Architecture search aims at automatically finding neural architectures that are competitive with architectures designed by human experts. While recent approaches have come close to matching the predictive performance of manually designed architectures for image recognition, these approaches are problematic under constrained resources for two reasons: first, the architecture search itself requires vast computational resources for most proposed methods. Secondly, the found neural architectures are solely optimized for high predictive performance without penalizing excessive resource consumption. We address the first shortcoming by proposing NASH, an architecture search which considerable reduces the computational resources required for training novel architectures by applying network morphisms and aggressive learning rate schedules. On CIFAR10, NASH finds architectures with errors below 4% in only 3 days. We address the second shortcoming by proposing Pareto-NASH, a method for multi-objective architecture search that allows approximating the Pareto-front of architectures under multiple objective, such as predictive performance and number of parameters, in a single run of the method. Within 56 GPU days of architecture search, Pareto-NASH finds a model with 4M parameters and test error of 3.5%, as well as a model with less than 1M parameters and test error of 4.6%.

The MIMIC-CXR dataset is (to date) the largest released chest x-ray dataset consisting of 473,064 chest x-rays and 206,574 radiology reports collected from 63,478 patients. We present the results of training and evaluating a collection of deep convolutional neural networks on this dataset to recognize multiple common thorax diseases. To the best of our knowledge, this is the first work that trains CNNs for this task on such a large collection of chest x-ray images, which is over four times the size of the largest previously released chest x-ray corpus (ChestX-Ray14). We describe and evaluate individual CNN models trained on frontal and lateral CXR view types. In addition, we present a novel DualNet architecture that emulates routine clinical practice by simultaneously processing both frontal and lateral CXR images obtained from a radiological exam. Our DualNet architecture shows improved performance in recognizing findings in CXR images when compared to applying separate baseline frontal and lateral classifiers.

Monitoring the biomedical literature for cases of Adverse Drug Reactions (ADRs) is a critically important and time consuming task in pharmacovigilance. The development of computer assisted approaches to aid this process in different forms has been the subject of many recent works. One particular area that has shown promise is the use of Deep Neural Networks, in particular, Convolutional Neural Networks (CNNs), for the detection of ADR relevant sentences. Using token-level convolutions and general purpose word embeddings, this architecture has shown good performance relative to more traditional models as well as Long Short Term Memory (LSTM) models. In this work, we evaluate and compare two different CNN architectures using the ADE corpus. In addition, we show that by de-duplicating the ADR relevant sentences, we can greatly reduce overoptimism in the classification results. Finally, we evaluate the use of word embeddings specifically developed for biomedical text and show that they lead to a better performance in this task.

Existing applications include a huge amount of knowledge that is out of reach for deep neural networks. This paper presents a novel approach for integrating calls to existing applications into deep learning architectures. Using this approach, we estimate each application's functionality with an estimator, which is implemented as a deep neural network (DNN). The estimator is then embedded into a base network that we direct into complying with the application's interface during an end-to-end optimization process. At inference time, we replace each estimator with its existing application counterpart and let the base network solve the task by interacting with the existing application. Using this 'Estimate and Replace' method, we were able to train a DNN end-to-end with less data and outperformed a matching DNN that did not interact with the external application.

Many recently trained neural networks employ large numbers of parameters to achieve good performance. One may intuitively use the number of parameters required as a rough gauge of the difficulty of a problem. But how accurate are such notions? How many parameters are really needed? In this paper we attempt to answer this question by training networks not in their native parameter space, but instead in a smaller, randomly oriented subspace. We slowly increase the dimension of this subspace, note at which dimension solutions first appear, and define this to be the intrinsic dimension of the objective landscape. The approach is simple to implement, computationally tractable, and produces several suggestive conclusions. Many problems have smaller intrinsic dimensions than one might suspect, and the intrinsic dimension for a given dataset varies little across a family of models with vastly different sizes. This latter result has the profound implication that once a parameter space is large enough to solve a problem, extra parameters serve directly to increase the dimensionality of the solution manifold. Intrinsic dimension allows some quantitative comparison of problem difficulty across supervised, reinforcement, and other types of learning where we conclude, for example, that solving the inverted pendulum problem is 100 times easier than classifying digits from MNIST, and playing Atari Pong from pixels is about as hard as classifying CIFAR-10. In addition to providing new cartography of the objective landscapes wandered by parameterized models, the method is a simple technique for constructively obtaining an upper bound on the minimum description length of a solution. A byproduct of this construction is a simple approach for compressing networks, in some cases by more than 100 times.

Semi-supervised learning (SSL) provides a powerful framework for leveraging unlabeled data when labels are limited or expensive to obtain. SSL algorithms based on deep neural networks have recently proven successful on standard benchmark tasks. However, we argue that these benchmarks fail to address many issues that these algorithms would face in real-world applications. After creating a unified reimplementation of various widely-used SSL techniques, we test them in a suite of experiments designed to address these issues. We find that the performance of simple baselines which do not use unlabeled data is often underreported, that SSL methods differ in sensitivity to the amount of labeled and unlabeled data, and that performance can degrade substantially when the unlabeled dataset contains out-of-class examples. To help guide SSL research towards real-world applicability, we make our unified reimplemention and evaluation platform publicly available.

The state of the art in video understanding suffers from two problems: (1) The major part of reasoning is performed locally in the video, therefore, it misses important relationships within actions that span several seconds. (2) While there are local methods with fast per-frame processing, the processing of the whole video is not efficient and hampers fast video retrieval or online classification of long-term activities. In this paper, we introduce a network architecture that takes long-term content into account and enables fast per-video processing at the same time. The architecture is based on merging long-term content already in the network rather than in a post-hoc fusion. Together with a sampling strategy, which exploits that neighboring frames are largely redundant, this yields high-quality action classification and video captioning at up to 230 videos per second, where each video can consist of a few hundred frames. The approach achieves competitive performance across all datasets while being 10x to 80x faster than state-of-the-art methods.

We present the MAC network, a novel fully differentiable neural network architecture, designed to facilitate explicit and expressive reasoning. MAC moves away from monolithic black-box neural architectures towards a design that encourages both transparency and versatility. The model approaches problems by decomposing them into a series of attention-based reasoning steps, each performed by a novel recurrent Memory, Attention, and Composition (MAC) cell that maintains a separation between control and memory. By stringing the cells together and imposing structural constraints that regulate their interaction, MAC effectively learns to perform iterative reasoning processes that are directly inferred from the data in an end-to-end approach. We demonstrate the model's strength, robustness and interpretability on the challenging CLEVR dataset for visual reasoning, achieving a new state-of-the-art 98.9% accuracy, halving the error rate of the previous best model. More importantly, we show that the model is computationally-efficient and data-efficient, in particular requiring 5x less data than existing models to achieve strong results.

Measuring the distance between concepts is an important field of study of Natural Language Processing, as it can be used to improve tasks related to the interpretation of those same concepts. WordNet, which includes a wide variety of concepts associated with words (i.e., synsets), is often used as a source for computing those distances. In this paper, we explore a distance for WordNet synsets based on visual features, instead of lexical ones. For this purpose, we extract the graphic features generated within a deep convolutional neural networks trained with ImageNet and use those features to generate a representative of each synset. Based on those representatives, we define a distance measure of synsets, which complements the traditional lexical distances. Finally, we propose some experiments to evaluate its performance and compare it with the current state-of-the-art.

Neural Sequence-to-Sequence models have proven to be accurate and robust for many sequence prediction tasks, and have become the standard approach for automatic translation of text. The models work in a five stage blackbox process that involves encoding a source sequence to a vector space and then decoding out to a new target sequence. This process is now standard, but like many deep learning methods remains quite difficult to understand or debug. In this work, we present a visual analysis tool that allows interaction with a trained sequence-to-sequence model through each stage of the translation process. The aim is to identify which patterns have been learned and to detect model errors. We demonstrate the utility of our tool through several real-world large-scale sequence-to-sequence use cases.