoder for Scalable Image Compression (DRASIC).

Long short-term memory networks (LSTMs) were introduced to combat vanishing gradients in simple recurrent neural networks (S-RNNs) by augmenting them with additive recurrent connections controlled by gates. We present an alternate view to explain the success of LSTMs: the gates themselves are powerful recurrent models that provide more representational power than previously appreciated. We do this by showing that the LSTM's gates can be decoupled from the embedded S-RNN, producing a restricted class of RNNs where the main recurrence computes an element-wise weighted sum of context-independent functions of the inputs. Experiments on a range of challenging NLP problems demonstrate that the simplified gate-based models work substantially better than S-RNNs, and often just as well as the original LSTMs, strongly suggesting that the gates are doing much more in practice than just alleviating vanishing gradients.

Gates do all the heavy lifting in LSTMs by computing element-wise weighted sums, and removing the internal simple RNN does not degrade model performance.

Machine learning algorithms designed to characterize, monitor, and intervene on human health (ML4H) are expected to perform safely and reliably when operating at scale, potentially outside strict human supervision. This requirement warrants a stricter attention to issues of reproducibility than other fields of machine learning. In this work, we conduct a systematic evaluation of over 100 recently published ML4H research papers along several dimensions related to reproducibility we identified. We find that the field of ML4H compares poorly to more established machine learning fields, particularly concerning data accessibility and code accessibility. Finally, drawing from success in other fields of science, we propose recommendations to data providers, academic publishers, and the ML4H research community in order to promote reproducible research moving forward.

By analyzing more than 300 papers in recent machine learning conferences, we found that Machine Learning for Health (ML4H) applications lag behind other machine learning fields in terms of reproducibility metrics.

We propose a solution for evaluation of mathematical expression. However, instead of designing a single end-to-end model we propose a Lego bricks style architecture. In this architecture instead of training a complex end-to-end neural network, many small networks can be trained independently each accomplishing one specific operation and acting a single lego brick. More difficult or complex task can then be solved using a combination of these smaller network. In this work we first identify 8 fundamental operations that are commonly used to solve arithmetic operations (such as 1 digit multiplication, addition, subtraction, sign calculator etc). These fundamental operations are then learned using simple feed forward neural networks. We then shows that different operations can be designed simply by reusing these smaller networks. As an example we reuse these smaller networks to develop larger and a more complex network to solve n-digit multiplication, n-digit division, and cross product. This bottom-up strategy not only introduces reusability, we also show that it allows to generalize for computations involving n-digits and we show results for up to 7 digit numbers. Unlike existing methods, our solution also generalizes for both positive as well as negative numbers.

We train many small networks each for a specific operation, these are then combined to perform complex operations

In standard generative adversarial network (SGAN), the discriminator estimates the probability that the input data is real. The generator is trained to increase the probability that fake data is real. We argue that it should also simultaneously decrease the probability that real data is real because 1) this would account for a priori knowledge that half of the data in the mini-batch is fake, 2) this would be observed with divergence minimization, and 3) in optimal settings, SGAN would be equivalent to integral probability metric (IPM) GANs. We show that this property can be induced by using a relativistic discriminator which estimate the probability that the given real data is more realistic than a randomly sampled fake data. We also present a variant in which the discriminator estimate the probability that the given real data is more realistic than fake data, on average. We generalize both approaches to non-standard GAN loss functions and we refer to them respectively as Relativistic GANs (RGANs) and Relativistic average GANs (RaGANs). We show that IPM-based GANs are a subset of RGANs which use the identity function. Empirically, we observe that 1) RGANs and RaGANs are significantly more stable and generate higher quality data samples than their non-relativistic counterparts, 2) Standard RaGAN with gradient penalty generate data of better quality than WGAN-GP while only requiring a single discriminator update per generator update (reducing the time taken for reaching the state-of-the-art by 400%), and 3) RaGANs are able to generate plausible high resolutions images (256x256) from a very small sample (N=2011), while GAN and LSGAN cannot; these images are of significantly better quality than the ones generated by WGAN-GP and SGAN with spectral normalization. The code is freely available on https://github.com/AlexiaJM/RelativisticGAN.

Improving the quality and stability of GANs using a relativistic discriminator; IPM GANs (such as WGAN-GP) are a special case.

Some of the most successful applications of deep reinforcement learning to challenging domains in discrete and continuous control have used policy gradient methods in the on-policy setting. However, policy gradients can suffer from large variance that may limit performance, and in practice require carefully tuned entropy regularization to prevent policy collapse. As an alternative to policy gradient algorithms, we introduce V-MPO, an on-policy adaptation of Maximum a Posteriori Policy Optimization (MPO) that performs policy iteration based on a learned state-value function. We show that V-MPO surpasses previously reported scores for both the Atari-57 and DMLab-30 benchmark suites in the multi-task setting, and does so reliably without importance weighting, entropy regularization, or population-based tuning of hyperparameters. On individual DMLab and Atari levels, the proposed algorithm can achieve scores that are substantially higher than has previously been reported. V-MPO is also applicable to problems with high-dimensional, continuous action spaces, which we demonstrate in the context of learning to control simulated humanoids with 22 degrees of freedom from full state observations and 56 degrees of freedom from pixel observations, as well as example OpenAI Gym tasks where V-MPO achieves substantially higher asymptotic scores than previously reported.

A state-value function-based version of MPO that achieves good results in a wide range of tasks in discrete and continuous control.

Turing complete computation and reasoning are often regarded as necessary pre- cursors to general intelligence. There has been a significant body of work studying neural networks that mimic general computation, but these networks fail to generalize to data distributions that are outside of their training set. We study this problem through the lens of fundamental computer science problems: sorting and graph processing. We modify the masking mechanism of a transformer in order to allow them to implement rudimentary functions with strong generalization. We call this model the Neural Execution Engine, and show that it learns, through supervision, to numerically compute the basic subroutines comprising these algorithms with near perfect accuracy. Moreover, it retains this level of accuracy while generalizing to unseen data and long sequences outside of the training distribution.

We propose neural execution engines (NEEs), which leverage a learned mask and supervised execution traces to mimic the functionality of subroutines and demonstrate strong generalization.

Meta-learning is a promising strategy for learning to efficiently learn within new tasks, using data gathered from a distribution of tasks. However, the meta-learning literature thus far has focused on the task segmented setting, where at train-time, offline data is assumed to be split according to the underlying task, and at test-time, the algorithms are optimized to learn in a single task. In this work, we enable the application of generic meta-learning algorithms to settings where this task segmentation is unavailable, such as continual online learning with a time-varying task. We present meta-learning via online changepoint analysis (MOCA), an approach which augments a meta-learning algorithm with a differentiable Bayesian changepoint detection scheme. The framework allows both training and testing directly on time series data without segmenting it into discrete tasks. We demonstrate the utility of this approach on a nonlinear meta-regression benchmark as well as two meta-image-classification benchmarks.

Bayesian changepoint detection enables meta-learning directly from time series data.

People with high-frequency hearing loss rely on hearing aids that employ frequency lowering algorithms. These algorithms shift some of the sounds from the high frequency band to the lower frequency band where the sounds become more perceptible for the people with the condition. Fricative phonemes have an important part of their content concentrated in high frequency bands. It is important that the frequency lowering algorithm is activated exactly for the duration of a fricative phoneme, and kept off at all other times. Therefore, timely (with zero delay) and accurate fricative phoneme detection is a key problem for high quality hearing aids. In this paper we present a deep learning based fricative phoneme detection algorithm that has zero detection delay and achieves state-of-the-art fricative phoneme detection accuracy on the TIMIT Speech Corpus. All reported results are reproducible and come with easy to use code that could serve as a baseline for future research.

A deep learning based approach for zero delay fricative phoneme detection

Sequence-to-sequence models with soft attention have been successfully applied to a wide variety of problems, but their decoding process incurs a quadratic time and space cost and is inapplicable to real-time sequence transduction. To address these issues, we propose Monotonic Chunkwise Attention (MoChA), which adaptively splits the input sequence into small chunks over which soft attention is computed. We show that models utilizing MoChA can be trained efficiently with standard backpropagation while allowing online and linear-time decoding at test time. When applied to online speech recognition, we obtain state-of-the-art results and match the performance of a model using an offline soft attention mechanism. In document summarization experiments where we do not expect monotonic alignments, we show significantly improved performance compared to a baseline monotonic attention-based model.

An online and linear-time attention mechanism that performs soft attention over adaptively-located chunks of the input sequence.

We present a framework for automatically ordering image patches that enables in-depth analysis of dataset relationship to learnability of a classification task using convolutional neural network. An image patch is a group of pixels residing in a continuous area contained in the sample. Our preliminary experimental results show that an informed smart shuffling of patches at a sample level can expedite training by exposing important features at early stages of training. In addition, we conduct systematic experiments and provide evidence that CNN’s generalization capabilities do not correlate with human recognizable features present in training samples. We utilized the framework not only to show that spatial locality of features within samples do not correlate with generalization, but also to expedite convergence while achieving similar generalization performance. Using multiple network architectures and datasets, we show that ordering image regions using mutual information measure between adjacent patches, enables CNNs to converge in a third of the total steps required to train the same network without patch ordering.

Develop new techniques that rely on patch reordering to enable detailed analysis of data-set relationship to training and generalization performances.

Producing agents that can generalize to a wide range of environments is a significant challenge in reinforcement learning. One method for overcoming this issue is domain randomization, whereby at the start of each training episode some parameters of the environment are randomized so that the agent is exposed to many possible variations. However, domain randomization is highly inefficient and may lead to policies with high variance across domains. In this work, we formalize the domain randomization problem, and show that minimizing the policy's Lipschitz constant with respect to the randomization parameters leads to low variance in the learned policies. We propose a method where the agent only needs to be trained on one variation of the environment, and its learned state representations are regularized during training to minimize this constant. We conduct experiments that demonstrate that our technique leads to more efficient and robust learning than standard domain randomization, while achieving equal generalization scores.

We produce reinforcement learning agents that generalize well to a wide range of environments using a novel regularization technique.

Claims from the fields of network neuroscience and connectomics suggest that topological models of the brain involving complex networks are of particular use and interest. The field of deep neural networks has mostly left inspiration from these claims out. In this paper, we propose three architectures and use each of them to explore the intersection of network neuroscience and deep learning in an attempt to bridge the gap between the two fields. Using the teachings from network neuroscience and connectomics, we show improvements over the ResNet architecture, we show a possible connection between early training and the spectral properties of the network, and we show the trainability of a DNN based on the neuronal network of C.Elegans.

We explore the intersection of network neurosciences and deep learning.

Creating a knowledge base that is accurate, up-to-date and complete remains a significant challenge despite substantial efforts in automated knowledge base construction. In this paper, we present Alexandria -- a system for unsupervised, high-precision knowledge base construction. Alexandria uses a probabilistic program to define a process of converting knowledge base facts into unstructured text. Using probabilistic inference, we can invert this program and so retrieve facts, schemas and entities from web text. The use of a probabilistic program allows uncertainty in the text to be propagated through to the retrieved facts, which increases accuracy and helps merge facts from multiple sources. Because Alexandria does not require labelled training data, knowledge bases can be constructed with the minimum of manual input. We demonstrate this by constructing a high precision (typically 97\%+) knowledge base for people from a single seed fact.

This paper presents a system for unsupervised, high-precision knowledge base construction using a probabilistic program to define a process of converting knowledge base facts into unstructured text.

Recent advances have made it possible to create deep complex-valued neural networks. Despite this progress, many challenging learning tasks have yet to leverage the power of complex representations. Building on recent advances, we propose a new deep complex-valued method for signal retrieval and extraction in the frequency domain. As a case study, we perform audio source separation in the Fourier domain. Our new method takes advantage of the convolution theorem which states that the Fourier transform of two convolved signals is the elementwise product of their Fourier transforms. Our novel method is based on a complex-valued version of Feature-Wise Linear Modulation (FiLM) and serves as the keystone of our proposed signal extraction method. We also introduce a new and explicit amplitude and phase-aware loss, which is scale and time invariant, taking into account the complex-valued components of the spectrogram. Using the Wall Street Journal Dataset, we compared our phase-aware loss to several others that operate both in the time and frequency domains and demonstrate the effectiveness of our proposed signal extraction method and proposed loss.

New Signal Extraction Method in the Fourier Domain

We propose an implementation of GNN that predicts and imitates the motion be- haviors from observed swarm trajectory data. The network’s ability to capture interaction dynamics in swarms is demonstrated through transfer learning. We finally discuss the inherent availability and challenges in the scalability of GNN, and proposed a method to improve it with layer-wise tuning and mixing of data enabled by padding.

Improve the scalability of graph neural networks on imitation learning and prediction of swarm motion

Embedding layers are commonly used to map discrete symbols into continuous embedding vectors that reflect their semantic meanings. Despite their effectiveness, the number of parameters in an embedding layer increases linearly with the number of symbols and poses a critical challenge on memory and storage constraints. In this work, we propose a generic and end-to-end learnable compression framework termed differentiable product quantization (DPQ). We present two instantiations of DPQ that leverage different approximation techniques to enable differentiability in end-to-end learning. Our method can readily serve as a drop-in alternative for any existing embedding layer. Empirically, DPQ offers significant compression ratios (14-238x) at negligible or no performance cost on 10 datasets across three different language tasks.

We propose a differentiable product quantization framework that can reduce the size of embedding layer in an end-to-end training at no performance cost.

For multi-valued functions---such as when the conditional distribution on targets given the inputs is multi-modal---standard regression approaches are not always desirable because they provide the conditional mean. Modal regression approaches aim to instead find the conditional mode, but are restricted to nonparametric approaches. Such approaches can be difficult to scale, and make it difficult to benefit from parametric function approximation, like neural networks, which can learn complex relationships between inputs and targets. In this work, we propose a parametric modal regression algorithm, by using the implicit function theorem to develop an objective for learning a joint parameterized function over inputs and targets. We empirically demonstrate on several synthetic problems that our method (i) can learn multi-valued functions and produce the conditional modes, (ii) scales well to high-dimensional inputs and (iii) is even more effective for certain unimodal problems, particularly for high frequency data where the joint function over inputs and targets can better capture the complex relationship between them. We conclude by showing that our method provides small improvements on two regression datasets that have asymmetric distributions over the targets.

We introduce a simple and novel modal regression algorithm which is easy to scale to large problems.

Deep reinforcement learning algorithms require large amounts of experience to learn an individual task. While in principle meta-reinforcement learning (meta-RL) algorithms enable agents to learn new skills from small amounts of experience, several major challenges preclude their practicality. Current methods rely heavily on on-policy experience, limiting their sample efficiency. They also lack mechanisms to reason about task uncertainty when adapting to new tasks, limiting their effectiveness in sparse reward problems. In this paper, we address these challenges by developing an off-policy meta-RL algorithm that disentangles task inference and control. In our approach, we perform online probabilistic filtering of latent task variables to infer how to solve a new task from small amounts of experience. This probabilistic interpretation enables posterior sampling for structured and efficient exploration. We demonstrate how to integrate these task variables with off-policy RL algorithms to achieve both meta-training and adaptation efficiency. Our method outperforms prior algorithms in sample efficiency by 20-100X as well as in asymptotic performance on several meta-RL benchmarks.

Sample efficient meta-RL by combining variational inference of probabilistic task variables with off-policy RL

Knowledge bases, massive collections of facts (RDF triples) on diverse topics, support vital modern applications. However, existing knowledge bases contain very little data compared to the wealth of information on the Web. This is because the industry standard in knowledge base creation and augmentation suffers from a serious bottleneck: they rely on domain experts to identify appropriate web sources to extract data from. Efforts to fully automate knowledge extraction have failed to improve this standard: these automated systems are able to retrieve much more data and from a broader range of sources, but they suffer from very low precision and recall. As a result, these large-scale extractions remain unexploited. In this paper, we present MIDAS, a system that harnesses the results of automated knowledge extraction pipelines to repair the bottleneck in industrial knowledge creation and augmentation processes. MIDAS automates the suggestion of good-quality web sources and describes what to extract with respect to augmenting an existing knowledge base. We make three major contributions. First, we introduce a novel concept, web source slices, to describe the contents of a web source. Second, we define a profit function to quantify the value of a web source slice with respect to augmenting an existing knowledge base. Third, we develop effective and highly-scalable algorithms to derive high-profit web source slices. We demonstrate that MIDAS produces high-profit results and outperforms the baselines significantly on both real-word and synthetic datasets.

This paper focuses on identifying high quality web sources for industrial knowledge base augmentation pipeline.

We explore the match prediction problem where one seeks to estimate the likelihood of a group of M items preferred over another, based on partial group comparison data. Challenges arise in practice. As existing state-of-the-art algorithms are tailored to certain statistical models, we have different best algorithms across distinct scenarios. Worse yet, we have no prior knowledge on the underlying model for a given scenario. These call for a unified approach that can be universally applied to a wide range of scenarios and achieve consistently high performances. To this end, we incorporate deep learning architectures so as to reflect the key structural features that most state-of-the-art algorithms, some of which are optimal in certain settings, share in common. This enables us to infer hidden models underlying a given dataset, which govern in-group interactions and statistical patterns of comparisons, and hence to devise the best algorithm tailored to the dataset at hand. Through extensive experiments on synthetic and real-world datasets, we evaluate our framework in comparison to state-of-the-art algorithms. It turns out that our framework consistently leads to the best performance across all datasets in terms of cross entropy loss and prediction accuracy, while the state-of-the-art algorithms suffer from inconsistent performances across different datasets. Furthermore, we show that it can be easily extended to attain satisfactory performances in rank aggregation tasks, suggesting that it can be adaptable for other tasks as well.

We investigate the merits of employing neural networks in the match prediction problem where one seeks to estimate the likelihood of a group of M items preferred over another, based on partial group comparison data.

Recurrent Neural Networks (RNNs) are designed to handle sequential data but suffer from vanishing or exploding gradients. Recent work on Unitary Recurrent Neural Networks (uRNNs) have been used to address this issue and in some cases, exceed the capabilities of Long Short-Term Memory networks (LSTMs). We propose a simpler and novel update scheme to maintain orthogonal recurrent weight matrices without using complex valued matrices. This is done by parametrizing with a skew-symmetric matrix using the Cayley transform. Such a parametrization is unable to represent matrices with negative one eigenvalues, but this limitation is overcome by scaling the recurrent weight matrix by a diagonal matrix consisting of ones and negative ones. The proposed training scheme involves a straightforward gradient calculation and update step. In several experiments, the proposed scaled Cayley orthogonal recurrent neural network (scoRNN) achieves superior results with fewer trainable parameters than other unitary RNNs.

A novel approach to maintain orthogonal recurrent weight matrices in a RNN.

A large number of natural language processing tasks exist to analyze syntax, semantics, and information content of human language. These seemingly very different tasks are usually solved by specially designed architectures. In this paper, we provide the simple insight that a great variety of tasks can be represented in a single unified format consisting of labeling spans and relations between spans, thus a single task-independent model can be used across different tasks. We perform extensive experiments to test this insight on 10 disparate tasks as broad as dependency parsing (syntax), semantic role labeling (semantics), relation extraction (information content), aspect based sentiment analysis (sentiment), and many others, achieving comparable performance as state-of-the-art specialized models. We further demonstrate benefits in multi-task learning. We convert these datasets into a unified format to build a benchmark, which provides a holistic testbed for evaluating future models for generalized natural language analysis.

We use a single model to solve a great variety of natural language analysis tasks by formulating them in a unified span-relation format.

Large matrix inversions have often been cited as a major impediment to scaling Gaussian process (GP) models. With the use of GPs as building blocks for ever more sophisticated Bayesian deep learning models, removing these impediments is a necessary step for achieving large scale results. We present a variational approximation for a wide range of GP models that does not require a matrix inverse to be performed at each optimisation step. Our bound instead directly parameterises a free matrix, which is an additional variational parameter. At the local maxima of the bound, this matrix is equal to the matrix inverse. We prove that our bound gives the same guarantees as earlier variational approximations. We demonstrate some beneficial properties of the bound experimentally, although significant wall clock time speed improvements will require future improvements in optimisation and implementation.

We present a variational lower bound for GP models that can be optimised without computing expensive matrix operations like inverses, while providing the same guarantees as existing variational approximations.

It has been shown that using geometric spaces with non-zero curvature instead of plain Euclidean spaces with zero curvature improves performance on a range of Machine Learning tasks for learning representations. Recent work has leveraged these geometries to learn latent variable models like Variational Autoencoders (VAEs) in spherical and hyperbolic spaces with constant curvature. While these approaches work well on particular kinds of data that they were designed for e.g.~tree-like data for a hyperbolic VAE, there exists no generic approach unifying all three models. We develop a Mixed-curvature Variational Autoencoder, an efficient way to train a VAE whose latent space is a product of constant curvature Riemannian manifolds, where the per-component curvature can be learned. This generalizes the Euclidean VAE to curved latent spaces, as the model essentially reduces to the Euclidean VAE if curvatures of all latent space components go to 0.

Variational Autoencoders with latent spaces modeled as products of constant curvature Riemannian manifolds improve on image reconstruction over single-manifold variants.

Machine learning algorithms for generating molecular structures offer a promising new approach to drug discovery. We cast molecular optimization as a translation problem, where the goal is to map an input compound to a target compound with improved biochemical properties. Remarkably, we observe that when generated molecules are iteratively fed back into the translator, molecular compound attributes improve with each step. We show that this finding is invariant to the choice of translation model, making this a "black box" algorithm. We call this method Black Box Recursive Translation (BBRT), a new inference method for molecular property optimization. This simple, powerful technique operates strictly on the inputs and outputs of any translation model. We obtain new state-of-the-art results for molecular property optimization tasks using our simple drop-in replacement with well-known sequence and graph-based models. Our method provides a significant boost in performance relative to its non-recursive peers with just a simple "``for" loop. Further, BBRT is highly interpretable, allowing users to map the evolution of newly discovered compounds from known starting points.

We introduce a black box algorithm for repeated optimization of compounds using a translation framework.

Deep Neural Networks (DNNs) are increasingly deployed in cloud servers and autonomous agents due to their superior performance. The deployed DNN is either leveraged in a white-box setting (model internals are publicly known) or a black-box setting (only model outputs are known) depending on the application. A practical concern in the rush to adopt DNNs is protecting the models against Intellectual Property (IP) infringement. We propose BlackMarks, the first end-to-end multi-bit watermarking framework that is applicable in the black-box scenario. BlackMarks takes the pre-trained unmarked model and the owner’s binary signature as inputs. The output is the corresponding marked model with specific keys that can be later used to trigger the embedded watermark. To do so, BlackMarks first designs a model-dependent encoding scheme that maps all possible classes in the task to bit ‘0’ and bit ‘1’. Given the owner’s watermark signature (a binary string), a set of key image and label pairs is designed using targeted adversarial attacks. The watermark (WM) is then encoded in the distribution of output activations of the DNN by fine-tuning the model with a WM-specific regularized loss. To extract the WM, BlackMarks queries the model with the WM key images and decodes the owner’s signature from the corresponding predictions using the designed encoding scheme. We perform a comprehensive evaluation of BlackMarks’ performance on MNIST, CIFAR-10, ImageNet datasets and corroborate its effectiveness and robustness. BlackMarks preserves the functionality of the original DNN and incurs negligible WM embedding overhead as low as 2.054%.

Proposing the first watermarking framework for multi-bit signature embedding and extraction using the outputs of the DNN.

Adversarial training provides a principled approach for training robust neural networks. From an optimization perspective, the adversarial training is essentially solving a minmax robust optimization problem. The outer minimization is trying to learn a robust classifier, while the inner maximization is trying to generate adversarial samples. Unfortunately, such a minmax problem is very difficult to solve due to the lack of convex-concave structure. This work proposes a new adversarial training method based on a general learning-to-learn framework. Specifically, instead of applying the existing hand-design algorithms for the inner problem, we learn an optimizer, which is parametrized as a convolutional neural network. At the same time, a robust classifier is learned to defense the adversarial attack generated by the learned optimizer. From the perspective of generative learning, our proposed method can be viewed as learning a deep generative model for generating adversarial samples, which is adaptive to the robust classification. Our experiments demonstrate that our proposed method significantly outperforms existing adversarial training methods on CIFAR-10 and CIFAR-100 datasets.

Don't know how to optimize? Then just learn to optimize!

In this work we introduce a new framework for performing temporal predictions in the presence of uncertainty. It is based on a simple idea of disentangling com- ponents of the future state which are predictable from those which are inherently unpredictable, and encoding the unpredictable components into a low-dimensional latent variable which is fed into the forward model. Our method uses a simple su- pervised training objective which is fast and easy to train. We evaluate it in the context of video prediction on multiple datasets and show that it is able to consi- tently generate diverse predictions without the need for alternating minimization over a latent space or adversarial training.

A simple and easy to train method for multimodal prediction in time series.

Conducting reinforcement-learning experiments can be a complex and timely process. A full experimental pipeline will typically consist of a simulation of an environment, an implementation of one or many learning algorithms, a variety of additional components designed to facilitate the agent-environment interplay, and any requisite analysis, plotting, and logging thereof. In light of this complexity, this paper introduces simple rl, a new open source library for carrying out reinforcement learning experiments in Python 2 and 3 with a focus on simplicity. The goal of simple_rl is to support seamless, reproducible methods for running reinforcement learning experiments. This paper gives an overview of the core design philosophy of the package, how it differs from existing libraries, and showcases its central features.

This paper introduces and motivates simple_rl, a new open source library for carrying out reinforcement learning experiments in Python 2 and 3 with a focus on simplicity.

Wasserstein GAN(WGAN) is a model that minimizes the Wasserstein distance between a data distribution and sample distribution. Recent studies have proposed stabilizing the training process for the WGAN and implementing the Lipschitz constraint. In this study, we prove the local stability of optimizing the simple gradient penalty $\mu$-WGAN(SGP $\mu$-WGAN) under suitable assumptions regarding the equilibrium and penalty measure $\mu$. The measure valued differentiation concept is employed to deal with the derivative of the penalty terms, which is helpful for handling abstract singular measures with lower dimensional support. Based on this analysis, we claim that penalizing the data manifold or sample manifold is the key to regularizing the original WGAN with a gradient penalty. Experimental results obtained with unintuitive penalty measures that satisfy our assumptions are also provided to support our theoretical results.

This paper deals with stability of simple gradient penalty $\mu$-WGAN optimization by introducing a concept of measure valued differentiation.

We present Random Partition Relaxation (RPR), a method for strong quantization of the parameters of convolutional neural networks to binary (+1/-1) and ternary (+1/0/-1) values. Starting from a pretrained model, we first quantize the weights and then relax random partitions of them to their continuous values for retraining before quantizing them again and switching to another weight partition for further adaptation. We empirically evaluate the performance of RPR with ResNet-18, ResNet-50 and GoogLeNet on the ImageNet classification task for binary and ternary weight networks. We show accuracies beyond the state-of-the-art for binary- and ternary-weight GoogLeNet and competitive performance for ResNet-18 and ResNet-50 using a SGD-based training method that can easily be integrated into existing frameworks.

State-of-the-art training method for binary and ternary weight networks based on alternating optimization of randomly relaxed weight partitions

Learning long-term dependencies is a key long-standing challenge of recurrent neural networks (RNNs). Hierarchical recurrent neural networks (HRNNs) have been considered a promising approach as long-term dependencies are resolved through shortcuts up and down the hierarchy. Yet, the memory requirements of Truncated Backpropagation Through Time (TBPTT) still prevent training them on very long sequences. In this paper, we empirically show that in (deep) HRNNs, propagating gradients back from higher to lower levels can be replaced by locally computable losses, without harming the learning capability of the network, over a wide range of tasks. This decoupling by local losses reduces the memory requirements of training by a factor exponential in the depth of the hierarchy in comparison to standard TBPTT.

We replace some gradients paths in hierarchical RNN's by an auxiliary loss. We show that this can reduce the memory cost while preserving performance.

In a typical deep learning approach to a computer vision task, Convolutional Neural Networks (CNNs) are used to extract features at varying levels of abstraction from an image and compress a high dimensional input into a lower dimensional decision space through a series of transformations. In this paper, we investigate how a class of input images is eventually compressed over the course of these transformations. In particular, we use singular value decomposition to analyze the relevant variations in feature space. These variations are formalized as the effective dimension of the embedding. We consider how the effective dimension varies across layers within class. We show that across datasets and architectures, the effective dimension of a class increases before decreasing further into the network, suggesting some sort of initial whitening transformation. Further, the decrease rate of the effective dimension deeper in the network corresponds with training performance of the model.

Neural networks that do a good job of classification project points into more spherical shapes before compressing them into fewer dimensions.

Deep learning methods have achieved high performance in sound recognition tasks. Deciding how to feed the training data is important for further performance improvement. We propose a novel learning method for deep sound recognition: Between-Class learning (BC learning). Our strategy is to learn a discriminative feature space by recognizing the between-class sounds as between-class sounds. We generate between-class sounds by mixing two sounds belonging to different classes with a random ratio. We then input the mixed sound to the model and train the model to output the mixing ratio. The advantages of BC learning are not limited only to the increase in variation of the training data; BC learning leads to an enlargement of Fisher’s criterion in the feature space and a regularization of the positional relationship among the feature distributions of the classes. The experimental results show that BC learning improves the performance on various sound recognition networks, datasets, and data augmentation schemes, in which BC learning proves to be always beneficial. Furthermore, we construct a new deep sound recognition network (EnvNet-v2) and train it with BC learning. As a result, we achieved a performance surpasses the human level.

We propose an novel learning method for deep sound recognition named BC learning.

Spatiotemporal forecasting has become an increasingly important prediction task in machine learning and statistics due to its vast applications, such as climate modeling, traffic prediction, video caching predictions, and so on. While numerous studies have been conducted, most existing works assume that the data from different sources or across different locations are equally reliable. Due to cost, accessibility, or other factors, it is inevitable that the data quality could vary, which introduces significant biases into the model and leads to unreliable prediction results. The problem could be exacerbated in black-box prediction models, such as deep neural networks. In this paper, we propose a novel solution that can automatically infer data quality levels of different sources through local variations of spatiotemporal signals without explicit labels. Furthermore, we integrate the estimate of data quality level with graph convolutional networks to exploit their efficient structures. We evaluate our proposed method on forecasting temperatures in Los Angeles.

We propose a method that infers the time-varying data quality level for spatiotemporal forecasting without explicitly assigned labels.

Human perception of 3D shapes goes beyond reconstructing them as a set of points or a composition of geometric primitives: we also effortlessly understand higher-level shape structure such as the repetition and reflective symmetry of object parts. In contrast, recent advances in 3D shape sensing focus more on low-level geometry but less on these higher-level relationships. In this paper, we propose 3D shape programs, integrating bottom-up recognition systems with top-down, symbolic program structure to capture both low-level geometry and high-level structural priors for 3D shapes. Because there are no annotations of shape programs for real shapes, we develop neural modules that not only learn to infer 3D shape programs from raw, unannotated shapes, but also to execute these programs for shape reconstruction. After initial bootstrapping, our end-to-end differentiable model learns 3D shape programs by reconstructing shapes in a self-supervised manner. Experiments demonstrate that our model accurately infers and executes 3D shape programs for highly complex shapes from various categories. It can also be integrated with an image-to-shape module to infer 3D shape programs directly from an RGB image, leading to 3D shape reconstructions that are both more accurate and more physically plausible.

We propose 3D shape programs, a structured, compositional shape representation. Our model learns to infer and execute shape programs to explain 3D shapes.

Deep Reinforcement Learning (Deep RL) has been receiving increasingly more attention thanks to its encouraging performance on a variety of control tasks. Yet, conventional regularization techniques in training neural networks (e.g., $L_2$ regularization, dropout) have been largely ignored in RL methods, possibly because agents are typically trained and evaluated in the same environment. In this work, we present the first comprehensive study of regularization techniques with multiple policy optimization algorithms on continuous control tasks. Interestingly, we find conventional regularization techniques on the policy networks can often bring large improvement on the task performance, and the improvement is typically more significant when the task is more difficult. We also compare with the widely used entropy regularization and find $L_2$ regularization is generally better. Our findings are further confirmed to be robust against the choice of training hyperparameters. We also study the effects of regularizing different components and find that only regularizing the policy network is typically enough. We hope our study provides guidance for future practices in regularizing policy optimization algorithms.

We show that conventional regularization methods (e.g., $L_2$, dropout), which have been largely ignored in RL methods, can be very effective in policy optimization.

We introduce FigureQA, a visual reasoning corpus of over one million question-answer pairs grounded in over 100,000 images. The images are synthetic, scientific-style figures from five classes: line plots, dot-line plots, vertical and horizontal bar graphs, and pie charts. We formulate our reasoning task by generating questions from 15 templates; questions concern various relationships between plot elements and examine characteristics like the maximum, the minimum, area-under-the-curve, smoothness, and intersection. To resolve, such questions often require reference to multiple plot elements and synthesis of information distributed spatially throughout a figure. To facilitate the training of machine learning systems, the corpus also includes side data that can be used to formulate auxiliary objectives. In particular, we provide the numerical data used to generate each figure as well as bounding-box annotations for all plot elements. We study the proposed visual reasoning task by training several models, including the recently proposed Relation Network as strong baseline. Preliminary results indicate that the task poses a significant machine learning challenge. We envision FigureQA as a first step towards developing models that can intuitively recognize patterns from visual representations of data.

We present a question-answering dataset, FigureQA, as a first step towards developing models that can intuitively recognize patterns from visual representations of data.

In this paper, I discuss some varieties of explanation that can arise in intelligent agents. I distinguish between process accounts, which address the detailed decisions made during heuristic search, and preference accounts, which clarify the ordering of alternatives independent of how they were generated. I also hypothesize which types of users will appreciate which types of explanation. In addition, I discuss three facets of multi-step decision making -- conceptual inference, plan generation, and plan execution -- in which explanations can arise. I also consider alternative ways to present questions to agents and for them provide their answers.

This position paper analyzes different types of self explanation that can arise in planning and related systems.

Generative deep learning has sparked a new wave of Super-Resolution (SR) algorithms that enhance single images with impressive aesthetic results, albeit with imaginary details. Multi-frame Super-Resolution (MFSR) offers a more grounded approach to the ill-posed problem, by conditioning on multiple low-resolution views. This is important for satellite monitoring of human impact on the planet -- from deforestation, to human rights violations -- that depend on reliable imagery. To this end, we present HighRes-net, the first deep learning approach to MFSR that learns its sub-tasks in an end-to-end fashion: (i) co-registration, (ii) fusion, (iii) up-sampling, and (iv) registration-at-the-loss. Co-registration of low-res views is learned implicitly through a reference-frame channel, with no explicit registration mechanism. We learn a global fusion operator that is applied recursively on an arbitrary number of low-res pairs. We introduce a registered loss, by learning to align the SR output to a ground-truth through ShiftNet. We show that by learning deep representations of multiple views, we can super-resolve low-resolution signals and enhance Earth observation data at scale. Our approach recently topped the European Space Agency's MFSR competition on real-world satellite imagery.

The first deep learning approach to MFSR to solve registration, fusion, up-sampling in an end-to-end manner.

Large mini-batch parallel SGD is commonly used for distributed training of deep networks. Approaches that use tightly-coupled exact distributed averaging based on AllReduce are sensitive to slow nodes and high-latency communication. In this work we show the applicability of Stochastic Gradient Push (SGP) for distributed training. SGP uses a gossip algorithm called PushSum for approximate distributed averaging, allowing for much more loosely coupled communications which can be beneficial in high-latency or high-variability scenarios. The tradeoff is that approximate distributed averaging injects additional noise in the gradient which can affect the train and test accuracies. We prove that SGP converges to a stationary point of smooth, non-convex objective functions. Furthermore, we validate empirically the potential of SGP. For example, using 32 nodes with 8 GPUs per node to train ResNet-50 on ImageNet, where nodes communicate over 10Gbps Ethernet, SGP completes 90 epochs in around 1.5 hours while AllReduce SGD takes over 5 hours, and the top-1 validation accuracy of SGP remains within 1.2% of that obtained using AllReduce SGD.

For distributed training over high-latency networks, use gossip-based approximate distributed averaging instead of exact distribute averaging like AllReduce.

In this paper, we extend the persona-based sequence-to-sequence (Seq2Seq) neural network conversation model to a multi-turn dialogue scenario by modifying the state-of-the-art hredGAN architecture to simultaneously capture utterance attributes such as speaker identity, dialogue topic, speaker sentiments and so on. The proposed system, phredGAN has a persona-based HRED generator (PHRED) and a conditional discriminator. We also explore two approaches to accomplish the conditional discriminator: (1) $phredGAN_a$, a system that passes the attribute representation as an additional input into a traditional adversarial discriminator, and (2) $phredGAN_d$, a dual discriminator system which in addition to the adversarial discriminator, collaboratively predicts the attribute(s) that generated the input utterance. To demonstrate the superior performance of phredGAN over the persona SeqSeq model, we experiment with two conversational datasets, the Ubuntu Dialogue Corpus (UDC) and TV series transcripts from the Big Bang Theory and Friends. Performance comparison is made with respect to a variety of quantitative measures as well as crowd-sourced human evaluation. We also explore the trade-offs from using either variant of $phredGAN$ on datasets with many but weak attribute modalities (such as with Big Bang Theory and Friends) and ones with few but strong attribute modalities (customer-agent interactions in Ubuntu dataset).

This paper develops an adversarial learning framework for neural conversation models with persona

We introduce bio-inspired artificial neural networks consisting of neurons that are additionally characterized by spatial positions. To simulate properties of biological systems we add the costs penalizing long connections and the proximity of neurons in a two-dimensional space. Our experiments show that in the case where the network performs two different tasks, the neurons naturally split into clusters, where each cluster is responsible for processing a different task. This behavior not only corresponds to the biological systems, but also allows for further insight into interpretability or continual learning.

Bio-inspired artificial neural networks, consisting of neurons positioned in a two-dimensional space, are capable of forming independent groups for performing different tasks.

The transformer has become a central model for many NLP tasks from translation to language modeling to representation learning. Its success demonstrates the effectiveness of stacked attention as a replacement for recurrence for many tasks. In theory attention also offers more insights into the model’s internal decisions; however, in practice when stacked it quickly becomes nearly as fully-connected as recurrent models. In this work, we propose an alternative transformer architecture, discrete transformer, with the goal of better separating out internal model decisions. The model uses hard attention to ensure that each step only depends on a fixed context. Additionally, the model uses a separate “syntactic” controller to separate out network structure from decision making. Finally we show that this approach can be further sparsified with direct regularization. Empirically, this approach is able to maintain the same level of performance on several datasets, while discretizing reasoning decisions over the data.

Discrete transformer which uses hard attention to ensure that each step only depends on a fixed context.

Deep predictive coding networks are neuroscience-inspired unsupervised learning models that learn to predict future sensory states. We build upon the PredNet implementation by Lotter, Kreiman, and Cox (2016) to investigate if predictive coding representations are useful to predict brain activity in the visual cortex. We use representational similarity analysis (RSA) to compare PredNet representations to functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG) data from the Algonauts Project (Cichy et al., 2019). In contrast to previous findings in the literature (Khaligh-Razavi & Kriegeskorte, 2014), we report empirical data suggesting that unsupervised models trained to predict frames of videos without further fine-tuning may outperform supervised image classification baselines in terms of correlation to spatial (fMRI) and temporal (MEG) data.

We show empirical evidence that predictive coding models yield representations more correlated to brain data than supervised image recognition models.

The incorporation of prior knowledge into learning is essential in achieving good performance based on small noisy samples. Such knowledge is often incorporated through the availability of related data arising from domains and tasks similar to the one of current interest. Ideally one would like to allow both the data for the current task and for previous related tasks to self-organize the learning system in such a way that commonalities and differences between the tasks are learned in a data-driven fashion. We develop a framework for learning multiple tasks simultaneously, based on sharing features that are common to all tasks, achieved through the use of a modular deep feedforward neural network consisting of shared branches, dealing with the common features of all tasks, and private branches, learning the specific unique aspects of each task. Once an appropriate weight sharing architecture has been established, learning takes place through standard algorithms for feedforward networks, e.g., stochastic gradient descent and its variations. The method deals with meta-learning (such as domain adaptation, transfer and multi-task learning) in a unified fashion, and can easily deal with data arising from different types of sources. Numerical experiments demonstrate the effectiveness of learning in domain adaptation and transfer learning setups, and provide evidence for the flexible and task-oriented representations arising in the network.

A generic framework for handling transfer and multi-task learning using pairs of autoencoders with task-specific and shared weights.

Deep neural networks and decision trees operate on largely separate paradigms; typically, the former performs representation learning with pre-specified architectures, while the latter is characterised by learning hierarchies over pre-specified features with data-driven architectures. We unite the two via adaptive neural trees (ANTs), a model that incorporates representation learning into edges, routing functions and leaf nodes of a decision tree, along with a backpropagation-based training algorithm that adaptively grows the architecture from primitive modules (e.g., convolutional layers). ANTs allow increased interpretability via hierarchical clustering, e.g., learning meaningful class associations, such as separating natural vs. man-made objects. We demonstrate this on classification and regression tasks, achieving over 99% and 90% accuracy on the MNIST and CIFAR-10 datasets, and outperforming standard neural networks, random forests and gradient boosted trees on the SARCOS dataset. Furthermore, ANT optimisation naturally adapts the architecture to the size and complexity of the training data.

We propose a framework to combine decision trees and neural networks, and show on image classification tasks that it enjoys the complementary benefits of the two approaches, while addressing the limitations of prior work.

While natural language processing systems often focus on a single language, multilingual transfer learning has the potential to improve performance, especially for low-resource languages. We introduce XLDA, cross-lingual data augmentation, a method that replaces a segment of the input text with its translation in another language. XLDA enhances performance of all 14 tested languages of the cross-lingual natural language inference (XNLI) benchmark. With improvements of up to 4.8, training with XLDA achieves state-of-the-art performance for Greek, Turkish, and Urdu. XLDA is in contrast to, and performs markedly better than, a more naive approach that aggregates examples in various languages in a way that each example is solely in one language. On the SQuAD question answering task, we see that XLDA provides a 1.0 performance increase on the English evaluation set. Comprehensive experiments suggest that most languages are effective as cross-lingual augmentors, that XLDA is robust to a wide range of translation quality, and that XLDA is even more effective for randomly initialized models than for pretrained models.

Translating portions of the input during training can improve cross-lingual performance.

Training conditional generative latent-variable models is challenging in scenarios where the conditioning signal is very strong and the decoder is expressive enough to generate a plausible output given only the condition; the generative model tends to ignore the latent variable, suffering from posterior collapse. We find, and empirically show, that one of the major reasons behind posterior collapse is rooted in the way that generative models are conditioned, i.e., through concatenation of the latent variable and the condition . To mitigate this problem, we propose to explicitly make the latent variables depend on the condition by unifying the conditioning and latent variable sampling, thus coupling them so as to prevent the model from discarding the root of variations . To achieve this, we develop a conditional Variational Autoencoder architecture that learns a distribution not only of the latent variables, but also of the condition, the latter acting as prior on the former . Our experiments on the challenging tasks of conditional human motion prediction and image captioning demonstrate the effectiveness of our approach at avoiding posterior collapse . Video results of our approach are anonymously provided in http://bit.ly/iclr2020

We propose a conditional variational autoencoder framework that mitigates the posterior collapse in scenarios where the conditioning signal strong enough for an expressive decoder to generate a plausible output from it.

We propose a study of the stability of several few-shot learning algorithms subject to variations in the hyper-parameters and optimization schemes while controlling the random seed. We propose a methodology for testing for statistical differences in model performances under several replications. To study this specific design, we attempt to reproduce results from three prominent papers: Matching Nets, Prototypical Networks, and TADAM. We analyze on the miniImagenet dataset on the standard classification task in the 5-ways, 5-shots learning setting at test time. We find that the selected implementations exhibit stability across random seed, and repeats.

We propose a study of the stability of several few-shot learning algorithms subject to variations in the hyper-parameters and optimization schemes while controlling the random seed.

We study the problem of representation learning in goal-conditioned hierarchical reinforcement learning. In such hierarchical structures, a higher-level controller solves tasks by iteratively communicating goals which a lower-level policy is trained to reach. Accordingly, the choice of representation -- the mapping of observation space to goal space -- is crucial. To study this problem, we develop a notion of sub-optimality of a representation, defined in terms of expected reward of the optimal hierarchical policy using this representation. We derive expressions which bound the sub-optimality and show how these expressions can be translated to representation learning objectives which may be optimized in practice. Results on a number of difficult continuous-control tasks show that our approach to representation learning yields qualitatively better representations as well as quantitatively better hierarchical policies, compared to existing methods.

We translate a bound on sub-optimality of representations to a practical training objective in the context of hierarchical reinforcement learning.

Heuristic search research often deals with finding algorithms for offline planning which aim to minimize the number of expanded nodes or planning time. In online planning, algorithms for real-time search or deadline-aware search have been considered before. However, in this paper, we are interested in the problem of {\em situated temporal planning} in which an agent's plan can depend on exogenous events in the external world, and thus it becomes important to take the passage of time into account during the planning process. Previous work on situated temporal planning has proposed simple pruning strategies, as well as complex schemes for a simplified version of the associated metareasoning problem. In this paper, we propose a simple metareasoning technique, called the crude greedy scheme, which can be applied in a situated temporal planner. Our empirical evaluation shows that the crude greedy scheme outperforms standard heuristic search based on cost-to-go estimates.

Metareasoning in a Situated Temporal Planner

Neural networks are vulnerable to small adversarial perturbations. Existing literature largely focused on understanding and mitigating the vulnerability of learned models. In this paper, we demonstrate an intriguing phenomenon about the most popular robust training method in the literature, adversarial training: Adversarial robustness, unlike clean accuracy, is sensitive to the input data distribution. Even a semantics-preserving transformations on the input data distribution can cause a significantly different robustness for the adversarial trained model that is both trained and evaluated on the new distribution. Our discovery of such sensitivity on data distribution is based on a study which disentangles the behaviors of clean accuracy and robust accuracy of the Bayes classifier. Empirical investigations further confirm our finding. We construct semantically-identical variants for MNIST and CIFAR10 respectively, and show that standardly trained models achieve comparable clean accuracies on them, but adversarially trained models achieve significantly different robustness accuracies. This counter-intuitive phenomenon indicates that input data distribution alone can affect the adversarial robustness of trained neural networks, not necessarily the tasks themselves. Lastly, we discuss the practical implications on evaluating adversarial robustness, and make initial attempts to understand this complex phenomenon.

Robustness performance of PGD trained models are sensitive to semantics-preserving transformation of image datasets, which implies the trickiness of evaluation of robust learning algorithms in practice.

Many tasks in natural language processing involve comparing two sentences to compute some notion of relevance, entailment, or similarity. Typically this comparison is done either at the word level or at the sentence level, with no attempt to leverage the inherent structure of the sentence. When sentence structure is used for comparison, it is obtained during a non-differentiable pre-processing step, leading to propagation of errors. We introduce a model of structured alignments between sentences, showing how to compare two sentences by matching their latent structures. Using a structured attention mechanism, our model matches possible spans in the first sentence to possible spans in the second sentence, simultaneously discovering the tree structure of each sentence and performing a comparison, in a model that is fully differentiable and is trained only on the comparison objective. We evaluate this model on two sentence comparison tasks: the Stanford natural language inference dataset and the TREC-QA dataset. We find that comparing spans results in superior performance to comparing words individually, and that the learned trees are consistent with actual linguistic structures.

Matching sentences by learning the latent constituency tree structures with a variant of the inside-outside algorithm embedded as a neural network layer.

Learning disentangled representation from any unlabelled data is a non-trivial problem. In this paper we propose Information Maximising Autoencoder (InfoAE) where the encoder learns powerful disentangled representation through maximizing the mutual information between the representation and given information in an unsupervised fashion. We have evaluated our model on MNIST dataset and achieved approximately 98.9 % test accuracy while using complete unsupervised training.

Learn disentangle representation in an unsupervised manner.

Effective training of neural networks requires much data. In the low-data regime, parameters are underdetermined, and learnt networks generalise poorly. Data Augmentation (Krizhevsky et al., 2012) alleviates this by using existing data more effectively. However standard data augmentation produces only limited plausible alternative data. Given there is potential to generate a much broader set of augmentations, we design and train a generative model to do data augmentation. The model, based on image conditional Generative Adversarial Networks, takes data from a source domain and learns to take any data item and generalise it to generate other within-class data items. As this generative process does not depend on the classes themselves, it can be applied to novel unseen classes of data. We show that a Data Augmentation Generative Adversarial Network (DAGAN) augments standard vanilla classifiers well. We also show a DAGAN can enhance few-shot learning systems such as Matching Networks. We demonstrate these approaches on Omniglot, on EMNIST having learnt the DAGAN on Omniglot, and VGG-Face data. In our experiments we can see over 13% increase in accuracy in the low-data regime experiments in Omniglot (from 69% to 82%), EMNIST (73.9% to 76%) and VGG-Face (4.5% to 12%); in Matching Networks for Omniglot we observe an increase of 0.5% (from 96.9% to 97.4%) and an increase of 1.8% in EMNIST (from 59.5% to 61.3%).

Conditional GANs trained to generate data augmented samples of their conditional inputs used to enhance vanilla classification and one shot learning systems such as matching networks and pixel distance

Answering questions about data can require understanding what parts of an input X influence the response Y. Finding such an understanding can be built by testing relationships between variables through a machine learning model. For example, conditional randomization tests help determine whether a variable relates to the response given the rest of the variables. However, randomization tests require users to specify test statistics. We formalize a class of proper test statistics that are guaranteed to select a feature when it provides information about the response even when the rest of the features are known. We show that f-divergences provide a broad class of proper test statistics. In the class of f-divergences, the KL-divergence yields an easy-to-compute proper test statistic that relates to the AMI. Questions of feature importance can be asked at the level of an individual sample. We show that estimators from the same AMI test can also be used to find important features in a particular instance. We provide an example to show that perfect predictive models are insufficient for instance-wise feature selection. We evaluate our method on several simulation experiments, on a genomic dataset, a clinical dataset for hospital readmission, and on a subset of classes in ImageNet. Our method outperforms several baselines in various simulated datasets, is able to identify biologically significant genes, can select the most important predictors of a hospital readmission event, and is able to identify distinguishing features in an image-classification task.

We develop a simple regression-based model-agnostic feature selection method to interpret data generating processes with FDR control, and outperform several popular baselines on several simulated, medical, and image datasets.

Supervised learning depends on annotated examples, which are taken to be the ground truth. But these labels often come from noisy crowdsourcing platforms, like Amazon Mechanical Turk. Practitioners typically collect multiple labels per example and aggregate the results to mitigate noise (the classic crowdsourcing problem). Given a fixed annotation budget and unlimited unlabeled data, redundant annotation comes at the expense of fewer labeled examples. This raises two fundamental questions: (1) How can we best learn from noisy workers? (2) How should we allocate our labeling budget to maximize the performance of a classifier? We propose a new algorithm for jointly modeling labels and worker quality from noisy crowd-sourced data. The alternating minimization proceeds in rounds, estimating worker quality from disagreement with the current model and then updating the model by optimizing a loss function that accounts for the current estimate of worker quality. Unlike previous approaches, even with only one annotation per example, our algorithm can estimate worker quality. We establish a generalization error bound for models learned with our algorithm and establish theoretically that it's better to label many examples once (vs less multiply) when worker quality exceeds a threshold. Experiments conducted on both ImageNet (with simulated noisy workers) and MS-COCO (using the real crowdsourced labels) confirm our algorithm's benefits.

A new approach for learning a model from noisy crowdsourced annotations.

Neural networks make mistakes. The reason why a mistake is made often remains a mystery. As such neural networks often are considered a black box. It would be useful to have a method that can give an explanation that is intuitive to a user as to why an image is misclassified. In this paper we develop a method for explaining the mistakes of a classifier model by visually showing what must be added to an image such that it is correctly classified. Our work combines the fields of adversarial examples, generative modeling and a correction technique based on difference target propagation to create an technique that creates explanations of why an image is misclassified. In this paper we explain our method and demonstrate it on MNIST and CelebA. This approach could aid in demystifying neural networks for a user.

New way of explaining why a neural network has misclassified an image

In the context of multi-task learning, neural networks with branched architectures have often been employed to jointly tackle the tasks at hand. Such ramified networks typically start with a number of shared layers, after which different tasks branch out into their own sequence of layers. Understandably, as the number of possible network configurations is combinatorially large, deciding what layers to share and where to branch out becomes cumbersome. Prior works have either relied on ad hoc methods to determine the level of layer sharing, which is suboptimal, or utilized neural architecture search techniques to establish the network design, which is considerably expensive. In this paper, we go beyond these limitations and propose a principled approach to automatically construct branched multi-task networks, by leveraging the employed tasks' affinities. Given a specific budget, i.e. number of learnable parameters, the proposed approach generates architectures, in which shallow layers are task-agnostic, whereas deeper ones gradually grow more task-specific. Extensive experimental analysis across numerous, diverse multi-tasking datasets shows that, for a given budget, our method consistently yields networks with the highest performance, while for a certain performance threshold it requires the least amount of learnable parameters.

A method for the automated construction of branched multi-task networks with strong experimental evaluation on diverse multi-tasking datasets.

Typical recent neural network designs are primarily convolutional layers, but the tricks enabling structured efficient linear layers (SELLs) have not yet been adapted to the convolutional setting. We present a method to express the weight tensor in a convolutional layer using diagonal matrices, discrete cosine transforms (DCTs) and permutations that can be optimised using standard stochastic gradient methods. A network composed of such structured efficient convolutional layers (SECL) outperforms existing low-rank networks and demonstrates competitive computational efficiency.

It's possible to substitute the weight matrix in a convolutional layer to train it as a structured efficient layer; performing as well as low-rank decomposition.

Blind document deblurring is a fundamental task in the field of document processing and restoration, having wide enhancement applications in optical character recognition systems, forensics, etc. Since this problem is highly ill-posed, supervised and unsupervised learning methods are well suited for this application. Using various techniques, extensive work has been done on natural-scene deblurring. However, these extracted features are not suitable for document images. We present SVDocNet, an end-to-end trainable U-Net based spatial recurrent neural network (RNN) for blind document deblurring where the weights of the RNNs are determined by different convolutional neural networks (CNNs). This network achieves state of the art performance in terms of both quantitative measures and qualitative results.

We present SVDocNet, an end-to-end trainable U-Net based spatial recurrent neural network (RNN) for blind document deblurring.

In contrast to the monolithic deep architectures used in deep learning today for computer vision, the visual cortex processes retinal images via two functionally distinct but interconnected networks: the ventral pathway for processing object-related information and the dorsal pathway for processing motion and transformations. Inspired by this cortical division of labor and properties of the magno- and parvocellular systems, we explore an unsupervised approach to feature learning that jointly learns object features and their transformations from natural videos. We propose a new convolutional bilinear sparse coding model that (1) allows independent feature transformations and (2) is capable of processing large images. Our learning procedure leverages smooth motion in natural videos. Our results show that our model can learn groups of features and their transformations directly from natural videos in a completely unsupervised manner. The learned "dynamic filters" exhibit certain equivariance properties, resemble cortical spatiotemporal filters, and capture the statistics of transitions between video frames. Our model can be viewed as one of the first approaches to demonstrate unsupervised learning of primary "capsules" (proposed by Hinton and colleagues for supervised learning) and has strong connections to the Lie group approach to visual perception.

We extend bilinear sparse coding and leverage video sequences to learn dynamic filters.

Conventional out-of-distribution (OOD) detection schemes based on variational autoencoder or Random Network Distillation (RND) are known to assign lower uncertainty to the OOD data than the target distribution. In this work, we discover that such conventional novelty detection schemes are also vulnerable to the blurred images. Based on the observation, we construct a novel RND-based OOD detector, SVD-RND, that utilizes blurred images during training. Our detector is simple, efficient in test time, and outperforms baseline OOD detectors in various domains. Further results show that SVD-RND learns a better target distribution representation than the baselines. Finally, SVD-RND combined with geometric transform achieves near-perfect detection accuracy in CelebA domain.

We propose a novel OOD detector that employ blurred images as adversarial examples . Our model achieve significant OOD detection performance in various domains.

Training large deep neural networks on massive datasets is  computationally very challenging. There has been recent surge in interest in using large batch stochastic optimization methods to tackle this issue. The most prominent algorithm in this line of research is LARS, which by  employing layerwise adaptive learning rates trains ResNet on ImageNet in a few minutes. However, LARS performs poorly for attention models like BERT, indicating that its performance gains are not consistent across tasks. In this paper, we first study a principled layerwise adaptation strategy to accelerate training of deep neural networks using large mini-batches. Using this strategy, we develop a new layerwise adaptive large batch optimization technique called LAMB; we then provide convergence analysis of LAMB as well as LARS, showing convergence to a stationary point in general nonconvex settings. Our empirical results demonstrate the superior performance of LAMB across various tasks such as BERT and ResNet-50 training with very little hyperparameter tuning. In particular, for BERT training, our optimizer enables use of very large batch sizes of 32868 without any degradation of performance.   By increasing the batch size to the memory limit of a TPUv3 Pod, BERT training time can be reduced from 3 days to just 76 minutes (Table 1).

A fast optimizer for general applications and large-batch training.

Model-agnostic meta-learning (MAML) is known as a powerful meta-learning method. However, MAML is notorious for being hard to train because of the existence of two learning rates. Therefore, in this paper, we derive the conditions that inner learning rate $\alpha$ and meta-learning rate $\beta$ must satisfy for MAML to converge to minima with some simplifications. We find that the upper bound of $\beta$ depends on $ \alpha$, in contrast to the case of using the normal gradient descent method. Moreover, we show that the threshold of $\beta$ increases as $\alpha$ approaches its own upper bound. This result is verified by experiments on various few-shot tasks and architectures; specifically, we perform sinusoid regression and classification of Omniglot and MiniImagenet datasets with a multilayer perceptron and a convolutional neural network. Based on this outcome, we present a guideline for determining the learning rates: first, search for the largest possible $\alpha$; next, tune $\beta$ based on the chosen value of $\alpha$.

We analyzed the role of two learning rates in model-agnostic meta-learning in convergence.

We present a neural framework for learning associations between interrelated groups of words such as the ones found in Subject-Verb-Object (SVO) structures. Our model induces a joint function-specific word vector space, where vectors of e.g. plausible SVO compositions lie close together. The model retains information about word group membership even in the joint space, and can thereby effectively be applied to a number of tasks reasoning over the SVO structure. We show the robustness and versatility of the proposed framework by reporting state-of-the-art results on the tasks of estimating selectional preference (i.e., thematic fit) and event similarity. The results indicate that the combinations of representations learned with our task-independent model outperform task-specific architectures from prior work, while reducing the number of parameters by up to 95%. The proposed framework is versatile and holds promise to support learning function-specific representations beyond the SVO structures.

Task-independent neural model for learning associations between interrelated groups of words.

The fabrication of semiconductor involves etching process to remove selected areas from wafers. However, the measurement of etched structure in micro-graph heavily relies on time-consuming manual routines. Traditional image processing usually demands on large number of annotated data and the performance is still poor. We treat this challenge as segmentation problem and use deep learning approach to detect masks of objects in etched structure of wafer. Then, we use simple image processing to carry out automatic measurement on the objects. We attempt Generative Adversarial Network (GAN) to generate more data to overcome the problem of very limited dataset. We download 10 SEM (Scanning Electron Microscope) images of 4 types from Internet, based on which we carry out our experiments. Our deep learning based method demonstrates superiority over image processing approach with mean accuracy reaching over 96% for the measurements, compared with the ground truth. To the best of our knowledge, it is the first time that deep learning has been applied in semiconductor industry for automatic measurement.

Using deep learning method to carry out automatic measurement of SEM images in semiconductor industry

Generating and scheduling activities is particularly challenging when considering both consumptive resources and complex resource interactions such as time-dependent resource usage.We present three methods of determining valid temporal placement intervals for an activity in a temporally grounded plan in the presence of such constraints. We introduce the Max Duration and Probe algorithms which are sound, but incomplete, and the Linear algorithm which is sound and complete for linear rate resource consumption. We apply these techniques to the problem of scheduling awakes for a planetary rover where the awake durations are affected by existing activities. We demonstrate how the Probe algorithm performs competitively with the Linear algorithm given an advantageous problem space and well-defined heuristics. We show that the Probe and Linear algorithms outperform the Max Duration algorithm empirically. We then empirically present the runtime differences between the three algorithms. The Probe algorithm is currently base-lined for use in the onboard scheduler for NASA’s next planetary rover, the Mars 2020 rover.

This paper describes and analyzes three methods to schedule non-fixed duration activities in the presence of consumptive resources.