Datasets:

dim

/

scitldr

like 0

Deep latent variable models have become a popular model choice due to the scalable learning algorithms introduced by (Kingma & Welling 2013, Rezende et al. 2014). These approaches maximize a variational lower bound on the intractable log likelihood of the observed data. Burda et al. (2015) introduced a multi-sample variational bound, IWAE, that is at least as tight as the standard variational lower bound and becomes increasingly tight as the number of samples increases. Counterintuitively, the typical inference network gradient estimator for the IWAE bound performs poorly as the number of samples increases (Rainforth et al. 2018, Le et al. 2018). Roeder et a. (2017) propose an improved gradient estimator, however, are unable to show it is unbiased. We show that it is in fact biased and that the bias can be estimated efficiently with a second application of the reparameterization trick. The doubly reparameterized gradient (DReG) estimator does not suffer as the number of samples increases, resolving the previously raised issues. The same idea can be used to improve many recently introduced training techniques for latent variable models. In particular, we show that this estimator reduces the variance of the IWAE gradient, the reweighted wake-sleep update (RWS) (Bornschein & Bengio 2014), and the jackknife variational inference (JVI) gradient (Nowozin 2018). Finally, we show that this computationally efficient, drop-in estimator translates to improved performance for all three objectives on several modeling tasks.

Doubly reparameterized gradient estimators provide unbiased variance reduction which leads to improved performance.

Zeroth-order optimization is the process of minimizing an objective $f(x)$, given oracle access to evaluations at adaptively chosen inputs $x$. In this paper, we present two simple yet powerful GradientLess Descent (GLD) algorithms that do not rely on an underlying gradient estimate and are numerically stable. We analyze our algorithm from a novel geometric perspective and we show that for {\it any monotone transform} of a smooth and strongly convex objective with latent dimension $k \ge n$, we present a novel analysis that shows convergence within an $\epsilon$-ball of the optimum in $O(kQ\log(n)\log(R/\epsilon))$ evaluations, where the input dimension is $n$, $R$ is the diameter of the input space and $Q$ is the condition number. Our rates are the first of its kind to be both 1) poly-logarithmically dependent on dimensionality and 2) invariant under monotone transformations. We further leverage our geometric perspective to show that our analysis is optimal. Both monotone invariance and its ability to utilize a low latent dimensionality are key to the empirical success of our algorithms, as demonstrated on synthetic and MuJoCo benchmarks.

Gradientless Descent is a provably efficient gradient-free algorithm that is monotone-invariant and fast for high-dimensional zero-th order optimization.

Many processes can be concisely represented as a sequence of events leading from a starting state to an end state. Given raw ingredients, and a finished cake, an experienced chef can surmise the recipe. Building upon this intuition, we propose a new class of visual generative models: goal-conditioned predictors (GCP). Prior work on video generation largely focuses on prediction models that only observe frames from the beginning of the video. GCP instead treats videos as start-goal transformations, making video generation easier by conditioning on the more informative context provided by the first and final frames. Not only do existing forward prediction approaches synthesize better and longer videos when modified to become goal-conditioned, but GCP models can also utilize structures that are not linear in time, to accomplish hierarchical prediction . To this end, we study both auto-regressive GCP models and novel tree-structured GCP models that generate frames recursively, splitting the video iteratively into finer and finer segments delineated by subgoals . In experiments across simulated and real datasets, our GCP methods generate high-quality sequences over long horizons . Tree-structured GCPs are also substantially easier to parallelize than auto-regressive GCPs, making training and inference very efficient, and allowing the model to train on sequences that are thousands of frames in length.Finally, we demonstrate the utility of GCP approaches for imitation learning in the setting without access to expert actions . Videos are on the supplementary website: https://sites.google.com/view/video-gcp

We propose a new class of visual generative models: goal-conditioned predictors. We show experimentally that conditioning on the goal allows to reduce uncertainty and produce predictions over much longer horizons.

Recent advances in computing technology and sensor design have made it easier to collect longitudinal or time series data from patients, resulting in a gigantic amount of available medical data. Most of the medical time series lack annotations or even when the annotations are available they could be subjective and prone to human errors. Earlier works have developed natural language processing techniques to extract concept annotations and/or clinical narratives from doctor notes. However, these approaches are slow and do not use the accompanying medical time series data. To address this issue, we introduce the problem of concept annotation for the medical time series data, i.e., the task of predicting and localizing medical concepts by using the time series data as input. We propose Relational Multi-Instance Learning (RMIL) - a deep Multi Instance Learning framework based on recurrent neural networks, which uses pooling functions and attention mechanisms for the concept annotation tasks. Empirical results on medical datasets show that our proposed models outperform various multi-instance learning models.

We propose a deep Multi Instance Learning framework based on recurrent neural networks which uses pooling functions and attention mechanisms for the concept annotation tasks.

The embedding layers transforming input words into real vectors are the key components of deep neural networks used in natural language processing. However, when the vocabulary is large, the corresponding weight matrices can be enormous, which precludes their deployment in a limited resource setting. We introduce a novel way of parametrizing embedding layers based on the Tensor Train (TT) decomposition, which allows compressing the model significantly at the cost of a negligible drop or even a slight gain in performance. We evaluate our method on a wide range of benchmarks in natural language processing and analyze the trade-off between performance and compression ratios for a wide range of architectures, from MLPs to LSTMs and Transformers.

Embedding layers are factorized with Tensor Train decomposition to reduce their memory footprint.

We note that common implementations of adaptive gradient algorithms, such as Adam, limit the potential benefit of weight decay regularization, because the weights do not decay multiplicatively (as would be expected for standard weight decay) but by an additive constant factor. We propose a simple way to resolve this issue by decoupling weight decay and the optimization steps taken w.r.t. the loss function. We provide empirical evidence that our proposed modification (i) decouples the optimal choice of weight decay factor from the setting of the learning rate for both standard SGD and Adam, and (ii) substantially improves Adam's generalization performance, allowing it to compete with SGD with momentum on image classification datasets (on which it was previously typically outperformed by the latter). We also demonstrate that longer optimization runs require smaller weight decay values for optimal results and introduce a normalized variant of weight decay to reduce this dependence. Finally, we propose a version of Adam with warm restarts (AdamWR) that has strong anytime performance while achieving state-of-the-art results on CIFAR-10 and ImageNet32x32. Our source code will become available after the review process.

Fixing weight decay regularization in adaptive gradient methods such as Adam

Lifelong learning is the problem of learning multiple consecutive tasks in a sequential manner where knowledge gained from previous tasks is retained and used for future learning. It is essential towards the development of intelligent machines that can adapt to their surroundings. In this work we focus on a lifelong learning approach to generative modeling where we continuously incorporate newly observed streaming distributions into our learnt model. We do so through a student-teacher architecture which allows us to learn and preserve all the distributions seen so far without the need to retain the past data nor the past models. Through the introduction of a novel cross-model regularizer, the student model leverages the information learnt by the teacher, which acts as a summary of everything seen till now. The regularizer has the additional benefit of reducing the effect of catastrophic interference that appears when we learn over streaming data. We demonstrate its efficacy on streaming distributions as well as its ability to learn a common latent representation across a complex transfer learning scenario.

Lifelong distributional learning through a student-teacher architecture coupled with a cross model posterior regularizer.

Three-dimensional geometric data offer an excellent domain for studying representation learning and generative modeling. In this paper, we look at geometric data represented as point clouds. We introduce a deep autoencoder (AE) network with excellent reconstruction quality and generalization ability. The learned representations outperform the state of the art in 3D recognition tasks and enable basic shape editing applications via simple algebraic manipulations, such as semantic part editing, shape analogies and shape interpolation. We also perform a thorough study of different generative models including GANs operating on the raw point clouds, significantly improved GANs trained in the fixed latent space our AEs and, Gaussian mixture models (GMM). Interestingly, GMMs trained in the latent space of our AEs produce samples of the best fidelity and diversity. To perform our quantitative evaluation of generative models, we propose simple measures of fidelity and diversity based on optimally matching between sets point clouds.

Deep autoencoders to learn a good representation for geometric 3D point-cloud data; Generative models for point clouds.

Despite the remarkable performance of deep neural networks (DNNs) on various tasks, they are susceptible to adversarial perturbations which makes it difficult to deploy them in real-world safety-critical applications. In this paper, we aim to obtain robust networks by sparsifying DNN's latent features sensitive to adversarial perturbation. Specifically, we define vulnerability at the latent feature space and then propose a Bayesian framework to prioritize/prune features based on their contribution to both the original and adversarial loss. We also suggest regularizing the features' vulnerability during training to improve robustness further. While such network sparsification has been primarily studied in the literature for computational efficiency and regularization effect of DNNs, we confirm that it is also useful to design a defense mechanism through quantitative evaluation and qualitative analysis. We validate our method, \emph{Adversarial Neural Pruning (ANP)} on multiple benchmark datasets, which results in an improvement in test accuracy and leads to state-of-the-art robustness. ANP also tackles the practical problem of obtaining sparse and robust networks at the same time, which could be crucial to ensure adversarial robustness on lightweight networks deployed to computation and memory-limited devices.

We propose a novel method for suppressing the vulnerability of latent feature space to achieve robust and compact networks.

In anomaly detection (AD), one seeks to identify whether a test sample is abnormal, given a data set of normal samples. A recent and promising approach to AD relies on deep generative models, such as variational autoencoders (VAEs),for unsupervised learning of the normal data distribution. In semi-supervised AD (SSAD), the data also includes a small sample of labeled anomalies. In this work,we propose two variational methods for training VAEs for SSAD. The intuitive idea in both methods is to train the encoder to ‘separate’ between latent vectors for normal and outlier data. We show that this idea can be derived from principled probabilistic formulations of the problem, and propose simple and effective algorithms. Our methods can be applied to various data types, as we demonstrate on SSAD datasets ranging from natural images to astronomy and medicine, and can be combined with any VAE model architecture. When comparing to state-of-the-art SSAD methods that are not specific to particular data types, we obtain marked improvement in outlier detection.

We proposed two VAE modifications that account for negative data examples, and used them for semi-supervised anomaly detection.

We introduce dynamic instance hardness (DIH) to facilitate the training of machine learning models. DIH is a property of each training sample and is computed as the running mean of the sample's instantaneous hardness as measured over the training history. We use DIH to evaluate how well a model retains knowledge about each training sample over time. We find that for deep neural nets (DNNs), the DIH of a sample in relatively early training stages reflects its DIH in later stages and as a result, DIH can be effectively used to reduce the set of training samples in future epochs. Specifically, during each epoch, only samples with high DIH are trained (since they are historically hard) while samples with low DIH can be safely ignored. DIH is updated each epoch only for the selected samples, so it does not require additional computation. Hence, using DIH during training leads to an appreciable speedup. Also, since the model is focused on the historically more challenging samples, resultant models are more accurate. The above, when formulated as an algorithm, can be seen as a form of curriculum learning, so we call our framework DIH curriculum learning (or DIHCL). The advantages of DIHCL, compared to other curriculum learning approaches, are: (1) DIHCL does not require additional inference steps over the data not selected by DIHCL in each epoch, (2) the dynamic instance hardness, compared to static instance hardness (e.g., instantaneous loss), is more stable as it integrates information over the entire training history up to the present time. Making certain mathematical assumptions, we formulate the problem of DIHCL as finding a curriculum that maximizes a multi-set function $f(\cdot)$, and derive an approximation bound for a DIH-produced curriculum relative to the optimal curriculum. Empirically, DIHCL-trained DNNs significantly outperform random mini-batch SGD and other recently developed curriculum learning methods in terms of efficiency, early-stage convergence, and final performance, and this is shown in training several state-of-the-art DNNs on 11 modern datasets.

New understanding of training dynamics and metrics of memorization hardness lead to efficient and provable curriculum learning.

This paper explores many immediate connections between adaptive control and machine learning, both through common update laws as well as common concepts. Adaptive control as a field has focused on mathematical rigor and guaranteed convergence. The rapid advances in machine learning on the other hand have brought about a plethora of new techniques and problems for learning. This paper elucidates many of the numerous common connections between both fields such that results from both may be leveraged together to solve new problems. In particular, a specific problem related to higher order learning is solved through insights obtained from these intersections.

History of parallel developments in update laws and concepts between adaptive control and optimization in machine learning.

Recurrent convolution (RC) shares the same convolutional kernels and unrolls them multiple times, which is originally proposed to model time-space signals. We suggest that RC can be viewed as a model compression strategy for deep convolutional neural networks. RC reduces the redundancy across layers and is complementary to most existing model compression approaches. However, the performance of an RC network can't match the performance of its corresponding standard one, i.e. with the same depth but independent convolutional kernels. This reduces the value of RC for model compression. In this paper, we propose a simple variant which improves RC networks: The batch normalization layers of an RC module are learned independently (not shared) for different unrolling steps. We provide insights on why this works. Experiments on CIFAR show that unrolling a convolutional layer several steps can improve the performance, thus indirectly plays a role in model compression.

Recurrent convolution for model compression and a trick for training it, that is learning independent BN layres over steps.

The visual world is vast and varied, but its variations divide into structured and unstructured factors. Structured factors, such as scale and orientation, admit clear theories and efficient representation design. Unstructured factors, such as what it is that makes a cat look like a cat, are too complicated to model analytically, and so require free-form representation learning. We compose structured Gaussian filters and free-form filters, optimized end-to-end, to factorize the representation for efficient yet general learning. Our experiments on dynamic structure, in which the structured filters vary with the input, equal the accuracy of dynamic inference with more degrees of freedom while improving efficiency. (Please see https://arxiv.org/abs/1904.11487 for the full edition.)

Dynamic receptive fields with spatial Gaussian structure are accurate and efficient.

It is widely known that well-designed perturbations can cause state-of-the-art machine learning classifiers to mis-label an image, with sufficiently small perturbations that are imperceptible to the human eyes. However, by detecting the inconsistency between the image and wrong label, the human observer would be alerted of the attack. In this paper, we aim to design attacks that not only make classifiers generate wrong labels, but also make the wrong labels imperceptible to human observers. To achieve this, we propose an algorithm called LabelFool which identifies a target label similar to the ground truth label and finds a perturbation of the image for this target label. We first find the target label for an input image by a probability model, then move the input in the feature space towards the target label. Subjective studies on ImageNet show that in the label space, our attack is much less recognizable by human observers, while objective experimental results on ImageNet show that we maintain similar performance in the image space as well as attack rates to state-of-the-art attack algorithms.

A trick on adversarial samples so that the mis-classified labels are imperceptible in the label space to human observers

This paper presents noise type/position classification of various impact noises generated in a building which is a serious conflict issue in apartment complexes. For this study, a collection of floor impact noise dataset is recorded with a single microphone. Noise types/positions are selected based on a report by the Floor Management Center under Korea Environmental Corporation. Using a convolutional neural networks based classifier, the impact noise signals converted to log-scaled Mel-spectrograms are classified into noise types or positions. Also, our model is evaluated on a standard environmental sound dataset ESC-50 to show extensibility on environmental sound classification.

This paper presents noise type/position classification of various impact noises generated in a building which is a serious conflict issue in apartment complexes

Recordings of neural circuits in the brain reveal extraordinary dynamical richness and high variability. At the same time, dimensionality reduction techniques generally uncover low-dimensional structures underlying these dynamics. What determines the dimensionality of activity in neural circuits? What is the functional role of dimensionality in behavior and task learning? In this work we address these questions using recurrent neural network (RNN) models. We find that, depending on the dynamics of the initial network, RNNs learn to increase and reduce dimensionality in a way that matches task demands. These findings shed light on fundamental dynamical mechanisms by which neural networks solve tasks with robust representations that generalize to new cases.

Recurrent Neural Networks learn to increase and reduce the dimensionality of their internal representation in a way that matches the task, depending on the dynamics of the initial network.

Domain adaptation addresses the common problem when the target distribution generating our test data drifts from the source (training) distribution. While absent assumptions, domain adaptation is impossible, strict conditions, e.g. covariate or label shift, enable principled algorithms. Recently-proposed domain-adversarial approaches consist of aligning source and target encodings, often motivating this approach as minimizing two (of three) terms in a theoretical bound on target error. Unfortunately, this minimization can cause arbitrary increases in the third term, e.g. they can break down under shifting label distributions. We propose asymmetrically-relaxed distribution alignment, a new approach that overcomes some limitations of standard domain-adversarial algorithms. Moreover, we characterize precise assumptions under which our algorithm is theoretically principled and demonstrate empirical benefits on both synthetic and real datasets.

Instead of strict distribution alignments in traditional deep domain adaptation objectives, which fails when target label distribution shifts, we propose to optimize a relaxed objective with new analysis, new algorithms, and experimental validation.

In this paper, we explore \textit{summary-to-article generation}: the task of generating long articles given a short summary, which provides finer-grained content control for the generated text. To prevent sequence-to-sequence (seq2seq) models from degenerating into language models and better controlling the long text to be generated, we propose a hierarchical generation approach which first generates a sketch of intermediate length based on the summary and then completes the article by enriching the generated sketch. To mitigate the discrepancy between the ``oracle'' sketch used during training and the noisy sketch generated during inference, we propose an end-to-end joint training framework based on multi-agent reinforcement learning. For evaluation, we use text summarization corpora by reversing their inputs and outputs, and introduce a novel evaluation method that employs a summarization system to summarize the generated article and test its match with the original input summary. Experiments show that our proposed hierarchical generation approach can generate a coherent and relevant article based on the given summary, yielding significant improvements upon conventional seq2seq models.

we explore the task of summary-to-article generation and propose a hierarchical generation scheme together with a jointly end-to-end reinforcement learning framework to train the hierarchical model.

When training a deep neural network for supervised image classification, one can broadly distinguish between two types of latent features of images that will drive the classification of class Y. Following the notation of Gong et al. (2016), we can divide features broadly into the classes of (i) “core” or “conditionally invariant” features X^ci whose distribution P(X^ci | Y) does not change substantially across domains and (ii) “style” or “orthogonal” features X^orth whose distribution P(X^orth | Y) can change substantially across domains. These latter orthogonal features would generally include features such as position, rotation, image quality or brightness but also more complex ones like hair color or posture for images of persons. We try to guard against future adversarial domain shifts by ideally just using the “conditionally invariant” features for classification. In contrast to previous work, we assume that the domain itself is not observed and hence a latent variable. We can hence not directly see the distributional change of features across different domains. We do assume, however, that we can sometimes observe a so-called identifier or ID variable. We might know, for example, that two images show the same person, with ID referring to the identity of the person. In data augmentation, we generate several images from the same original image, with ID referring to the relevant original image. The method requires only a small fraction of images to have an ID variable. We provide a causal framework for the problem by adding the ID variable to the model of Gong et al. (2016). However, we are interested in settings where we cannot observe the domain directly and we treat domain as a latent variable. If two or more samples share the same class and identifier, (Y, ID)=(y,i), then we treat those samples as counterfactuals under different style interventions on the orthogonal or style features. Using this grouping-by-ID approach, we regularize the network to provide near constant output across samples that share the same ID by penalizing with an appropriate graph Laplacian. This is shown to substantially improve performance in settings where domains change in terms of image quality, brightness, color changes, and more complex changes such as changes in movement and posture. We show links to questions of interpretability, fairness and transfer learning.

We propose counterfactual regularization to guard against adversarial domain shifts arising through shifts in the distribution of latent "style features" of images.

Gradient-based meta-learning algorithms require several steps of gradient descent to adapt to newly incoming tasks. This process becomes more costly as the number of samples increases. Moreover, the gradient updates suffer from several sources of noise leading to a degraded performance . In this work, we propose a meta-learning algorithm equipped with the GradiEnt Component COrrections, aGECCO cell for short, which generates a multiplicative corrective low-rank matrix which (after vectorization) corrects the estimated gradients . GECCO contains a simple decoder-like network with learnable parameters, an attention module and a so-called context input parameter . The context parameter of GECCO is updated to generate a low-rank corrective term for the network gradients . As a result, meta-learning requires only a few of gradient updates to absorb new task (often, a single update is sufficient in the few-shot scenario ). While previous approaches address this problem by altering the learning rates, factorising network parameters or directly learning feature corrections from features and/or gradients, GECCO is an off-the-shelf generator-like unit that performs element-wise gradient corrections without the need to ‘observe’ the features and/or the gradients directly . We show that our GECCO (i) accelerates learning, (ii) performs robust corrections of the gradients corrupted by a noise, and (iii) leads to notable improvements over existing gradient-based meta-learning algorithms.

We propose a meta-learner to adapt quickly on multiple tasks even one step in a few-shot setting.

Discriminative question answering models can overfit to superficial biases in datasets, because their loss function saturates when any clue makes the answer likely. We introduce generative models of the joint distribution of questions and answers, which are trained to explain the whole question, not just to answer it. Our question answering (QA) model is implemented by learning a prior over answers, and a conditional language model to generate the question given the answer—allowing scalable and interpretable many-hop reasoning as the question is generated word-by-word. Our model achieves competitive performance with specialised discriminative models on the SQUAD and CLEVR benchmarks, indicating that it is a more general architecture for language understanding and reasoning than previous work. The model greatly improves generalisation both from biased training data and to adversarial testing data, achieving a new state-of-the-art on ADVERSARIAL SQUAD. We will release our code.

Question answering models that model the joint distribution of questions and answers can learn more than discriminative models

In this paper, we turn our attention to the interworking between the activation functions and the batch normalization, which is a virtually mandatory technique to train deep networks currently. We propose the activation function Displaced Rectifier Linear Unit (DReLU) by conjecturing that extending the identity function of ReLU to the third quadrant enhances compatibility with batch normalization. Moreover, we used statistical tests to compare the impact of using distinct activation functions (ReLU, LReLU, PReLU, ELU, and DReLU) on the learning speed and test accuracy performance of standardized VGG and Residual Networks state-of-the-art models. These convolutional neural networks were trained on CIFAR-100 and CIFAR-10, the most commonly used deep learning computer vision datasets. The results showed DReLU speeded up learning in all models and datasets. Besides, statistical significant performance assessments (p<0.05) showed DReLU enhanced the test accuracy presented by ReLU in all scenarios. Furthermore, DReLU showed better test accuracy than any other tested activation function in all experiments with one exception, in which case it presented the second best performance. Therefore, this work demonstrates that it is possible to increase performance replacing ReLU by an enhanced activation function.

A new activation function called Displaced Rectifier Linear Unit is proposed. It is showed to enhance the training and inference performance of batch normalized convolutional neural networks.

Encoding the input scale information explicitly into the representation learned by a convolutional neural network (CNN) is beneficial for many vision tasks especially when dealing with multiscale input signals. We study, in this paper, a scale-equivariant CNN architecture with joint convolutions across the space and the scaling group, which is shown to be both sufficient and necessary to achieve scale-equivariant representations. To reduce the model complexity and computational burden, we decompose the convolutional filters under two pre-fixed separable bases and truncate the expansion to low-frequency components. A further benefit of the truncated filter expansion is the improved deformation robustness of the equivariant representation. Numerical experiments demonstrate that the proposed scale-equivariant neural network with decomposed convolutional filters (ScDCFNet) achieves significantly improved performance in multiscale image classification and better interpretability than regular CNNs at a reduced model size.

We construct scale-equivariant convolutional neural networks in the most general form with both computational efficiency and proved deformation robustness.

In this paper, we diagnose deep neural networks for 3D point cloud processing to explore the utility of different network architectures. We propose a number of hypotheses on the effects of specific network architectures on the representation capacity of DNNs. In order to prove the hypotheses, we design five metrics to diagnose various types of DNNs from the following perspectives, information discarding, information concentration, rotation robustness, adversarial robustness, and neighborhood inconsistency. We conduct comparative studies based on such metrics to verify the hypotheses, which may shed new lights on the architectural design of neural networks. Experiments demonstrated the effectiveness of our method. The code will be released when this paper is accepted.

We diagnose deep neural networks for 3D point cloud processing to explore the utility of different network architectures.

In this work we construct flexible joint distributions from low-dimensional conditional semi-implicit distributions. Explicitly defining the structure of the approximation allows to make the variational lower bound tighter, resulting in more accurate inference.

Utilizing the structure of distributions improves semi-implicit variational inference

Imitation learning from human-expert demonstrations has been shown to be greatly helpful for challenging reinforcement learning problems with sparse environment rewards. However, it is very difficult to achieve similar success without relying on expert demonstrations. Recent works on self-imitation learning showed that imitating the agent's own past good experience could indirectly drive exploration in some environments, but these methods often lead to sub-optimal and myopic behavior. To address this issue, we argue that exploration in diverse directions by imitating diverse trajectories, instead of focusing on limited good trajectories, is more desirable for the hard-exploration tasks. We propose a new method of learning a trajectory-conditioned policy to imitate diverse trajectories from the agent's own past experiences and show that such self-imitation helps avoid myopic behavior and increases the chance of finding a globally optimal solution for hard-exploration tasks, especially when there are misleading rewards. Our method significantly outperforms existing self-imitation learning and count-based exploration methods on various hard-exploration tasks with local optima. In particular, we report a state-of-the-art score of more than 20,000 points on Montezumas Revenge without using expert demonstrations or resetting to arbitrary states.

Self-imitation learning of diverse trajectories with trajectory-conditioned policy

We present a method that trains large capacity neural networks with significantly improved accuracy and lower dynamic computational cost. This is achieved by gating the deep-learning architecture on a fine-grained-level. Individual convolutional maps are turned on/off conditionally on features in the network. To achieve this, we introduce a new residual block architecture that gates convolutional channels in a fine-grained manner. We also introduce a generally applicable tool batch-shaping that matches the marginal aggregate posteriors of features in a neural network to a pre-specified prior distribution. We use this novel technique to force gates to be more conditional on the data. We present results on CIFAR-10 and ImageNet datasets for image classification, and Cityscapes for semantic segmentation. Our results show that our method can slim down large architectures conditionally, such that the average computational cost on the data is on par with a smaller architecture, but with higher accuracy. In particular, on ImageNet, our ResNet50 and ResNet34 gated networks obtain 74.60% and 72.55% top-1 accuracy compared to the 69.76% accuracy of the baseline ResNet18 model, for similar complexity. We also show that the resulting networks automatically learn to use more features for difficult examples and fewer features for simple examples.

A method that trains large capacity neural networks with significantly improved accuracy and lower dynamic computational cost

With a view to bridging the gap between deep learning and symbolic AI, we present a novel end-to-end neural network architecture that learns to form propositional representations with an explicitly relational structure from raw pixel data. In order to evaluate and analyse the architecture, we introduce a family of simple visual relational reasoning tasks of varying complexity. We show that the proposed architecture, when pre-trained on a curriculum of such tasks, learns to generate reusable representations that better facilitate subsequent learning on previously unseen tasks when compared to a number of baseline architectures. The workings of a successfully trained model are visualised to shed some light on how the architecture functions.

We present an end-to-end differentiable architecture that learns to map pixels to predicates, and evaluate it on a suite of simple relational reasoning tasks

In natural language inference, the semantics of some words do not affect the inference. Such information is considered superficial and brings overfitting. How can we represent and discard such superficial information? In this paper, we use first order logic (FOL) - a classic technique from meaning representation language – to explain what information is superficial for a given sentence pair. Such explanation also suggests two inductive biases according to its properties. We proposed a neural network-based approach that utilizes the two inductive biases. We obtain substantial improvements over extensive experiments.

We use neural networks to project superficial information out for natural language inference by defining and identifying the superficial information from the perspective of first-order logic.

We propose an approach to training machine learning models that are fair in the sense that their performance is invariant under certain perturbations to the features. For example, the performance of a resume screening system should be invariant under changes to the name of the applicant. We formalize this intuitive notion of fairness by connecting it to the original notion of individual fairness put forth by Dwork et al and show that the proposed approach achieves this notion of fairness. We also demonstrate the effectiveness of the approach on two machine learning tasks that are susceptible to gender and racial biases.

Algorithm for training individually fair classifier using adversarial robustness

In this paper, we propose a Seed-Augment-Train/Transfer (SAT) framework that contains a synthetic seed image dataset generation procedure for languages with different numeral systems using freely available open font file datasets. This seed dataset of images is then augmented to create a purely synthetic training dataset, which is in turn used to train a deep neural network and test on held-out real world handwritten digits dataset spanning five Indic scripts, Kannada, Tamil, Gujarati, Malayalam, and Devanagari. We showcase the efficacy of this approach both qualitatively, by training a Boundary-seeking GAN (BGAN) that generates realistic digit images in the five languages, and also qualitatively by testing a CNN trained on the synthetic data on the real-world datasets. This establishes not only an interesting nexus between the font-datasets-world and transfer learning but also provides a recipe for universal-digit classification in any script.

Is seeding and augmentation all you need for classifying digits in any language?

An important research direction in machine learning has centered around developing meta-learning algorithms to tackle few-shot learning. An especially successful algorithm has been Model Agnostic Meta-Learning (MAML), a method that consists of two optimization loops, with the outer loop finding a meta-initialization, from which the inner loop can efficiently learn new tasks. Despite MAML's popularity, a fundamental open question remains -- is the effectiveness of MAML due to the meta-initialization being primed for rapid learning (large, efficient changes in the representations) or due to feature reuse, with the meta initialization already containing high quality features? We investigate this question, via ablation studies and analysis of the latent representations, finding that feature reuse is the dominant factor. This leads to the ANIL (Almost No Inner Loop) algorithm, a simplification of MAML where we remove the inner loop for all but the (task-specific) head of the underlying neural network. ANIL matches MAML's performance on benchmark few-shot image classification and RL and offers computational improvements over MAML. We further study the precise contributions of the head and body of the network, showing that performance on the test tasks is entirely determined by the quality of the learned features, and we can remove even the head of the network (the NIL algorithm). We conclude with a discussion of the rapid learning vs feature reuse question for meta-learning algorithms more broadly.

The success of MAML relies on feature reuse from the meta-initialization, which also yields a natural simplification of the algorithm, with the inner loop removed for the network body, as well as other insights on the head and body.

Model training remains a dominant financial cost and time investment in machine learning applications. Developing and debugging models often involve iterative training, further exacerbating this issue. With growing interest in increasingly complex models, there is a need for techniques that help to reduce overall training effort. While incremental training can save substantial time and cost by training an existing model on a small subset of data, little work has explored policies for determining when incremental training provides adequate model performance versus full retraining. We provide a method-agnostic algorithm for deciding when to incrementally train versus fully train. We call this setting of non-deterministic full- or incremental training ``Mixed Setting Training". Upon evaluation in slot-filling tasks, we find that this algorithm provides a bounded error, avoids catastrophic forgetting, and results in a significant speedup over a policy of always fully training.

We provide a method-agnostic algorithm for deciding when to incrementally train versus fully train and it provides a significant speedup over fully training and avoids catastrophic forgetting

Neural networks have succeeded in many reasoning tasks. Empirically, these tasks require specialized network structures, e.g., Graph Neural Networks (GNNs) perform well on many such tasks, while less structured networks fail. Theoretically, there is limited understanding of why and when a network structure generalizes better than other equally expressive ones. We develop a framework to characterize which reasoning tasks a network can learn well, by studying how well its structure aligns with the algorithmic structure of the relevant reasoning procedure. We formally define algorithmic alignment and derive a sample complexity bound that decreases with better alignment. This framework explains the empirical success of popular reasoning models and suggests their limitations. We unify seemingly different reasoning tasks, such as intuitive physics, visual question answering, and shortest paths, via the lens of a powerful algorithmic paradigm, dynamic programming (DP). We show that GNNs can learn DP and thus solve these tasks. On several reasoning tasks, our theory aligns with empirical results.

We develop a theoretical framework to characterize which reasoning tasks a neural network can learn well.

Cell-cell interactions have an integral role in tumorigenesis as they are critical in governing immune responses. As such, investigating specific cell-cell interactions has the potential to not only expand upon the understanding of tumorigenesis, but also guide clinical management of patient responses to cancer immunotherapies. A recent imaging technique for exploring cell-cell interactions, multiplexed ion beam imaging by time-of-flight (MIBI-TOF), allows for cells to be quantified in 36 different protein markers at sub-cellular resolutions in situ as high resolution multiplexed images. To explore the MIBI images, we propose a GAN for multiplexed data with protein specific attention. By conditioning image generation on cell types, sizes, and neighborhoods through semantic segmentation maps, we are able to observe how these factors affect cell-cell interactions simultaneously in different protein channels. Furthermore, we design a set of metrics and offer the first insights towards cell spatial orientations, cell protein expressions, and cell neighborhoods. Our model, cell-cell interaction GAN (CCIGAN), outperforms or matches existing image synthesis methods on all conventional measures and significantly outperforms on biologically motivated metrics. To our knowledge, we are the first to systematically model multiple cellular protein behaviors and interactions under simulated conditions through image synthesis.

We explore cell-cell interactions across tumor environment contexts observed in highly multiplexed images, by image synthesis using a novel attention GAN architecture.

Machine learning models for question-answering (QA), where given a question and a passage, the learner must select some span in the passage as an answer, are known to be brittle. By inserting a single nuisance sentence into the passage, an adversary can fool the model into selecting the wrong span. A promising new approach for QA decomposes the task into two stages: (i) select relevant sentences from the passage; and (ii) select a span among those sentences. Intuitively, if the sentence selector excludes the offending sentence, then the downstream span selector will be robust. While recent work has hinted at the potential robustness of two-stage QA, these methods have never, to our knowledge, been explicitly combined with adversarial training. This paper offers a thorough empirical investigation of adversarial robustness, demonstrating that although the two-stage approach lags behind single-stage span selection, adversarial training improves its performance significantly, leading to an improvement of over 22 points in F1 score over the adversarially-trained single-stage model.

A two-stage approach consisting of sentence selection followed by span selection can be made more robust to adversarial attacks in comparison to a single-stage model trained on full context.

The aim of this study is to introduce a formal framework for analysis and synthesis of driver assistance systems. It applies formal methods to the verification of a stochastic human driver model built using the cognitive architecture ACT-R, and then bootstraps safety in semi-autonomous vehicles through the design of provably correct Advanced Driver Assistance Systems. The main contributions include the integration of probabilistic ACT-R models in the formal analysis of semi-autonomous systems and an abstraction technique that enables a finite representation of a large dimensional, continuous system in the form of a Markov model. The effectiveness of the method is illustrated in several case studies under various conditions.

Verification of a human driver model based on a cognitive architecture and synthesis of a correct-by-construction ADAS from it.

In contrast to the older writing system of the 19th century, modern Hawaiian orthography employs characters for long vowels and glottal stops. These extra characters account for about one-third of the phonemes in Hawaiian, so including them makes a big difference to reading comprehension and pronunciation. However, transliterating between older and newer texts is a laborious task when performed manually. We introduce two related methods to help solve this transliteration problem automatically, given that there were not enough data to train an end-to-end deep learning model. One approach is implemented, end-to-end, using finite state transducers (FSTs). The other is a hybrid deep learning approach which approximately composes an FST with a recurrent neural network (RNN). We find that the hybrid approach outperforms the end-to-end FST by partitioning the original problem into one part that can be modelled by hand, using an FST, and into another part, which is easily solved by an RNN trained on the available data.

A novel, hybrid deep learning approach provides the best solution to a limited-data problem (which is important to the conservation of the Hawaiian language)

In many real-world settings, a learning model must perform few-shot classification: learn to classify examples from unseen classes using only a few labeled examples per class. Additionally, to be safely deployed, it should have the ability to detect out-of-distribution inputs: examples that do not belong to any of the classes. While both few-shot classification and out-of-distribution detection are popular topics, their combination has not been studied. In this work, we propose tasks for out-of-distribution detection in the few-shot setting and establish benchmark datasets, based on four popular few-shot classification datasets. Then, we propose two new methods for this task and investigate their performance. In sum, we establish baseline out-of-distribution detection results using standard metrics on new benchmark datasets and show improved results with our proposed methods.

We quantitatively study out-of-distribution detection in few-shot setting, establish baseline results with ProtoNet, MAML, ABML, and improved upon them.

While modern generative models are able to synthesize high-fidelity, visually appealing images, successfully generating examples that are useful for recognition tasks remains an elusive goal. To this end, our key insight is that the examples should be synthesized to recover classifier decision boundaries that would be learned from a large amount of real examples. More concretely, we treat a classifier trained on synthetic examples as ''student'' and a classifier trained on real examples as ''teacher''. By introducing knowledge distillation into a meta-learning framework, we encourage the generative model to produce examples in a way that enables the student classifier to mimic the behavior of the teacher. To mitigate the potential gap between student and teacher classifiers, we further propose to distill the knowledge in a progressive manner, either by gradually strengthening the teacher or weakening the student. We demonstrate the use of our model-agnostic distillation approach to deal with data scarcity, significantly improving few-shot learning performance on miniImageNet and ImageNet1K benchmarks.

This paper introduces progressive knowledge distillation for learning generative models that are recognition task oriented

Deep neural networks provide state-of-the-art performance for many applications of interest. Unfortunately they are known to be vulnerable to adversarial examples, formed by applying small but malicious perturbations to the original inputs. Moreover, the perturbations can transfer across models: adversarial examples generated for a specific model will often mislead other unseen models. Consequently the adversary can leverage it to attack against the deployed black-box systems. In this work, we demonstrate that the adversarial perturbation can be decomposed into two components: model-specific and data-dependent one, and it is the latter that mainly contributes to the transferability. Motivated by this understanding, we propose to craft adversarial examples by utilizing the noise reduced gradient (NRG) which approximates the data-dependent component. Experiments on various classification models trained on ImageNet demonstrates that the new approach enhances the transferability dramatically. We also find that low-capacity models have more powerful attack capability than high-capacity counterparts, under the condition that they have comparable test performance. These insights give rise to a principled manner to construct adversarial examples with high success rates and could potentially provide us guidance for designing effective defense approaches against black-box attacks.

We propose a new method for enhancing the transferability of adversarial examples by using the noise-reduced gradient.

We present the iterative two-pass decomposition flow to accelerate existing convolutional neural networks (CNNs). The proposed rank selection algorithm can effectively determine the proper ranks of the target convolutional layers for the low rank approximation. Our two-pass CP-decomposition helps prevent from the instability problem. The iterative flow makes the decomposition of the deeper networks systematic. The experiment results shows that VGG16 can be accelerated with a 6.2x measured speedup while the accuracy drop remains only 1.2%.

We present the iterative two-pass CP decomposition flow to effectively accelerate existing convolutional neural networks (CNNs).

We introduce LiPopt, a polynomial optimization framework for computing increasingly tighter upper bound on the Lipschitz constant of neural networks. The underlying optimization problems boil down to either linear (LP) or semidefinite (SDP) programming. We show how to use the sparse connectivity of a network, to significantly reduce the complexity of computation. This is specially useful for convolutional as well as pruned neural networks. We conduct experiments on networks with random weights as well as networks trained on MNIST, showing that in the particular case of the $\ell_\infty$-Lipschitz constant, our approach yields superior estimates as compared to other baselines available in the literature.

LP-based upper bounds on the Lipschitz constant of Neural Networks

Although few-shot learning research has advanced rapidly with the help of meta-learning, its practical usefulness is still limited because most of the researches assumed that all meta-training and meta-testing examples came from a single domain. We propose a simple but effective way for few-shot classification in which a task distribution spans multiple domains including previously unseen ones during meta-training. The key idea is to build a pool of embedding models which have their own metric spaces and to learn to select the best one for a particular task through multi-domain meta-learning. This simplifies task-specific adaptation over a complex task distribution as a simple selection problem rather than modifying the model with a number of parameters at meta-testing time. Inspired by common multi-task learning techniques, we let all models in the pool share a base network and add a separate modulator to each model to refine the base network in its own way. This architecture allows the pool to maintain representational diversity and each model to have domain-invariant representation as well. Experiments show that our selection scheme outperforms other few-shot classification algorithms when target tasks could come from many different domains. They also reveal that aggregating outputs from all constituent models is effective for tasks from unseen domains showing the effectiveness of our framework.

We address multi-domain few-shot classification by building multiple models to represent this complex task distribution in a collective way and simplifying task-specific adaptation as a selection problem from these pre-trained models.

Still in 2019, many scanned documents come into businesses in non-digital format. Text to be extracted from real world documents is often nestled inside rich formatting, such as tabular structures or forms with fill-in-the-blank boxes or underlines whose ink often touches or even strikes through the ink of the text itself. Such ink artifacts can severely interfere with the performance of recognition algorithms or other downstream processing tasks. In this work, we propose DeepErase, a neural preprocessor to erase ink artifacts from text images. We devise a method to programmatically augment text images with real artifacts, and use them to train a segmentation network in an weakly supervised manner. In additional to high segmentation accuracy, we show that our cleansed images achieve a significant boost in downstream recognition accuracy by popular OCR software such as Tesseract 4.0. We test DeepErase on out-of-distribution datasets (NIST SDB) of scanned IRS tax return forms and achieve double-digit improvements in recognition accuracy over baseline for both printed and handwritten text.

Neural-based removal of document ink artifacts (underlines, smudges, etc.) using no manually annotated training data

Black-box adversarial attacks require a large number of attempts before finding successful adversarial examples that are visually indistinguishable from the original input. Current approaches relying on substitute model training, gradient estimation or genetic algorithms often require an excessive number of queries. Therefore, they are not suitable for real-world systems where the maximum query number is limited due to cost. We propose a query-efficient black-box attack which uses Bayesian optimisation in combination with Bayesian model selection to optimise over the adversarial perturbation and the optimal degree of search space dimension reduction. We demonstrate empirically that our method can achieve comparable success rates with 2-5 times fewer queries compared to previous state-of-the-art black-box attacks.

We propose a query-efficient black-box attack which uses Bayesian optimisation in combination with Bayesian model selection to optimise over the adversarial perturbation and the optimal degree of search space dimension reduction.

Learning multimodal representations is a fundamentally complex research problem due to the presence of multiple heterogeneous sources of information. Although the presence of multiple modalities provides additional valuable information, there are two key challenges to address when learning from multimodal data: 1) models must learn the complex intra-modal and cross-modal interactions for prediction and 2) models must be robust to unexpected missing or noisy modalities during testing. In this paper, we propose to optimize for a joint generative-discriminative objective across multimodal data and labels. We introduce a model that factorizes representations into two sets of independent factors: multimodal discriminative and modality-specific generative factors. Multimodal discriminative factors are shared across all modalities and contain joint multimodal features required for discriminative tasks such as sentiment prediction. Modality-specific generative factors are unique for each modality and contain the information required for generating data. Experimental results show that our model is able to learn meaningful multimodal representations that achieve state-of-the-art or competitive performance on six multimodal datasets. Our model demonstrates flexible generative capabilities by conditioning on independent factors and can reconstruct missing modalities without significantly impacting performance. Lastly, we interpret our factorized representations to understand the interactions that influence multimodal learning.

We propose a model to learn factorized multimodal representations that are discriminative, generative, and interpretable.

The successful application of flexible, general learning algorithms to real-world robotics applications is often limited by their poor data-efficiency. To address the challenge, domains with more than one dominant task of interest encourage the sharing of information across tasks to limit required experiment time. To this end, we investigate compositional inductive biases in the form of hierarchical policies as a mechanism for knowledge transfer across tasks in reinforcement learning (RL). We demonstrate that this type of hierarchy enables positive transfer while mitigating negative interference. Furthermore, we demonstrate the benefits of additional incentives to efficiently decompose task solutions. Our experiments show that these incentives are naturally given in multitask learning and can be easily introduced for single objectives. We design an RL algorithm that enables stable and fast learning of structured policies and the effective reuse of both behavior components and transition data across tasks in an off-policy setting. Finally, we evaluate our algorithm in simulated environments as well as physical robot experiments and demonstrate substantial improvements in data data-efficiency over competitive baselines.

We develop a hierarchical, actor-critic algorithm for compositional transfer by sharing policy components and demonstrate component specialization and related direct benefits in multitask domains as well as its adaptation for single tasks.

In this paper, we study the representational power of deep neural networks (DNN) that belong to the family of piecewise-linear (PWL) functions, based on PWL activation units such as rectifier or maxout. We investigate the complexity of such networks by studying the number of linear regions of the PWL function. Typically, a PWL function from a DNN can be seen as a large family of linear functions acting on millions of such regions. We directly build upon the work of Mont´ufar et al. (2014), Mont´ufar (2017), and Raghu et al. (2017) by refining the upper and lower bounds on the number of linear regions for rectified and maxout networks. In addition to achieving tighter bounds, we also develop a novel method to perform exact numeration or counting of the number of linear regions with a mixed-integer linear formulation that maps the input space to output. We use this new capability to visualize how the number of linear regions change while training DNNs.

We empirically count the number of linear regions of rectifier networks and refine upper and lower bounds.

Convolutional neural networks memorize part of their training data, which is why strategies such as data augmentation and drop-out are employed to mitigate over- fitting. This paper considers the related question of “membership inference”, where the goal is to determine if an image was used during training. We con- sider membership tests over either ensembles of samples or individual samples. First, we show how to detect if a dataset was used to train a model, and in particular whether some validation images were used at train time. Then, we introduce a new approach to infer membership when a few of the top layers are not available or have been fine-tuned, and show that lower layers still carry information about the training samples. To support our findings, we conduct large-scale experiments on Imagenet and subsets of YFCC-100M with modern architectures such as VGG and Resnet.

We analyze the memorization properties by a convnet of the training set and propose several use-cases where we can extract some information about the training set.

While Generative Adversarial Networks (GANs) have empirically produced impressive results on learning complex real-world distributions, recent works have shown that they suffer from lack of diversity or mode collapse. The theoretical work of Arora et al. (2017a) suggests a dilemma about GANs’ statistical properties: powerful discriminators cause overfitting, whereas weak discriminators cannot detect mode collapse. By contrast, we show in this paper that GANs can in principle learn distributions in Wasserstein distance (or KL-divergence in many cases) with polynomial sample complexity, if the discriminator class has strong distinguishing power against the particular generator class (instead of against all possible generators). For various generator classes such as mixture of Gaussians, exponential families, and invertible and injective neural networks generators, we design corresponding discriminators (which are often neural nets of specific architectures) such that the Integral Probability Metric (IPM) induced by the discriminators can provably approximate the Wasserstein distance and/or KL-divergence. This implies that if the training is successful, then the learned distribution is close to the true distribution in Wasserstein distance or KL divergence, and thus cannot drop modes. Our preliminary experiments show that on synthetic datasets the test IPM is well correlated with KL divergence or the Wasserstein distance, indicating that the lack of diversity in GANs may be caused by the sub-optimality in optimization instead of statistical inefficiency.

GANs can in principle learn distributions sample-efficiently, if the discriminator class is compact and has strong distinguishing power against the particular generator class.

Understanding the optimization trajectory is critical to understand training of deep neural networks. We show how the hyperparameters of stochastic gradient descent influence the covariance of the gradients (K) and the Hessian of the training loss (H) along this trajectory. Based on a theoretical model, we predict that using a high learning rate or a small batch size in the early phase of training leads SGD to regions of the parameter space with (1) reduced spectral norm of K, and (2) improved conditioning of K and H. We show that the point on the trajectory after which these effects hold, which we refer to as the break-even point, is reached early during training. We demonstrate these effects empirically for a range of deep neural networks applied to multiple different tasks. Finally, we apply our analysis to networks with batch normalization (BN) layers and find that it is necessary to use a high learning rate to achieve loss smoothing effects attributed previously to BN alone.

In the early phase of training of deep neural networks there exists a "break-even point" which determines properties of the entire optimization trajectory.

Graph Convolution Network (GCN) has been recognized as one of the most effective graph models for semi-supervised learning, but it extracts merely the first-order or few-order neighborhood information through information propagation, which suffers performance drop-off for deeper structure. Existing approaches that deal with the higher-order neighbors tend to take advantage of adjacency matrix power. In this paper, we assume a seemly trivial condition that the higher-order neighborhood information may be similar to that of the first-order neighbors. Accordingly, we present an unsupervised approach to describe such similarities and learn the weight matrices of higher-order neighbors automatically through Lasso that minimizes the feature loss between the first-order and higher-order neighbors, based on which we formulate the new convolutional filter for GCN to learn the better node representations. Our model, called higher-order weighted GCN (HWGCN), has achieved the state-of-the-art results on a number of node classification tasks over Cora, Citeseer and Pubmed datasets.

We propose HWGCN to mix the relevant neighborhood information at different orders to better learn node representations.

The performance of deep neural networks is often attributed to their automated, task-related feature construction. It remains an open question, though, why this leads to solutions with good generalization, even in cases where the number of parameters is larger than the number of samples. Back in the 90s, Hochreiter and Schmidhuber observed that flatness of the loss surface around a local minimum correlates with low generalization error. For several flatness measures, this correlation has been empirically validated. However, it has recently been shown that existing measures of flatness cannot theoretically be related to generalization: if a network uses ReLU activations, the network function can be reparameterized without changing its output in such a way that flatness is changed almost arbitrarily. This paper proposes a natural modification of existing flatness measures that results in invariance to reparameterization. The proposed measures imply a robustness of the network to changes in the input and the hidden layers. Connecting this feature robustness to generalization leads to a generalized definition of the representativeness of data. With this, the generalization error of a model trained on representative data can be bounded by its feature robustness which depends on our novel flatness measure.

We introduce a novel measure of flatness at local minima of the loss surface of deep neural networks which is invariant with respect to layer-wise reparameterizations and we connect flatness to feature robustness and generalization.

Bayesian methods have been successfully applied to sparsify weights of neural networks and to remove structure units from the networks, e. g. neurons. We apply and further develop this approach for gated recurrent architectures. Specifically, in addition to sparsification of individual weights and neurons, we propose to sparsify preactivations of gates and information flow in LSTM. It makes some gates and information flow components constant, speeds up forward pass and improves compression. Moreover, the resulting structure of gate sparsity is interpretable and depends on the task.

We propose to sparsify preactivations of gates and information flow in LSTM to make them constant and boost the neuron sparsity level

Improving the accuracy of numerical methods remains a central challenge in many disciplines and is especially important for nonlinear simulation problems. A representative example of such problems is fluid flow, which has been thoroughly studied to arrive at efficient simulations of complex flow phenomena. This paper presents a data-driven approach that learns to improve the accuracy of numerical solvers. The proposed method utilizes an advanced numerical scheme with a fine simulation resolution to acquire reference data. We, then, employ a neural network that infers a correction to move a coarse thus quickly obtainable result closer to the reference data. We provide insights into the targeted learning problem with different learning approaches: fully supervised learning methods with a naive and an optimized data acquisition as well as an unsupervised learning method with a differentiable Navier-Stokes solver. While our approach is very general and applicable to arbitrary partial differential equation models, we specifically highlight gains in accuracy for fluid flow simulations.

We introduce a neural network approach to assist partial differential equation solvers.

A patient’s health information is generally fragmented across silos. Though it is technically feasible to unite data for analysis in a manner that underpins a rapid learning healthcare system, privacy concerns and regulatory barriers limit data centralization. Machine learning can be conducted in a federated manner on patient datasets with the same set of variables, but separated across sites of care. But federated learning cannot handle the situation where different data types for a given patient are separated vertically across different organizations. We call methods that enable machine learning model training on data separated by two or more degrees “confederated machine learning.” We built and evaluated a confederated machine learning model to stratify the risk of accidental falls among the elderly.

a confederated learning method that train model from horizontally and vertically separated medical data

Existing neural networks are vulnerable to "adversarial examples"---created by adding maliciously designed small perturbations in inputs to induce a misclassification by the networks. The most investigated defense strategy is adversarial training which augments training data with adversarial examples. However, applying single-step adversaries in adversarial training does not support the robustness of the networks, instead, they will even make the networks to be overfitted. In contrast to the single-step, multi-step training results in the state-of-the-art performance on MNIST and CIFAR10, yet it needs a massive amount of time. Therefore, we propose a method, Stochastic Quantized Activation (SQA) that solves overfitting problems in single-step adversarial training and fastly achieves the robustness comparable to the multi-step. SQA attenuates the adversarial effects by providing random selectivity to activation functions and allows the network to learn robustness with only single-step training. Throughout the experiment, our method demonstrates the state-of-the-art robustness against one of the strongest white-box attacks as PGD training, but with much less computational cost. Finally, we visualize the learning process of the network with SQA to handle strong adversaries, which is different from existing methods.

This paper proposes Stochastic Quantized Activation that solves overfitting problems in FGSM adversarial training and fastly achieves the robustness comparable to multi-step training.

Neural activity is highly variable in response to repeated stimuli. We used an open dataset, the Allen Brain Observatory, to quantify the distribution of responses to repeated natural movie presentations. A large fraction of responses are best fit by log-normal distributions or Gaussian mixtures with two components. These distributions are similar to those from units in deep neural networks with dropout. Using a separate set of electrophysiological recordings, we constructed a population coupling model as a control for state-dependent activity fluctuations and found that the model residuals also show non-Gaussian distributions. We then analyzed responses across trials from multiple sections of different movie clips and observed that the noise in cortex aligns better with in-clip versus out-of-clip stimulus variations. We argue that noise is useful for generalization when it moves along representations of different exemplars in-class, similar to the structure of cortical noise.

We study the structure of noise in the brain and find it may help generalization by moving representations along in-class stimulus variations.

Unsupervised domain adaptation has received significant attention in recent years. Most of existing works tackle the closed-set scenario, assuming that the source and target domains share the exactly same categories. In practice, nevertheless, a target domain often contains samples of classes unseen in source domain (i.e., unknown class). The extension of domain adaptation from closed-set to such open-set situation is not trivial since the target samples in unknown class are not expected to align with the source. In this paper, we address this problem by augmenting the state-of-the-art domain adaptation technique, Self-Ensembling, with category-agnostic clusters in target domain. Specifically, we present Self-Ensembling with Category-agnostic Clusters (SE-CC) --- a novel architecture that steers domain adaptation with the additional guidance of category-agnostic clusters that are specific to target domain. These clustering information provides domain-specific visual cues, facilitating the generalization of Self-Ensembling for both closed-set and open-set scenarios. Technically, clustering is firstly performed over all the unlabeled target samples to obtain the category-agnostic clusters, which reveal the underlying data space structure peculiar to target domain. A clustering branch is capitalized on to ensure that the learnt representation preserves such underlying structure by matching the estimated assignment distribution over clusters to the inherent cluster distribution for each target sample. Furthermore, SE-CC enhances the learnt representation with mutual information maximization. Extensive experiments are conducted on Office and VisDA datasets for both open-set and closed-set domain adaptation, and superior results are reported when comparing to the state-of-the-art approaches.

We present a new design, i.e., Self-Ensembling with Category-agnostic Clusters, for both closed-set and open-set domain adaptation.

We present Spectral Inference Networks, a framework for learning eigenfunctions of linear operators by stochastic optimization. Spectral Inference Networks generalize Slow Feature Analysis to generic symmetric operators, and are closely related to Variational Monte Carlo methods from computational physics. As such, they can be a powerful tool for unsupervised representation learning from video or graph-structured data. We cast training Spectral Inference Networks as a bilevel optimization problem, which allows for online learning of multiple eigenfunctions. We show results of training Spectral Inference Networks on problems in quantum mechanics and feature learning for videos on synthetic datasets. Our results demonstrate that Spectral Inference Networks accurately recover eigenfunctions of linear operators and can discover interpretable representations from video in a fully unsupervised manner.

We show how to learn spectral decompositions of linear operators with deep learning, and use it for unsupervised learning without a generative model.

The Tensor-Train factorization (TTF) is an efficient way to compress large weight matrices of fully-connected layers and recurrent layers in recurrent neural networks (RNNs). However, high Tensor-Train ranks for all the core tensors of parameters need to be element-wise fixed, which results in an unnecessary redundancy of model parameters. This work applies Riemannian stochastic gradient descent (RSGD) to train core tensors of parameters in the Riemannian Manifold before finding vectors of lower Tensor-Train ranks for parameters. The paper first presents the RSGD algorithm with a convergence analysis and then tests it on more advanced Tensor-Train RNNs such as bi-directional GRU/LSTM and Encoder-Decoder RNNs with a Tensor-Train attention model. The experiments on digit recognition and machine translation tasks suggest the effectiveness of the RSGD algorithm for Tensor-Train RNNs.

Applying the Riemannian SGD (RSGD) algorithm for training Tensor-Train RNNs to further reduce model parameters.

In this paper, we consider the problem of learning control policies that optimize areward function while satisfying constraints due to considerations of safety, fairness, or other costs. We propose a new algorithm - Projection Based ConstrainedPolicy Optimization (PCPO), an iterative method for optimizing policies in a two-step process - the first step performs an unconstrained update while the secondstep reconciles the constraint violation by projection the policy back onto the constraint set. We theoretically analyze PCPO and provide a lower bound on rewardimprovement, as well as an upper bound on constraint violation for each policy update. We further characterize the convergence of PCPO with projection basedon two different metrics - L2 norm and Kullback-Leibler divergence. Our empirical results over several control tasks demonstrate that our algorithm achievessuperior performance, averaging more than 3.5 times less constraint violation andaround 15% higher reward compared to state-of-the-art methods.

We propose a new algorithm that learns constraint-satisfying policies, and provide theoretical analysis and empirical demonstration in the context of reinforcement learning with constraints.

Deep networks face challenges of ensuring their robustness against inputs that cannot be effectively represented by information learned from training data. We attribute this vulnerability to the limitations inherent to activation-based representation. To complement the learned information from activation-based representation, we propose utilizing a gradient-based representation that explicitly focuses on missing information. In addition, we propose a directional constraint on the gradients as an objective during training to improve the characterization of missing information. To validate the effectiveness of the proposed approach, we compare the anomaly detection performance of gradient-based and activation-based representations. We show that the gradient-based representation outperforms the activation-based representation by 0.093 in CIFAR-10 and 0.361 in CURE-TSR datasets in terms of AUROC averaged over all classes. Also, we propose an anomaly detection algorithm that uses the gradient-based representation, denoted as GradCon, and validate its performance on three benchmarking datasets. The proposed method outperforms the majority of the state-of-the-art algorithms in CIFAR-10, MNIST, and fMNIST datasets with an average AUROC of 0.664, 0.973, and 0.934, respectively.

We propose a gradient-based representation for characterizing information that deep networks have not learned.

Medical images may contain various types of artifacts with different patterns and mixtures, which depend on many factors such as scan setting, machine condition, patients’ characteristics, surrounding environment, etc. However, existing deep learning based artifact reduction methods are restricted by their training set with specific predetermined artifact type and pattern. As such, they have limited clinical adoption. In this paper, we introduce a “Zero-Shot” medical image Artifact Reduction (ZSAR) framework, which leverages the power of deep learning but without using general pre-trained networks or any clean image reference. Specifically, we utilize the low internal visual entropy of an image and train a light-weight image-specific artifact reduction network to reduce artifacts in an image at test-time. We use Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) as vehicles to show that ZSAR can reduce artifacts better than state-of-the-art both qualitatively and quantitatively, while using shorter execution time. To the best of our knowledge, this is the first deep learning framework that reduces artifacts in medical images without using a priori training set.

We introduce a “Zero-Shot” medical image Artifact Reduction framework, which leverages the power of deep learning but without using general pre-trained networks or any clean image reference.

Attribution methods provide insights into the decision-making of machine learning models like artificial neural networks. For a given input sample, they assign a relevance score to each individual input variable, such as the pixels of an image. In this work we adapt the information bottleneck concept for attribution. By adding noise to intermediate feature maps we restrict the flow of information and can quantify (in bits) how much information image regions provide. We compare our method against ten baselines using three different metrics on VGG-16 and ResNet-50, and find that our methods outperform all baselines in five out of six settings. The method’s information-theoretic foundation provides an absolute frame of reference for attribution values (bits) and a guarantee that regions scored close to zero are not necessary for the network's decision.

We apply the informational bottleneck concept to attribution.

Recurrent Neural Networks (RNNs) are used in state-of-the-art models in domains such as speech recognition, machine translation, and language modelling. Sparsity is a technique to reduce compute and memory requirements of deep learning models. Sparse RNNs are easier to deploy on devices and high-end server processors. Even though sparse operations need less compute and memory relative to their dense counterparts, the speed-up observed by using sparse operations is less than expected on different hardware platforms. In order to address this issue, we investigate two different approaches to induce block sparsity in RNNs: pruning blocks of weights in a layer and using group lasso regularization with pruning to create blocks of weights with zeros. Using these techniques, we can create block-sparse RNNs with sparsity ranging from 80% to 90% with a small loss in accuracy. This technique allows us to reduce the model size by roughly 10x. Additionally, we can prune a larger dense network to recover this loss in accuracy while maintaining high block sparsity and reducing the overall parameter count. Our technique works with a variety of block sizes up to 32x32. Block-sparse RNNs eliminate overheads related to data storage and irregular memory accesses while increasing hardware efficiency compared to unstructured sparsity.

We show the RNNs can be pruned to induce block sparsity which improves speedup for sparse operations on existing hardware

Value iteration networks are an approximation of the value iteration (VI) algorithm implemented with convolutional neural networks to make VI fully differentiable. In this work, we study these networks in the context of robot motion planning, with a focus on applications to planetary rovers. The key challenging task in learning-based motion planning is to learn a transformation from terrain observations to a suitable navigation reward function. In order to deal with complex terrain observations and policy learning, we propose a value iteration recurrence, referred to as the soft value iteration network (SVIN). SVIN is designed to produce more effective training gradients through the value iteration network. It relies on a soft policy model, where the policy is represented with a probability distribution over all possible actions, rather than a deterministic policy that returns only the best action. We demonstrate the effectiveness of the proposed method in robot motion planning scenarios. In particular, we study the application of SVIN to very challenging problems in planetary rover navigation and present early training results on data gathered by the Curiosity rover that is currently operating on Mars.

We propose an improvement to value iteration networks, with applications to planetary rover path planning.

Transformer networks have lead to important progress in language modeling and machine translation. These models include two consecutive modules, a feed-forward layer and a self-attention layer. The latter allows the network to capture long term dependencies and are often regarded as the key ingredient in the success of Transformers. Building upon this intuition, we propose a new model that solely consists of attention layers. More precisely, we augment the self-attention layers with persistent memory vectors that play a similar role as the feed-forward layer. Thanks to these vectors, we can remove the feed-forward layer without degrading the performance of a transformer. Our evaluation shows the benefits brought by our model on standard character and word level language modeling benchmarks.

A novel attention layer that combines self-attention and feed-forward sublayers of Transformer networks.

This work views neural networks as data generating systems and applies anomalous pattern detection techniques on that data in order to detect when a network is processing a group of anomalous inputs. Detecting anomalies is a critical component for multiple machine learning problems including detecting the presence of adversarial noise added to inputs. More broadly, this work is a step towards giving neural networks the ability to detect groups of out-of-distribution samples. This work introduces ``Subset Scanning methods from the anomalous pattern detection domain to the task of detecting anomalous inputs to neural networks. Subset Scanning allows us to answer the question: "``Which subset of inputs have larger-than-expected activations at which subset of nodes? " Framing the adversarial detection problem this way allows us to identify systematic patterns in the activation space that span multiple adversarially noised images. Such images are ``"weird together". Leveraging this common anomalous pattern, we show increased detection power as the proportion of noised images increases in a test set. Detection power and accuracy results are provided for targeted adversarial noise added to CIFAR-10 images on a 20-layer ResNet using the Basic Iterative Method attack.

We efficiently find a subset of images that have higher than expected activations for some subset of nodes. These images appear more anomalous and easier to detect when viewed as a group.

Stability is a fundamental property of dynamical systems, yet to this date it has had little bearing on the practice of recurrent neural networks. In this work, we conduct a thorough investigation of stable recurrent models. Theoretically, we prove stable recurrent neural networks are well approximated by feed-forward networks for the purpose of both inference and training by gradient descent. Empirically, we demonstrate stable recurrent models often perform as well as their unstable counterparts on benchmark sequence tasks. Taken together, these findings shed light on the effective power of recurrent networks and suggest much of sequence learning happens, or can be made to happen, in the stable regime. Moreover, our results help to explain why in many cases practitioners succeed in replacing recurrent models by feed-forward models.

Stable recurrent models can be approximated by feed-forward networks and empirically perform as well as unstable models on benchmark tasks.

Weight-sharing plays a significant role in the success of many deep neural networks, by increasing memory efficiency and incorporating useful inductive priors about the problem into the network. But understanding how weight-sharing can be used effectively in general is a topic that has not been studied extensively. Chen et al. (2015) proposed HashedNets, which augments a multi-layer perceptron with a hash table, as a method for neural network compression. We generalize this method into a framework (ArbNets) that allows for efficient arbitrary weight-sharing, and use it to study the role of weight-sharing in neural networks. We show that common neural networks can be expressed as ArbNets with different hash functions. We also present two novel hash functions, the Dirichlet hash and the Neighborhood hash, and use them to demonstrate experimentally that balanced and deterministic weight-sharing helps with the performance of a neural network.

Studied the role of weight sharing in neural networks using hash functions, found that a balanced and deterministic hash function helps network performance.

We introduce Neural Markov Logic Networks (NMLNs), a statistical relational learning system that borrows ideas from Markov logic. Like Markov Logic Networks (MLNs), NMLNs are an exponential-family model for modelling distributions over possible worlds, but unlike MLNs, they do not rely on explicitly specified first-order logic rules. Instead, NMLNs learn an implicit representation of such rules as a neural network that acts as a potential function on fragments of the relational structure. Interestingly, any MLN can be represented as an NMLN. Similarly to recently proposed Neural theorem provers (NTPs) (Rocktaschel at al. 2017), NMLNs can exploit embeddings of constants but, unlike NTPs, NMLNs work well also in their absence. This is extremely important for predicting in settings other than the transductive one. We showcase the potential of NMLNs on knowledge-base completion tasks and on generation of molecular (graph) data.

We introduce a statistical relational learning system that borrows ideas from Markov logic but learns an implicit representation of rules as a neural network.

Using variational Bayes neural networks, we develop an algorithm capable of accumulating knowledge into a prior from multiple different tasks. This results in a rich prior capable of few-shot learning on new tasks. The posterior can go beyond the mean field approximation and yields good uncertainty on the performed experiments. Analysis on toy tasks show that it can learn from significantly different tasks while finding similarities among them. Experiments on Mini-Imagenet reach state of the art with 74.5% accuracy on 5 shot learning. Finally, we provide two new benchmarks, each showing a failure mode of existing meta learning algorithms such as MAML and prototypical Networks.

A scalable method for learning an expressive prior over neural networks across multiple tasks.

Sequential data often originates from diverse environments. Across them exist both shared regularities and environment specifics. To learn robust cross-environment descriptions of sequences we introduce disentangled state space models (DSSM). In the latent space of DSSM environment-invariant state dynamics is explicitly disentangled from environment-specific information governing that dynamics. We empirically show that such separation enables robust prediction, sequence manipulation and environment characterization. We also propose an unsupervised VAE-based training procedure to learn DSSM as Bayesian filters. In our experiments, we demonstrate state-of-the-art performance in controlled generation and prediction of bouncing ball video sequences across varying gravitational influences.

DISENTANGLED STATE SPACE MODELS

In this work, we approach one-shot and few-shot learning problems as methods for finding good prototypes for each class, where these prototypes are generalizable to new data samples and classes. We propose a metric learner that learns a Bregman divergence by learning its underlying convex function. Bregman divergences are a good candidate for this framework given they are the only class of divergences with the property that the best representative of a set of points is given by its mean. We propose a flexible extension to prototypical networks to enable joint learning of the embedding and the divergence, while preserving computational efficiency. Our preliminary results are comparable with the prior work on the Omniglot and Mini-ImageNet datasets, two standard benchmarks for one-shot and few-shot learning. We argue that our model can be used for other tasks that involve metric learning or tasks that require approximate convexity such as structured prediction and data completion.

Bregman divergence learning for few-shot learning.

Motivated by the flexibility of biological neural networks whose connectivity structure changes significantly during their lifetime,we introduce the Unrestricted Recursive Network (URN) and demonstrate that it can exhibit similar flexibility during training via gradient descent. We show empirically that many of the different neural network structures commonly used in practice today (including fully connected, locally connected and residual networks of differ-ent depths and widths) can emerge dynamically from the same URN.These different structures can be derived using gradient descent on a single general loss function where the structure of the data and the relative strengths of various regulator terms determine the structure of the emergent network. We show that this loss function and the regulators arise naturally when considering the symmetries of the network as well as the geometric properties of the input data.

We introduce a network framework which can modify its structure during training and show that it can converge to various ML network archetypes such as MLPs and LCNs.

We present CROSSGRAD , a method to use multi-domain training data to learn a classifier that generalizes to new domains. CROSSGRAD does not need an adaptation phase via labeled or unlabeled data, or domain features in the new domain. Most existing domain adaptation methods attempt to erase domain signals using techniques like domain adversarial training. In contrast, CROSSGRAD is free to use domain signals for predicting labels, if it can prevent overfitting on training domains. We conceptualize the task in a Bayesian setting, in which a sampling step is implemented as data augmentation, based on domain-guided perturbations of input instances. CROSSGRAD jointly trains a label and a domain classifier on examples perturbed by loss gradients of each other’s objectives. This enables us to directly perturb inputs, without separating and re-mixing domain signals while making various distributional assumptions. Empirical evaluation on three different applications where this setting is natural establishes that (1) domain-guided perturbation provides consistently better generalization to unseen domains, compared to generic instance perturbation methods, and (2) data augmentation is a more stable and accurate method than domain adversarial training.

Domain guided augmentation of data provides a robust and stable method of domain generalization

We present sketch-rnn, a recurrent neural network able to construct stroke-based drawings of common objects. The model is trained on a dataset of human-drawn images representing many different classes. We outline a framework for conditional and unconditional sketch generation, and describe new robust training methods for generating coherent sketch drawings in a vector format.

We investigate alternative to traditional pixel image modelling approaches, and propose a generative model for vector images.

Wilson et al. (2017) showed that, when the stepsize schedule is properly designed, stochastic gradient generalizes better than ADAM (Kingma & Ba, 2014). In light of recent work on hypergradient methods (Baydin et al., 2018), we revisit these claims to see if such methods close the gap between the most popular optimizers. As a byproduct, we analyze the true benefit of these hypergradient methods compared to more classical schedules, such as the fixed decay of Wilson et al. (2017). In particular, we observe they are of marginal help since their performance varies significantly when tuning their hyperparameters. Finally, as robustness is a critical quality of an optimizer, we provide a sensitivity analysis of these gradient based optimizers to assess how challenging their tuning is.

We provide a study trying to see how the recent online learning rate adaptation extends the conclusion made by Wilson et al. 2018 about adaptive gradient methods, along with comparison and sensitivity analysis.

Despite an ever growing literature on reinforcement learning algorithms and applications, much less is known about their statistical inference. In this paper, we investigate the large-sample behaviors of the Q-value estimates with closed-form characterizations of the asymptotic variances. This allows us to efficiently construct confidence regions for Q-value and optimal value functions, and to develop policies to minimize their estimation errors. This also leads to a policy exploration strategy that relies on estimating the relative discrepancies among the Q estimates. Numerical experiments show superior performances of our exploration strategy than other benchmark approaches.

We investigate the large-sample behaviors of the Q-value estimates and proposed an efficient exploration strategy that relies on estimating the relative discrepancies among the Q estimates.

We perform completely unsupervised one-sided image to image translation between a source domain $X$ and a target domain $Y$ such that we preserve relevant underlying shared semantics (e.g., class, size, shape, etc). In particular, we are interested in a more difficult case than those typically addressed in the literature, where the source and target are ``far" enough that reconstruction-style or pixel-wise approaches fail. We argue that transferring (i.e., \emph{translating}) said relevant information should involve both discarding source domain-specific information while incorporate target domain-specific information, the latter of which we model with a noisy prior distribution. In order to avoid the degenerate case where the generated samples are only explained by the prior distribution, we propose to minimize an estimate of the mutual information between the generated sample and the sample from the prior distribution. We discover that the architectural choices are an important factor to consider in order to preserve the shared semantic between $X$ and $Y$. We show state of the art results on the MNIST to SVHN task for unsupervised image to image translation.

We train an image to image translation network that take as input the source image and a sample from a prior distribution to generate a sample from the target distribution

Identifying salient points in images is a crucial component for visual odometry, Structure-from-Motion or SLAM algorithms. Recently, several learned keypoint methods have demonstrated compelling performance on challenging benchmarks. However, generating consistent and accurate training data for interest-point detection in natural images still remains challenging, especially for human annotators. We introduce IO-Net (i.e. InlierOutlierNet), a novel proxy task for the self-supervision of keypoint detection, description and matching. By making the sampling of inlier-outlier sets from point-pair correspondences fully differentiable within the keypoint learning framework, we show that are able to simultaneously self-supervise keypoint description and improve keypoint matching. Second, we introduce KeyPointNet, a keypoint-network architecture that is especially amenable to robust keypoint detection and description. We design the network to allow local keypoint aggregation to avoid artifacts due to spatial discretizations commonly used for this task, and we improve fine-grained keypoint descriptor performance by taking advantage of efficient sub-pixel convolutions to upsample the descriptor feature-maps to a higher operating resolution. Through extensive experiments and ablative analysis, we show that the proposed self-supervised keypoint learning method greatly improves the quality of feature matching and homography estimation on challenging benchmarks over the state-of-the-art.

Learning to extract distinguishable keypoints from a proxy task, outlier rejection.

We study the role of intrinsic motivation as an exploration bias for reinforcement learning in sparse-reward synergistic tasks, which are tasks where multiple agents must work together to achieve a goal they could not individually. Our key idea is that a good guiding principle for intrinsic motivation in synergistic tasks is to take actions which affect the world in ways that would not be achieved if the agents were acting on their own. Thus, we propose to incentivize agents to take (joint) actions whose effects cannot be predicted via a composition of the predicted effect for each individual agent. We study two instantiations of this idea, one based on the true states encountered, and another based on a dynamics model trained concurrently with the policy. While the former is simpler, the latter has the benefit of being analytically differentiable with respect to the action taken. We validate our approach in robotic bimanual manipulation tasks with sparse rewards; we find that our approach yields more efficient learning than both 1) training with only the sparse reward and 2) using the typical surprise-based formulation of intrinsic motivation, which does not bias toward synergistic behavior. Videos are available on the project webpage: https://sites.google.com/view/iclr2020-synergistic.

We propose a formulation of intrinsic motivation that is suitable as an exploration bias in multi-agent sparse-reward synergistic tasks, by encouraging agents to affect the world in ways that would not be achieved if they were acting individually.

A general graph-structured neural network architecture operates on graphs through two core components: (1) complex enough message functions; (2) a fixed information aggregation process. In this paper, we present the Policy Message Passing algorithm, which takes a probabilistic perspective and reformulates the whole information aggregation as stochastic sequential processes. The algorithm works on a much larger search space, utilizes reasoning history to perform inference, and is robust to noisy edges. We apply our algorithm to multiple complex graph reasoning and prediction tasks and show that our algorithm consistently outperforms state-of-the-art graph-structured models by a significant margin.

An probabilistic inference algorithm driven by neural network for graph-structured models

Deep multitask networks, in which one neural network produces multiple predictive outputs, are more scalable and often better regularized than their single-task counterparts. Such advantages can potentially lead to gains in both speed and performance, but multitask networks are also difficult to train without finding the right balance between tasks. We present a novel gradient normalization (GradNorm) technique which automatically balances the multitask loss function by directly tuning the gradients to equalize task training rates. We show that for various network architectures, for both regression and classification tasks, and on both synthetic and real datasets, GradNorm improves accuracy and reduces overfitting over single networks, static baselines, and other adaptive multitask loss balancing techniques. GradNorm also matches or surpasses the performance of exhaustive grid search methods, despite only involving a single asymmetry hyperparameter $\alpha$. Thus, what was once a tedious search process which incurred exponentially more compute for each task added can now be accomplished within a few training runs, irrespective of the number of tasks. Ultimately, we hope to demonstrate that gradient manipulation affords us great control over the training dynamics of multitask networks and may be one of the keys to unlocking the potential of multitask learning.

We show how you can boost performance in a multitask network by tuning an adaptive multitask loss function that is learned through directly balancing network gradients.

Image segmentation aims at grouping pixels that belong to the same object or region. At the heart of image segmentation lies the problem of determining whether a pixel is inside or outside a region, which we denote as the "insideness" problem. Many Deep Neural Networks (DNNs) variants excel in segmentation benchmarks, but regarding insideness, they have not been well visualized or understood: What representations do DNNs use to address the long-range relationships of insideness? How do architectural choices affect the learning of these representations? In this paper, we take the reductionist approach by analyzing DNNs solving the insideness problem in isolation, i.e. determining the inside of closed (Jordan) curves. We demonstrate analytically that state-of-the-art feed-forward and recurrent architectures can implement solutions of the insideness problem for any given curve. Yet, only recurrent networks could learn these general solutions when the training enforced a specific "routine" capable of breaking down the long-range relationships. Our results highlights the need for new training strategies that decompose the learning into appropriate stages, and that lead to the general class of solutions necessary for DNNs to understand insideness.

DNNs for image segmentation can implement solutions for the insideness problem but only some recurrent nets could learn them with a specific type of supervision.

We address the challenging problem of deep representation learning--the efficient adaption of a pre-trained deep network to different tasks. Specifically, we propose to explore gradient-based features. These features are gradients of the model parameters with respect to a task-specific loss given an input sample. Our key innovation is the design of a linear model that incorporates both gradient features and the activation of the network. We show that our model provides a local linear approximation to a underlying deep model, and discuss important theoretical insight. Moreover, we present an efficient algorithm for the training and inference of our model without computing the actual gradients. Our method is evaluated across a number of representation learning tasks on several datasets and using different network architectures. We demonstrate strong results in all settings. And our results are well-aligned with our theoretical insight.

Given a pre-trained model, we explored the per-sample gradients of the model parameters relative to a task-specific loss, and constructed a linear model that combines gradients of model parameters and the activation of the model.

Recovering 3D geometry shape, albedo and lighting from a single image has wide applications in many areas, which is also a typical ill-posed problem. In order to eliminate the ambiguity, face prior knowledge like linear 3D morphable models (3DMM) learned from limited scan data are often adopted to the reconstruction process. However, methods based on linear parametric models cannot generalize well for facial images in the wild with various ages, ethnicity, expressions, poses, and lightings. Recent methods aim to learn a nonlinear parametric model using convolutional neural networks (CNN) to regress the face shape and texture directly. However, the models were only trained on a dataset that is generated from a linear 3DMM. Moreover, the identity and expression representations are entangled in these models, which hurdles many facial editing applications. In this paper, we train our model with adversarial loss in a semi-supervised manner on hybrid batches of unlabeled and labeled face images to exploit the value of large amounts of unlabeled face images from unconstrained photo collections. A novel center loss is introduced to make sure that different facial images from the same person have the same identity shape and albedo. Besides, our proposed model disentangles identity, expression, pose, and lighting representations, which improves the overall reconstruction performance and facilitates facial editing applications, e.g., expression transfer. Comprehensive experiments demonstrate that our model produces high-quality reconstruction compared to state-of-the-art methods and is robust to various expression, pose, and lighting conditions.

We train our face reconstruction model with adversarial loss in semi-supervised manner on hybrid batches of unlabeled and labeled face images to exploit the value of large amounts of unlabeled face images from unconstrained photo collections.

Human conversations naturally evolve around related entities and connected concepts, while may also shift from topic to topic. This paper presents ConceptFlow, which leverages commonsense knowledge graphs to explicitly model such conversation flows for better conversation response generation. ConceptFlow grounds the conversation inputs to the latent concept space and represents the potential conversation flow as a concept flow along the commonsense relations. The concept is guided by a graph attention mechanism that models the possibility of the conversation evolving towards different concepts. The conversation response is then decoded using the encodings of both utterance texts and concept flows, integrating the learned conversation structure in the concept space. Our experiments on Reddit conversations demonstrate the advantage of ConceptFlow over previous commonsense aware dialog models and fine-tuned GPT-2 models, while using much fewer parameters but with explicit modeling of conversation structures.

This paper presents ConceptFlow that explicitly models the conversation flow in commonsense knowledge graph for better conversation generation.

Biological neural networks face homeostatic and resource constraints that restrict the allowed configurations of connection weights. If a constraint is tight it defines a very small solution space, and the size of these constraint spaces determines their potential overlap with the solutions for computational tasks. We study the geometry of the solution spaces for constraints on neurons' total synaptic weight and on individual synaptic weights, characterizing the connection degrees (numbers of partners) that maximize the size of these solution spaces. We then hypothesize that the size of constraints' solution spaces could serve as a cost function governing neural circuit development. We develop analytical approximations and bounds for the model evidence of the maximum entropy degree distributions under these cost functions. We test these on a published electron microscopic connectome of an associative learning center in the fly brain, finding evidence for a developmental progression in circuit structure.

We examine the hypothesis that the entropy of solution spaces for constraints on synaptic weights (the "flexibility" of the constraint) could serve as a cost function for neural circuit development.

In this preliminary work, we study the generalization properties of infinite ensembles of infinitely-wide neural networks. Amazingly, this model family admits tractable calculations for many information-theoretic quantities. We report analytical and empirical investigations in the search for signals that correlate with generalization.

Infinite ensembles of infinitely wide neural networks are an interesting model family from an information theoretic perspective.

Learning multilingual representations of text has proven a successful method for many cross-lingual transfer learning tasks. There are two main paradigms for learning such representations: (1) alignment, which maps different independently trained monolingual representations into a shared space, and (2) joint training, which directly learns unified multilingual representations using monolingual and cross-lingual objectives jointly. In this paper, we first conduct direct comparisons of representations learned using both of these methods across diverse cross-lingual tasks. Our empirical results reveal a set of pros and cons for both methods, and show that the relative performance of alignment versus joint training is task-dependent. Stemming from this analysis, we propose a simple and novel framework that combines these two previously mutually-exclusive approaches. Extensive experiments on various tasks demonstrate that our proposed framework alleviates limitations of both approaches, and outperforms existing methods on the MUSE bilingual lexicon induction (BLI) benchmark. We further show that our proposed framework can generalize to contextualized representations and achieves state-of-the-art results on the CoNLL cross-lingual NER benchmark.

We conduct a comparative study of cross-lingual alignment vs joint training methods and unify these two previously exclusive paradigms in a new framework.

Large number of weights in deep neural networks make the models difficult to be deployed in low memory environments such as, mobile phones, IOT edge devices as well as "inferencing as a service" environments on the cloud. Prior work has considered reduction in the size of the models, through compression techniques like weight pruning, filter pruning, etc. or through low-rank decomposition of the convolution layers. In this paper, we demonstrate the use of multiple techniques to achieve not only higher model compression but also reduce the compute resources required during inferencing. We do filter pruning followed by low-rank decomposition using Tucker decomposition for model compression. We show that our approach achieves upto 57\% higher model compression when compared to either Tucker Decomposition or Filter pruning alone at similar accuracy for GoogleNet. Also, it reduces the Flops by upto 48\% thereby making the inferencing faster.

Combining orthogonal model compression techniques to get significant reduction in model size and number of flops required during inferencing.

We review the limitations of BLEU and ROUGE -- the most popular metrics used to assess reference summaries against hypothesis summaries, and introduce JAUNE: a set of criteria for what a good metric should behave like and propose concrete ways to use recent Transformers-based Language Models to assess reference summaries against hypothesis summaries.

Introduces JAUNE: a methodology to replace BLEU and ROUGE score with multidimensional, model-based evaluators for assessing summaries

This paper presents a new Graph Neural Network (GNN) type using feature-wise linear modulation (FiLM). Many standard GNN variants propagate information along the edges of a graph by computing ``messages'' based only on the representation of the source of each edge. In GNN-FiLM, the representation of the target node of an edge is additionally used to compute a transformation that can be applied to all incoming messages, allowing feature-wise modulation of the passed information. Results of experiments comparing different GNN architectures on three tasks from the literature are presented, based on re-implementations of baseline methods. Hyperparameters for all methods were found using extensive search, yielding somewhat surprising results: differences between baseline models are smaller than reported in the literature. Nonetheless, GNN-FiLM outperforms baseline methods on a regression task on molecular graphs and performs competitively on other tasks.

new GNN formalism + extensive experiments; showing differences between GGNN/GCN/GAT are smaller than thought

To deal simultaneously with both, the attributed network embedding and clustering, we propose a new model. It exploits both content and structure information, capitalising on their simultaneous use. The proposed model relies on the approximation of the relaxed continuous embedding solution by the true discrete clustering one. Thereby, we show that incorporating an embedding representation provides simpler and more interpretable solutions. Experiment results demonstrate that the proposed algorithm performs better, in terms of clustering and embedding, than the state-of-art algorithms, including deep learning methods devoted to similar tasks for attributed network datasets with different proprieties.

This paper propose a novel matrix decomposition framework for simultaneous attributed network data embedding and clustering.

We propose a learned image-guided rendering technique that combines the benefits of image-based rendering and GAN-based image synthesis. The goal of our method is to generate photo-realistic re-renderings of reconstructed objects for virtual and augmented reality applications (e.g., virtual showrooms, virtual tours and sightseeing, the digital inspection of historical artifacts). A core component of our work is the handling of view-dependent effects. Specifically, we directly train an object-specific deep neural network to synthesize the view-dependent appearance of an object. As input data we are using an RGB video of the object. This video is used to reconstruct a proxy geometry of the object via multi-view stereo. Based on this 3D proxy, the appearance of a captured view can be warped into a new target view as in classical image-based rendering. This warping assumes diffuse surfaces, in case of view-dependent effects, such as specular highlights, it leads to artifacts. To this end, we propose EffectsNet, a deep neural network that predicts view-dependent effects. Based on these estimations, we are able to convert observed images to diffuse images. These diffuse images can be projected into other views. In the target view, our pipeline reinserts the new view-dependent effects. To composite multiple reprojected images to a final output, we learn a composition network that outputs photo-realistic results. Using this image-guided approach, the network does not have to allocate capacity on ``remembering'' object appearance, instead it learns how to combine the appearance of captured images. We demonstrate the effectiveness of our approach both qualitatively and quantitatively on synthetic as well as on real data.

We propose a learned image-guided rendering technique that combines the benefits of image-based rendering and GAN-based image synthesis while considering view-dependent effects.

We evaluate the distribution learning capabilities of generative adversarial networks by testing them on synthetic datasets. The datasets include common distributions of points in $R^n$ space and images containing polygons of various shapes and sizes. We find that by and large GANs fail to faithfully recreate point datasets which contain discontinous support or sharp bends with noise. Additionally, on image datasets, we find that GANs do not seem to learn to count the number of objects of the same kind in an image. We also highlight the apparent tension between generalization and learning in GANs.

GANs are evaluated on synthetic datasets