Everyday we are exposed to various chemicals via food additives, cleaning and cosmetic products and medicines -- and some of them might be toxic. However testing the toxicity of all existing compounds by biological experiments is neither financially nor logistically feasible. Therefore the government agencies NIH, EPA and FDA launched the Tox21 Data Challenge within the "Toxicology in the 21st Century" (Tox21) initiative. The goal of this challenge was to assess the performance of computational methods in predicting the toxicity of chemical compounds. State of the art toxicity prediction methods build upon specifically-designed chemical descriptors developed over decades. Though Deep Learning is new to the field and was never applied to toxicity prediction before, it clearly outperformed all other participating methods. In this application paper we show that deep nets automatically learn features resembling well-established toxicophores. In total, our Deep Learning approach won both of the panel-challenges (nuclear receptors and stress response) as well as the overall Grand Challenge, and thereby sets a new standard in tox prediction.

Catastrophic forgetting is a problem faced by many machine learning models and algorithms. When trained on one task, then trained on a second task, many machine learning models "forget" how to perform the first task. This is widely believed to be a serious problem for neural networks. Here, we investigate the extent to which the catastrophic forgetting problem occurs for modern neural networks, comparing both established and recent gradient-based training algorithms and activation functions. We also examine the effect of the relationship between the first task and the second task on catastrophic forgetting. We find that it is always best to train using the dropout algorithm--the dropout algorithm is consistently best at adapting to the new task, remembering the old task, and has the best tradeoff curve between these two extremes. We find that different tasks and relationships between tasks result in very different rankings of activation function performance. This suggests the choice of activation function should always be cross-validated.

Inspired by recent successes of deep learning in computer vision and speech recognition, we propose a novel framework to encode time series data as different types of images, namely, Gramian Angular Fields (GAF) and Markov Transition Fields (MTF). This enables the use of techniques from computer vision for classification. Using a polar coordinate system, GAF images are represented as a Gramian matrix where each element is the trigonometric sum (i.e., superposition of directions) between different time intervals. MTF images represent the first order Markov transition probability along one dimension and temporal dependency along the other. We used Tiled Convolutional Neural Networks (tiled CNNs) on 12 standard datasets to learn high-level features from individual GAF, MTF, and GAF-MTF images that resulted from combining GAF and MTF representations into a single image. The classification results of our approach are competitive with five stateof-the-art approaches. An analysis of the features and weights learned via tiled CNNs explains why the approach works.

Recently it has been shown that deep neural networks can learn to play Atari games by directly observing raw pixels of the playing area. We show how apprenticeship learning can be applied in this setting so that an agent can learn to perform a task (i.e. play a game) by observing the expert, without any explicitly provided knowledge of the game’s internal state or objectives.

Why does Deep Learning work? What representations does it capture? How do higher-order representations emerge? We study these questions from the perspective of group theory, thereby opening a new approach towards a theory of Deep learning. One factor behind the recent resurgence of the subject is a key algorithmic step called pre-training: first search for a good generative model for the input samples, and repeat the process one layer at a time. We show deeper implications of this simple principle, by establishing a connection with the interplay of orbits and stabilizers of group actions. Although the neural networks themselves may not form groups, we show the existence of {\em shadow} groups whose elements serve as close approximations. Over the shadow groups, the pre-training step, originally introduced as a mechanism to better initialize a network, becomes equivalent to a search for features with minimal orbits. Intuitively, these features are in a way the {\em simplest}. Which explains why a deep learning network learns simple features first. Next, we show how the same principle, when repeated in the deeper layers, can capture higher order representations, and why representation complexity increases as the layers get deeper.

Training very deep networks is an important open problem in machine learning. One of many difficulties is that the norm of the back-propagated error gradient can grow or decay exponentially. Here we show that training very deep feed-forward networks (FFNs) is not as difficult as previously thought. Unlike when back-propagation is applied to a recurrent network, application to an FFN amounts to multiplying the error gradient by a different random matrix at each layer. We show that the successive application of correctly scaled random matrices to an initial vector results in a random walk of the log of the norm of the resulting vectors, and we compute the scaling that makes this walk unbiased. The variance of the random walk grows only linearly with network depth and is inversely proportional to the size of each layer. Practically, this implies a gradient whose log-norm scales with the square root of the network depth and shows that the vanishing gradient problem can be mitigated by increasing the width of the layers. Mathematical analyses and experimental results using stochastic gradient descent to optimize tasks related to the MNIST and TIMIT datasets are provided to support these claims. Equations for the optimal matrix scaling are provided for the linear and ReLU cases.

Top-performing deep architectures are trained on massive amounts of labeled data. In the absence of labeled data for a certain task, domain adaptation often provides an attractive option given that labeled data of similar nature but from a different domain (e.g. synthetic images) are available. Here, we propose a new approach to domain adaptation in deep architectures that can be trained on large amount of labeled data from the source domain and large amount of unlabeled data from the target domain (no labeled target-domain data is necessary). As the training progresses, the approach promotes the emergence of "deep" features that are (i) discriminative for the main learning task on the source domain and (ii) invariant with respect to the shift between the domains. We show that this adaptation behaviour can be achieved in almost any feed-forward model by augmenting it with few standard layers and a simple new gradient reversal layer. The resulting augmented architecture can be trained using standard backpropagation. Overall, the approach can be implemented with little effort using any of the deep-learning packages. The method performs very well in a series of image classification experiments, achieving adaptation effect in the presence of big domain shifts and outperforming previous state-of-the-art on Office datasets.

A generative Bayesian model is developed for deep (multi-layer) convolutional dictionary learning. A novel probabilistic pooling operation is integrated into the deep model, yielding efficient bottom-up and top-down probabilistic learning. After learning the deep convolutional dictionary, testing is implemented via deconvolutional inference. To speed up this inference, a new statistical approach is proposed to project the top-layer dictionary elements to the data level. Following this, only one layer of deconvolution is required during testing. Experimental results demonstrate powerful capabilities of the model to learn multi-layer features from images. Excellent classification results are obtained on both the MNIST and Caltech 101 datasets.

Recently, fully-connected and convolutional neural networks have been trained to achieve state-of-the-art performance on a wide variety of tasks such as speech recognition, image classification, natural language processing, and bioinformatics. For classification tasks, most of these "deep learning" models employ the softmax activation function for prediction and minimize cross-entropy loss. In this paper, we demonstrate a small but consistent advantage of replacing the softmax layer with a linear support vector machine. Learning minimizes a margin-based loss instead of the cross-entropy loss. While there have been various combinations of neural nets and SVMs in prior art, our results using L2-SVMs show that by simply replacing softmax with linear SVMs gives significant gains on popular deep learning datasets MNIST, CIFAR-10, and the ICML 2013 Representation Learning Workshop's face expression recognition challenge.

Recurrent Neural Networks (RNNs) with Long Short-Term Memory units (LSTM) are widely used because they are expressive and are easy to train. Our interest lies in empirically evaluating the expressiveness and the learnability of LSTMs in the sequence-to-sequence regime by training them to evaluate short computer programs, a domain that has traditionally been seen as too complex for neural networks. We consider a simple class of programs that can be evaluated with a single left-to-right pass using constant memory. Our main result is that LSTMs can learn to map the character-level representations of such programs to their correct outputs. Notably, it was necessary to use curriculum learning, and while conventional curriculum learning proved ineffective, we developed a new variant of curriculum learning that improved our networks' performance in all experimental conditions. The improved curriculum had a dramatic impact on an addition problem, making it possible to train an LSTM to add two 9-digit numbers with 99% accuracy.