Recent works have demonstrated that deep learning on graphs is vulnerable to adversarial attacks, in that imperceptible perturbations on input data can lead to dramatic performance deterioration. In this paper, we focus on the underlying problem of learning robust representations on graphs via mutual information. In contrast to previous works measure the task-specific robustness based on the label space, we here take advantage of the representation space to study a task-free robustness measure given the joint input space w.r.t graph topology and node attributes. We formulate this problem as a constrained saddle point optimization problem and solve it efficiently in a reduced search space. Furthermore, we provably establish theoretical connections between our task-free robustness measure and the robustness of downstream classifiers. Extensive experiments demonstrate that our proposed method is able to enhance robustness against adversarial attacks on graphs, yet even increases natural accuracy.

Photoacoustic computed tomography (PACT) reconstructs the initial pressure distribution from raw PA signals. Standard reconstruction always induces artifacts using limited-view signals, which are influenced by limited angle coverage of transducers, finite bandwidth, and uncertain heterogeneous biological tissue. Recently, supervised deep learning has been used to overcome limited-view problem that requires ground-truth. However, even full-view sampling still induces artifacts that cannot be used to train the model. It causes a dilemma that we could not acquire perfect ground-truth in practice. To reduce the dependence on the quality of ground-truth, in this paper, for the first time, we propose a beyond supervised reconstruction framework (BSR-Net) based on deep learning to compensate the limited-view issue by feeding limited-view position-wise data. A quarter position-wise data is fed into model and outputs a group full-view data. Specifically, our method introduces a residual structure, which generates beyond supervised reconstruction result, whose artifacts are drastically reduced in the output compared to ground-truth. Moreover, two novel losses are designed to restrain the artifacts. The numerical and in-vivo results have demonstrated the performance of our method to reconstruct the full-view image without artifacts.

Learning task-agnostic dynamics models in high-dimensional observation spaces can be challenging for model-based RL agents. We propose a novel way to learn latent world models by learning to predict sequences of future actions conditioned on task completion. These task-conditioned models adaptively focus modeling capacity on task-relevant dynamics, while simultaneously serving as an effective heuristic for planning with sparse rewards. We evaluate our method on challenging visual goal completion tasks and show a substantial increase in performance compared to prior model-free approaches. Deep reinforcement learning has proven to be a powerful and effective framework for solving a diversity of challenging decision-making problems (Silver et al., 2017a; Berner et al., 2019). However these algorithms are typically trained to maximize a single reward function, ignoring information that is not directly relevant to the associated task at hand. This way of learning is in stark contrast to how humans learn (Tenenbaum, 2018). Without being prompted by a specific task, humans can still explore their environment, practice achieving imaginary goals, and in so doing learn about the dynamics of the environment. When subsequently presented with a novel task, humans can utilize this learned knowledge to bootstrap learning -- a property we would like our artificial agents to have. In this work, we investigate one way to bridge this gap by learning world models (Ha & Schmidhuber, 2018) that enable the realization of previously unseen tasks. By modeling the task-agnostic dynamics of an environment, an agent can make predictions about how its own actions may affect the environment state without the need for additional samples from the environment. Prior work has shown that by using powerful function approximators to model environment dynamics, training an agent entirely within its own world models can result in large gains in sample efficiency (Ha & Schmidhuber, 2018).

Thanks to the use of convolution and pooling layers, convolutional neural networks were for a long time thought to be shift-invariant. However, recent works have shown that the output of a CNN can change significantly with small shifts in input: a problem caused by the presence of downsampling (stride) layers. The existing solutions rely either on data augmentation or on anti-aliasing, both of which have limitations and neither of which enables perfect shift invariance. Additionally, the gains obtained from these methods do not extend to image patterns not seen during training. To address these challenges, we propose adaptive polyphase sampling (APS), a simple sub-sampling scheme that allows convolutional neural networks to achieve 100% consistency in classification performance under shifts, without any loss in accuracy. With APS the networks exhibit perfect consistency to shifts even before training, making it the first approach that makes convolutional neural networks truly shift invariant.

Bayesian neural networks (BNNs) have been long considered an ideal, yet unscalable solution for improving the robustness and the predictive uncertainty of deep neural networks. While they could capture more accurately the posterior distribution of the network parameters, most BNN approaches are either limited to small networks or rely on constraining assumptions such as parameter independence. These drawbacks have enabled prominence of simple, but computationally heavy approaches such as Deep Ensembles, whose training and testing costs increase linearly with the number of networks. In this work we aim for efficient deep BNNs amenable to complex computer vision architectures, e.g. ResNet50 DeepLabV3+, and tasks, e.g. semantic segmentation, with fewer assumptions on the parameters. We achieve this by leveraging variational autoencoders (VAEs) to learn the interaction and the latent distribution of the parameters at each network layer. Our approach, Latent-Posterior BNN (LP-BNN), is compatible with the recent BatchEnsemble method, leading to highly efficient ({in terms of computation and} memory during both training and testing) ensembles. LP-BNN s attain competitive results across multiple metrics in several challenging benchmarks for image classification, semantic segmentation and out-of-distribution detection.

Though deep learning models have taken on commercial and political relevance, many aspects of their training and operation remain poorly understood. This has sparked interest in "science of deep learning" projects, many of which are run at scale and require enormous amounts of time, money, and electricity. But how much of this research really needs to occur at scale? In this paper, we introduce MNIST-1D: a minimalist, low-memory, and low-compute alternative to classic deep learning benchmarks. The training examples are 20 times smaller than MNIST examples yet they differentiate more clearly between linear, nonlinear, and convolutional models which attain 32, 68, and 94% accuracy respectively (these models obtain 94, 99+, and 99+% on MNIST). Then we present example use cases which include measuring the spatial inductive biases of lottery tickets, observing deep double descent, and metalearning an activation function.

The current dominant paradigm in sensorimotor control, whether imitation or reinforcement learning, is to train policies directly in raw action spaces such as torque, joint angle, or end-effector position. This forces the agent to make decisions individually at each timestep in training, and hence, limits the scalability to continuous, high-dimensional, and long-horizon tasks. In contrast, research in classical robotics has, for a long time, exploited dynamical systems as a policy representation to learn robot behaviors via demonstrations. These techniques, however, lack the flexibility and generalizability provided by deep learning or reinforcement learning and have remained under-explored in such settings. In this work, we begin to close this gap and embed the structure of a dynamical system into deep neural network-based policies by reparameterizing action spaces via second-order differential equations. We propose Neural Dynamic Policies (NDPs) that make predictions in trajectory distribution space as opposed to prior policy learning methods where actions represent the raw control space. The embedded structure allows end-to-end policy learning for both reinforcement and imitation learning setups. We show that NDPs outperform the prior state-of-the-art in terms of either efficiency or performance across several robotic control tasks for both imitation and reinforcement learning setups.

In the recent past, psychological stress has been increasingly observed in humans, and early detection is crucial to prevent health risks. Stress detection using on-device deep learning algorithms has been on the rise owing to advancements in pervasive computing. However, an important challenge that needs to be addressed is handling unlabeled data in real-time via suitable ground truthing techniques (like Active Learning), which should help establish affective states (labels) while also selecting only the most informative data points to query from an oracle. In this paper, we propose a framework with capabilities to represent model uncertainties through approximations in Bayesian Neural Networks using Monte-Carlo (MC) Dropout. This is combined with suitable acquisition functions for active learning. Empirical results on a popular stress and affect detection dataset experimented on a Raspberry Pi 2 indicate that our proposed framework achieves a considerable efficiency boost during inference, with a substantially low number of acquired pool points during active learning across various acquisition functions. Variation Ratios achieves an accuracy of 90.38% which is comparable to the maximum test accuracy achieved while training on about 40% lesser data.

In this paper, we propose a learning algorithm that enables a model to quickly exploit commonalities among related tasks from an unseen task distribution, before quickly adapting to specific tasks from that same distribution. We investigate how learning with different task distributions can first improve adaptability by meta-finetuning on related tasks before improving goal task generalization with finetuning. Synthetic regression experiments validate the intuition that learning to meta-learn improves adaptability and consecutively generalization. The methodology, setup, and hypotheses in this proposal were positively evaluated by peer review before conclusive experiments were carried out.

The success of deep learning models has led to a lot of recent interest in understanding the properties of "interpolating" neural network models, that achieve (near-)zero training loss [Zha 17a; Bel 19]. One aspect of understanding these models is to theoretically characterize how first-order gradient methods (with appropriate random initialization) seem to reliably find interpolating solutions to non-convex optimization problems. In this paper, we show that, under two sets of conditions, training fixed-width two-layer networks with gradient descent drives the logistic loss to zero. The networks have smooth "Huberized" ReLUs [Tat 20, see (1) and Figure 1] and the output weights are not trained. The first result only requires the assumption that the initial loss is small, but does not require any assumption about either the width of the network or the number of samples. It guarantees that if the initial loss is small then gradient descent drives the logistic loss to zero. For our second result we assume that the inputs come from four clusters, two per class, and that the clusters corresponding to the opposite labels are appropriately separated. Under these assumptions, we show that random Gaussian initialization along with a single step of gradient descent is enough to guarantee that the loss reduces sufficiently that the first result applies. A few proof ideas that facilitate our results are as follows: under our first set of assumptions, when the loss is small, we show that the negative gradient aligns well with the parameter vector. 1