We investigate deep Bayesian neural networks with Gaussian priors on the weights and ReLU-like nonlinearities, shedding light on novel sparsity-inducing mechanisms at the level of the units of the network, both pre- and post-nonlinearities. The main thrust of the paper is to establish that the units prior distribution becomes increasingly heavy-tailed with depth. We show that first layer units are Gaussian, second layer units are sub-Exponential, and we introduce sub-Weibull distributions to characterize the deeper layers units. Bayesian neural networks with Gaussian priors are well known to induce the weight decay penalty on the weights. In contrast, our result indicates a more elaborate regularisation scheme at the level of the units, ranging from convex penalties for the first two layers - weight decay for the first and Lasso for the second - to non convex penalties for deeper layers. Thus, despite weight decay does not allow for the weights to be set exactly to zero, sparse solutions tend to be selected for the units from the second layer onward. This result provides new theoretical insight on deep Bayesian neural networks, underpinning their natural shrinkage properties and practical potential.

We present the Tamed Cross Entropy (TCE) loss function, a robust derivative of the standard Cross Entropy (CE) loss used in deep learning for classification tasks. However, unlike other robust losses, the TCE loss is designed to exhibit the same training properties than the CE loss in noiseless scenarios. Therefore, the TCE loss requires no modification on the training regime compared to the CE loss and, in consequence, can be applied in all applications where the CE loss is currently used. We evaluate the TCE loss using the ResNet architecture on four image datasets that we artificially contaminated with various levels of label noise. The TCE loss outperforms the CE loss in every tested scenario.

Abstract--Diverse fault types, fast re-closures and complicated transient states after a fault event make real-time fault location in power grids challenging. Existing localization techniques in this area rely on simplistic assumptions, such as static loads, or require much higher sampling rates or total measurement availability. This paper proposes a data-driven localization method based on a Convolutional Neural Network (CNN) classifier using bus voltages. Unlike prior data-driven methods, the proposed classifier is based on features with physical interpretations that are described in details. The accuracy of our CNN based localization tool is demonstrably superior to other machine learning classifiers in the literature. T o further improve the location performance, a novel phasor measurement units (PMU) placement strategy is proposed and validated against other methods. A significant aspect of our methodology is that under very low observability ( 7% of buses), the algorithm is still able to localize the faulted line to a small neighborhood with high probability. The performance of our scheme is validated through simulations of faults of various types in the IEEE 68-bus power system under varying load conditions, system observability and measurement quality. Efficient fault localization is an integral part of the system restoration, and it is necessary for improving power system stability and reliability.

Network pruning is widely used for reducing the heavy computational cost of deep models. A typical pruning algorithm is a three-stage pipeline, i.e., training (a large model), pruning and fine-tuning. During pruning, according to a certain criterion, redundant weights are pruned and important weights are kept to best preserve the accuracy. In this work, we make several surprising observations which contradict common beliefs. For all the six state-of-the-art pruning algorithms we examined, fine-tuning a pruned model only gives comparable or even worse performance than training that model with randomly initialized weights. For pruning algorithms which assume a predefined target network architecture, one can get rid of the full pipeline and directly train the target network from scratch. Our observations are consistent for a wide variety of pruning algorithms with multiple network architectures, datasets, and tasks. Our results have several implications: 1) training a large, over-parameterized model is not necessary to obtain an efficient final model, 2) learned "important" weights of the large model are not necessarily useful for the small pruned model, 3) the pruned architecture itself, rather than a set of inherited "important" weights, is what leads to the efficiency benefit in the final model, which suggests that some pruning algorithms could be seen as performing network architecture search. Over-parameterization is a widely-recognized property of deep neural networks (Denton et al., 2014; Ba & Caruana, 2014), which leads to high computational cost and high memory footprint.

Generative Adversarial Networks (GANs) have experienced a recent surge in popularity, performing competitively in a variety of tasks, especially in computer vision. However, GAN training has shown limited success in natural language processing. This is largely because sequences of text are discrete, and thus gradients cannot propagate from the discriminator to the generator. Recent solutions use reinforcement learning to propagate approximate gradients to the generator, but this is inefficient to train. We propose to utilize an autoencoder to learn a low-dimensional representation of sentences. A GAN is then trained to generate its own vectors in this space, which decode to realistic utterances. We report both random and interpolated samples from the generator. Visualization of sentence vectors indicate our model correctly learns the latent space of the autoencoder. Both human ratings and BLEU scores show that our model generates realistic text against competitive baselines.

We propose a segmentation framework that uses deep neural networks and introduce two innovations. First, we describe a biophysics-based domain adaptation method. Second, we propose an automatic method to segment white and gray matter, and cerebrospinal fluid, in addition to tumorous tissue. Regarding our first innovation, we use a domain adaptation framework that combines a novel multispecies biophysical tumor growth model with a generative adversarial model to create realistic looking synthetic multimodal MR images with known segmentation. Regarding our second innovation, we propose an automatic approach to enrich available segmentation data by computing the segmentation for healthy tissues. This segmentation, which is done using diffeomorphic image registration between the BraTS training data and a set of prelabeled atlases, provides more information for training and reduces the class imbalance problem. Our overall approach is not specific to any particular neural network and can be used in conjunction with existing solutions. We demonstrate the performance improvement using a 2D U-Net for the BraTS'18 segmentation challenge. Our biophysics based domain adaptation achieves better results, as compared to the existing state-of-the-art GAN model used to create synthetic data for training.

We present Piano Genie, an intelligent controller which allows non-musicians to improvise on the piano. With Piano Genie, a user performs on a simple interface with eight buttons, and their performance is decoded into the space of plausible piano music in real time. To learn a suitable mapping procedure for this problem, we train recurrent neural network autoencoders with discrete bottlenecks: an encoder learns an appropriate sequence of buttons corresponding to a piano piece, and a decoder learns to map this sequence back to the original piece. During performance, we substitute a user's input for the encoder output, and play the decoder's prediction each time the user presses a button. To improve the interpretability of Piano Genie's performance mechanics, we impose musically-salient constraints over the encoder's outputs.

Data augmentation is commonly used to encode invariances in learning methods. However, this process is often performed in an inefficient manner, as artificial examples are created by applying a number of transformations to all points in the training set. The resulting explosion of the dataset size can be an issue in terms of storage and training costs, as well as in selecting and tuning the optimal set of transformations to apply. In this work, we demonstrate that it is possible to significantly reduce the number of data points included in data augmentation while realizing the same accuracy and invariance benefits of augmenting the entire dataset. We propose a novel set of subsampling policies, based on model influence and loss, that can achieve a 90% reduction in augmentation set size while maintaining the accuracy gains of standard data augmentation.

Highly expressive models such as deep neural networks (DNNs) have been widely applied to various applications and achieved increasing success. However, recent studies show that such machine learning models appear to be vulnerable against adversarial examples. So far adversarial examples have been heavily explored for 2D images, while few works have conducted to understand vulnerabilities of 3D objects which exist in real world, where 3D objects are projected to 2D domains by photo taking for different learning (recognition) tasks. In this paper, we consider adversarial behaviors in practical scenarios by manipulating the shape and texture of a given 3D mesh representation of an object. Our goal is to project the optimized "adversarial meshes" to 2D with a photorealistic renderer, and still able to mislead different machine learning models. Extensive experiments show that by generating unnoticeable 3D adversarial perturbation on shape or texture for a 3D mesh, the corresponding projected 2D instance can either lead classifiers to misclassify the victim object as an arbitrary malicious target, or hide any target object within the scene from object detectors. We conduct human studies to show that our optimized adversarial 3D perturbation is highly unnoticeable for human vision systems. In addition to the subtle perturbation for a given 3D mesh, we also propose to synthesize a realistic 3D mesh and put in a scene mimicking similar rendering conditions and therefore attack different machine learning models. In-depth analysis of transferability among various 3D renderers and vulnerable regions of meshes are provided to help better understand adversarial behaviors in real-world.

A key question for machine learning approaches in particle physics is how to best represent and learn from collider events. As an event is intrinsically a variable-length unordered set of particles, we build upon recent machine learning efforts to learn directly from sets of features. Adapting and specializing the "Deep Sets" framework to particle physics, we introduce Energy Flow Networks, which respect infrared and collinear safety by construction. We also develop Particle Flow Networks, which allow for general energy dependence and the inclusion of additional particle-level information such as charge and flavor. These networks feature a per-particle internal (latent) representation, and summing over all particles yields an overall event-level latent representation. We show how this latent space decomposition unifies existing event representations based on detector images and radiation moments. To demonstrate the power and simplicity of this set-based approach, we apply these networks to the collider task of discriminating quark jets from gluon jets, finding similar or improved performance compared to existing methods. We also show how the learned event representation can be directly visualized, providing insight into the inner workings of the model. These architectures lend themselves to efficiently processing and analyzing events for a wide variety of tasks at the Large Hadron Collider. Implementations and examples of our architectures are available online in our EnergyFlow package.