Besides accuracy, the storage of convolutional neural networks (CNN) models is another important factor considering limited hardware resources in practical applications. For example, autonomous driving requires the design of accurate yet fast CNN for low latency in object detection and classification. To fulfill the need, we aim at obtaining CNN models with both high testing accuracy and small size/storage to address resource constraints in many embedded systems. In particular, this paper focuses on proposing a generic reinforcement learning based model compression approach in a two-stage compression pipeline: pruning and quantization. The first stage of compression, i.e., pruning, is achieved via exploiting deep reinforcement learning (DRL) to co-learn the accuracy of CNN models updated after layer-wise channel pruning on a testing dataset and the FLOPs, number of floating point operations in each layer, updated after kernel-wise variational pruning using information dropout. Layer-wise channel pruning is to remove unimportant kernels from the input channel dimension while kernel-wise variational pruning is to remove unimportant kernels from the 2D-kernel dimensions, namely, height and width. The second stage, i.e., quantization, is achieved via a similar DRL approach but focuses on obtaining the optimal weight bits for individual layers. We further conduct experimental results on CIFAR-10 and ImageNet datasets. For the CIFAR-10 dataset, the proposed method can reduce the size of VGGNet by 9x from 20.04MB to 2.2MB with 0.2% accuracy increase. For the ImageNet dataset, the proposed method can reduce the size of VGG-16 by 33x from 138MB to 4.14MB with no accuracy loss.

Generalization of deep networks has been of great interest in recent years, resulting in a number of theoretically and empirically motivated complexity measures. However, most papers proposing such measures study only a small set of models, leaving open the question of whether the conclusion drawn from those experiments would remain valid in other settings. We present the first large scale study of generalization in deep networks. We investigate more then 40 complexity measures taken from both theoretical bounds and empirical studies. We train over 10,000 convolutional networks by systematically varying commonly used hyperparameters. Hoping to uncover potentially causal relationships between each measure and generalization, we analyze carefully controlled experiments and show surprising failures of some measures as well as promising measures for further research.

Achieving fusion of deep learning with combinatorial algorithms promises transformative changes to artificial intelligence. One possible approach is to introduce combinatorial building blocks into neural networks. Such end-to-end architectures have the potential to tackle combinatorial problems on raw input data such as ensuring global consistency in multi-object tracking or route planning on maps in robotics. In this work, we present a method that implements an efficient backward pass through blackbox implementations of combinatorial solvers with linear objective functions. We provide both theoretical and experimental backing. In particular, we incorporate the Gurobi MIP solver, Blossom V algorithm, and Dijkstra's algorithm into architectures that extract suitable features from raw inputs for the traveling salesman problem, the min-cost perfect matching problem and the shortest path problem.

The genomic profile underlying an individual tumor can be highly informative in the creation of a personalized cancer treatment strategy for a given patient; a practice known as precision oncology. This involves next generation sequencing of a tumor sample and the subsequent identification of genomic aberrations, such as somatic mutations, to provide potential candidates of targeted therapy. The identification of these aberrations from sequencing noise and germline variant background poses a classic classification-style problem. This has been previously broached with many different supervised machine learning methods, including deep-learning neural networks. However, these neural networks have thus far not been tailored to give any indication of confidence in the mutation call, meaning an oncologist could be targeting a mutation with a low probability of being true. To address this, we present here a deep bayesian recurrent neural network for cancer variant calling, which shows no degradation in performance compared to standard neural networks. This approach enables greater flexibility through different priors to avoid overfitting to a single dataset. We will be incorporating this approach into software for oncologists to obtain safe, robust, and statistically confident somatic mutation calls for precision oncology treatment choices.

We consider the problem of compressed sensing and of (real-valued) phase retrieval with random measurement matrix. We derive sharp asymptotics for the information-theoretically optimal performance and for the best known polynomial algorithm for an ensemble of generative priors consisting of fully connected deep neural networks with random weight matrices and arbitrary activations. We compare the performance to sparse separable priors and conclude that generative priors might be advantageous in terms of algorithmic performance. In particular, while sparsity does not allow to perform compressive phase retrieval efficiently close to its information-theoretic limit, it is found that under the random generative prior compressed phase retrieval becomes tractable.

Deep metric learning employs deep neural networks to embed instances into a metric space such that distances between instances of the same class are small and distances between instances from different classes are large. In most existing deep metric learning techniques, the embedding of an instance is given by a feature vector produced by a deep neural network and Euclidean distance or cosine similarity defines distances between these vectors. In this paper, we study deep distributional embeddings of sequences, where the embedding of a sequence is given by the distribution of learned deep features across the sequence. This has the advantage of capturing statistical information about the distribution of patterns within the sequence in the embedding. When embeddings are distributions rather than vectors, measuring distances between embeddings involves comparing their respective distributions. We propose a distance metric based on Wasserstein distances between the distributions and a corresponding loss function for metric learning, which leads to a novel end-to-end trainable embedding model. We empirically observe that distributional embeddings outperform standard vector embeddings and that training with the proposed Wasserstein metric outperforms training with other distance functions.

Handwriting-based gender classification is a well-researched problem that has been approached mainly by traditional machine learning techniques. In this paper, we propose a novel deep learning-based approach for this task. Specifically, we present a convolutional neural network (CNN), which performs automatic feature extraction from a given handwritten image, followed by classification of the writer's gender. Also, we introduce a new dataset of labeled handwritten samples, in Hebrew and English, of 405 participants. Comparing the gender classification accuracy on this dataset against human examiners, our results show that the proposed deep learning-based approach is substantially more accurate than that of humans.

Adversarial Training (AT) and Virtual Adversarial Training (VAT) are the regularization techniques that train Deep Neural Networks (DNNs) with adversarial examples generated by adding small but worst-case perturbations to input examples. In this paper, we propose xAT and xVAT, new adversarial training algorithms, that generate multiplicative perturbations to input examples for robust training of DNNs. Such perturbations are much more perceptible and interpretable than their additive counterparts exploited by AT and VAT. Furthermore, the multiplicative perturbations can be generated transductively or inductively while the standard AT and VAT only support a transductive implementation. W e conduct a series of experiments that analyze the behavior of the multiplicative perturbations and demonstrate that xAT and xVAT match or outperform state-of-the-art classification accuracies across multiple established benchmarks while being about 30% faster than their additive counterparts. Furthermore, the resulting DNNs also demonstrate distinct weight distributions. 1. Introduction Over the past few years, Deep Neural Networks (DNNs) have achieved state-of-the-art performance on a wide range of learning tasks. However, the success of DNNs has a high reliance on large sets of labeled examples; when trained on small datasets, DNNs plague to overfitting if not regularized properly. For many practical applications, collecting a large amount of labeled examples is very expensive and/or time-consuming. To address this issue, researchers have investigated a host of techniques, such as Dropout [24], A T [4, 25], V A T [14], and Mixup [29], to regularize the training of DNNs.

Event ticket price prediction is important to marketing strategy for any sports team or musical ensemble. An accurate prediction model can help the marketing team to make promotion plan more effectively and efficiently. However, given all the historical transaction records, it is challenging to predict the sale price of the remaining seats at any future timestamp, not only because that the sale price is relevant to a lot of features (seat locations, date-to-event of the transaction, event date, team performance, etc.), but also because of the temporal and spatial sparsity in the dataset. For a game/concert, the ticket selling price of one seat is only observable once at the time of sale. Furthermore, some seats may not even be purchased (therefore no record available). In fact, data sparsity is commonly encountered in many prediction problems. Here, we propose a bi-level optimizing deep neural network to address the curse of spatio-temporal sparsity. Specifically, we introduce coarsening and refining layers, and design a bi-level loss function to integrate different level of loss for better prediction accuracy. Our model can discover the interrelations among ticket sale price, seat locations, selling time, event information, etc. Experiments show that our proposed model outperforms other benchmark methods in real-world ticket selling price prediction.

We present DeepFPC, a novel deep neural network designed by unfolding the iterations of the fixed-point continuation algorithm with one-sided l1-norm (FPC-l1), which has been proposed for solving the 1-bit compressed sensing problem. The network architecture resembles that of deep residual learning and incorporates prior knowledge about the signal structure (i.e., sparsity), thereby offering interpretability by design. Once DeepFPC is properly trained, a sparse signal can be recovered fast and accurately from quantized measurements. The proposed model is evaluated in the task of direction-of-arrival (DOA) estimation and is shown to outperform state-of-the-art algorithms, namely, the iterative FPC-l1 algorithm and the 1-bit MUSIC method.