Deep neural networks (DNNs) have begun to have a pervasive impact on various applications of machine learning. However, the problem of finding an optimal DNN architecture for large applications is challenging. Common approaches go for deeper and larger DNN architectures but may incur substantial redundancy. To address these problems, we introduce a network growth algorithm that complements network pruning to learn both weights and compact DNN architectures during training. We propose a DNN synthesis tool (NeST) that combines both methods to automate the generation of compact and accurate DNNs. NeST starts with a randomly initialized sparse network called the seed architecture. It iteratively tunes the architecture with gradient-based growth and magnitude-based pruning of neurons and connections. Our experimental results show that NeST yields accurate, yet very compact DNNs, with a wide range of seed architecture selection. For the LeNet-300-100 (LeNet-5) architecture, we reduce network parameters by 70.2x (74.3x) and floating-point operations (FLOPs) by 79.4x (43.7x). For the AlexNet and VGG-16 architectures, we reduce network parameters (FLOPs) by 15.7x (4.6x) and 30.2x (8.6x), respectively. NeST's grow-and-prune paradigm delivers significant additional parameter and FLOPs reduction relative to pruning-only methods.

A variety of methods existing for generating synthetic electronic health records (EHRs), but they are not capable of generating unstructured text, like emergency department (ED) chief complaints, history of present illness or progress notes. Here, we use the encoder-decoder model, a deep learning algorithm that features in many contemporary machine translation systems, to generate synthetic chief complaints from discrete variables in EHRs, like age group, gender, and discharge diagnosis. After being trained end-to-end on authentic records, the model can generate realistic chief complaint text that preserves much of the epidemiological information in the original data. As a side effect of the model's optimization goal, these synthetic chief complaints are also free of relatively uncommon abbreviation and misspellings, and they include none of the personally-identifiable information (PII) that was in the training data, suggesting it may be used to support the de-identification of text in EHRs. When combined with algorithms like generative adversarial networks (GANs), our model could be used to generate fully-synthetic EHRs, facilitating data sharing between healthcare providers and researchers and improving our ability to develop machine learning methods tailored to the information in healthcare data. 1 Introduction The wide adoption of electronic health record (EHR) systems has led to the creation of large amounts of healthcare data. Although these data are primarily used to improve patient outcomes and streamline the delivery of care (healthit.gov), Because they contain personally identifiable patient information, however, much of which is protected under the Health Insurance Portability and Accountability Act (HIPAA), these data are often difficult for providers to share with investigators outside their organizations, limiting their feasibility for use in research.

Extraction of missing attribute values is to find values describing an attribute of interest from a free text input. Most past related work on extraction of missing attribute values work with a closed world assumption with the possible set of values known beforehand, or use dictionaries of values and hand-crafted features. How can we discover new attribute values that we have never seen before? Can we do this with limited human annotation or supervision? We study this problem in the context of product catalogs that often have missing values for many attributes of interest. In this work, we leverage product profile information such as titles and descriptions to discover missing values of product attributes. We develop a novel deep tagging model OpenTag for this extraction problem with the following contributions: (1) we formalize the problem as a sequence tagging task, and propose a joint model exploiting recurrent neural networks (specifically, bidirectional LSTM) to capture context and semantics, and Conditional Random Fields (CRF) to enforce tagging consistency, (2) we develop a novel attention mechanism to provide interpretable explanation for our model's decisions, (3) we propose a novel sampling strategy exploring active learning to reduce the burden of human annotation. OpenTag does not use any dictionary or hand-crafted features as in prior works. Extensive experiments in real-life datasets in different domains show that OpenTag with our active learning strategy discovers new attribute values from as few as 150 annotated samples (reduction in 3.3x amount of annotation effort) with a high F-score of 83%, outperforming state-of-the-art models.

Episodic memory is a psychology term which refers to the ability to recall specific events from the past. We suggest one advantage of this particular type of memory is the ability to easily assign credit to a specific state when remembered information is found to be useful. Inspired by this idea, and the increasing popularity of external memory mechanisms to handle long-term dependencies in deep learning systems, we propose a novel algorithm which uses a reservoir sampling procedure to maintain an external memory consisting of a fixed number of past states. The algorithm allows a deep reinforcement learning agent to learn online to preferentially remember those states which are found to be useful to recall later on. Critically this method allows for efficient online computation of gradient estimates with respect to the write process of the external memory. Thus unlike most prior mechanisms for external memory it is feasible to use in an online reinforcement learning setting. Much of reinforcement learning (RL) theory is based on the assumption that the environment has the Markov property, meaning that future states are independent of past states given the present state. This implies the agent has all the information it needs to make an optimal decision at each time and therefore has no need to remember the past. This is however not realistic in general, realistic problems often require significant information from the past to make an informed decision in the present, and there is often no obvious way to incorporate the relevant information into an expanded present state.

In this work, we present theoretical results on the convergence of non-convex accelerated gradient descent in matrix factorization models. The technique is applied to matrix sensing problems with squared loss, for the estimation of a rank $r$ optimal solution $X^\star \in \mathbb{R}^{n \times n}$. We show that the acceleration leads to linear convergence rate, even under non-convex settings where the variable $X$ is represented as $U U^\top$ for $U \in \mathbb{R}^{n \times r}$. Our result has the same dependence on the condition number of the objective --and the optimal solution-- as that of the recent results on non-accelerated algorithms. However, acceleration is observed in practice, both in synthetic examples and in two real applications: neuronal multi-unit activities recovery from single electrode recordings, and quantum state tomography on quantum computing simulators.

Exploiting sparsity enables hardware systems to run neural networks faster and more energy-efficiently. However, most prior sparsity-centric optimization techniques only accelerate the forward pass of neural networks and usually require an even longer training process with iterative pruning and retraining. We observe that artificially inducing sparsity in the gradients of the gates in an LSTM cell has little impact on the training quality. Further, we can enforce structured sparsity in the gate gradients to make the LSTM backward pass up to 45% faster than the state-of-the-art dense approach and 168% faster than the state-of-the-art sparsifying method on modern GPUs. Though the structured sparsifying method can impact the accuracy of a model, this performance gap can be eliminated by mixing our sparse training method and the standard dense training method. Experimental results show that the mixed method can achieve comparable results in a shorter time span than using purely dense training.

A semi-recurrent hybrid VAE-GAN model for generating sequential data is introduced. In order to consider the spatial correlation of the data in each frame of the generated sequence, CNNs are utilized in the encoder, generator, and discriminator. The subsequent frames are sampled from the latent distributions obtained by encoding the previous frames. As a result, the dependencies between the frames are maintained. Two testing frameworks for synthesizing a sequence with any number of frames are also proposed. The promising experimental results on piano music generation indicates the potential of the proposed framework in modeling other sequential data such as video.

Machine learning is currently dominated by largely experimental work focused on improvements in a few key tasks. However, the impressive accuracy numbers of the best performing models are questionable because the same test sets have been used to select these models for multiple years now. To understand the danger of overfitting, we measure the accuracy of CIFAR-10 classifiers by creating a new test set of truly unseen images. Although we ensure that the new test set is as close to the original data distribution as possible, we find a large drop in accuracy (4% to 10%) for a broad range of deep learning models. Yet more recent models with higher original accuracy show a smaller drop and better overall performance, indicating that this drop is likely not due to overfitting based on adaptivity. Instead, we view our results as evidence that current accuracy numbers are brittle and susceptible to even minute natural variations in the data distribution.

In this paper, we present a technique for unsupervised learning of visual representations. Specifically, we train a model for foreground and background classification task, in the process of which it learns visual representations. Foreground and background patches for training come af- ter mining for such patches from hundreds and thousands of unlabelled videos available on the web which we ex- tract using a proposed patch extraction algorithm. With- out using any supervision, with just using 150, 000 unla- belled videos and the PASCAL VOC 2007 dataset, we train a object recognition model that achieves 45.3 mAP which is close to the best performing unsupervised feature learn- ing technique whereas better than many other proposed al- gorithms. The code for patch extraction is implemented in Matlab and available open source at the following link .

Understanding the nature of representations learned by supervised machine learning models is a significant goal in the machine learning community. We present a method for feature interpretation that makes use of recent advances in autoregressive density estimation models to invert model representations. We train generative inversion models to express a distribution over input features conditioned on intermediate model representations. Insights into the invariances learned by supervised models can be gained by viewing samples from these inversion models. In addition, we can use these inversion models to estimate the mutual information between a model's inputs and its intermediate representations, thus quantifying the amount of information preserved by the network at different stages. Using this method we examine the types of information preserved at different layers of convolutional neural networks, and explore the invariances induced by different architectural choices. Finally we show that the mutual information between inputs and network layers decreases over the course of training, supporting recent work by Shwartz-Ziv and Tishby (2017) on the information bottleneck theory of deep learning.