We propose several ways of reusing subword embeddings and other weights in subword-aware neural language models. The proposed techniques do not benefit a competitive character-aware model, but some of them improve the performance of syllable- and morpheme-aware models while showing significant reductions in model sizes. We discover a simple hands-on principle: in a multi-layer input embedding model, layers should be tied consecutively bottom-up if reused at output. Our best morpheme-aware model with properly reused weights beats the competitive word-level model by a large margin across multiple languages and has 20%-87% fewer parameters.

Asynchronous distributed machine learning solutions have proven very effective so far, but always assuming perfectly functioning workers. In practice, some of the workers can however exhibit Byzantine behavior, caused by hardware failures, software bugs, corrupt data, or even malicious attacks. We introduce \emph{Kardam}, the first distributed asynchronous stochastic gradient descent (SGD) algorithm that copes with Byzantine workers. Kardam consists of two complementary components: a filtering and a dampening component. The first is scalar-based and ensures resilience against $\frac{1}{3}$ Byzantine workers. Essentially, this filter leverages the Lipschitzness of cost functions and acts as a self-stabilizer against Byzantine workers that would attempt to corrupt the progress of SGD. The dampening component bounds the convergence rate by adjusting to stale information through a generic gradient weighting scheme. We prove that Kardam guarantees almost sure convergence in the presence of asynchrony and Byzantine behavior, and we derive its convergence rate. We evaluate Kardam on the CIFAR-100 and EMNIST datasets and measure its overhead with respect to non Byzantine-resilient solutions. We empirically show that Kardam does not introduce additional noise to the learning procedure but does induce a slowdown (the cost of Byzantine resilience) that we both theoretically and empirically show to be less than $f/n$, where $f$ is the number of Byzantine failures tolerated and $n$ the total number of workers. Interestingly, we also empirically observe that the dampening component is interesting in its own right for it enables to build an SGD algorithm that outperforms alternative staleness-aware asynchronous competitors in environments with honest workers.

We present Deep Voice 3, a fully-convolutional attention-based neural text-to-speech (TTS) system. Deep Voice 3 matches state-of-the-art neural speech synthesis systems in naturalness while training ten times faster. We scale Deep Voice 3 to data set sizes unprecedented for TTS, training on more than eight hundred hours of audio from over two thousand speakers. In addition, we identify common error modes of attention-based speech synthesis networks, demonstrate how to mitigate them, and compare several different waveform synthesis methods. We also describe how to scale inference to ten million queries per day on one single-GPU server.

In recent years, Deep Reinforcement Learning has made impressive advances in solving several important benchmark problems for sequential decision making. Many control applications use a generic multilayer perceptron (MLP) for non-vision parts of the policy network. In this work, we propose a new neural network architecture for the policy network representation that is simple yet effective. The proposed Structured Control Net (SCN) splits the generic MLP into two separate sub-modules: a nonlinear control module and a linear control module. Intuitively, the nonlinear control is for forward-looking and global control, while the linear control stabilizes the local dynamics around the residual of global control. We hypothesize that this will bring together the benefits of both linear and nonlinear policies: improve training sample efficiency, final episodic reward, and generalization of learned policy, while requiring a smaller network and being generally applicable to different training methods. We validated our hypothesis with competitive results on simulations from OpenAI MuJoCo, Roboschool, Atari, and a custom 2D urban driving environment, with various ablation and generalization tests, trained with multiple black-box and policy gradient training methods. The proposed architecture has the potential to improve upon broader control tasks by incorporating problem specific priors into the architecture. As a case study, we demonstrate much improved performance for locomotion tasks by emulating the biological central pattern generators (CPGs) as the nonlinear part of the architecture.

We explore efficient neural architecture search methods and show that a simple yet powerful evolutionary algorithm can discover new architectures with excellent performance. Our approach combines a novel hierarchical genetic representation scheme that imitates the modularized design pattern commonly adopted by human experts, and an expressive search space that supports complex topologies. Our algorithm efficiently discovers architectures that outperform a large number of manually designed models for image classification, obtaining top-1 error of 3.6% on CIFAR-10 and 20.3% when transferred to ImageNet, which is competitive with the best existing neural architecture search approaches.

We present a deep learning model, DE-LSTM, for the simulation of a stochastic process with underlying nonlinear dynamics. The deep learning model aims to approximate the probability density function of a stochastic process via numerical discretization and the underlying nonlinear dynamics is modeled by the Long Short-Term Memory (LSTM) network. After the numerical discretization by a softmax function, the function estimation problem is solved by a multi-label classification problem. A penalized maximum log likelihood method is proposed to impose smoothness in the predicted probability distribution. It is shown that LSTM is a state space model, where the internal dynamics consists of a system of relaxation processes. A sequential Monte Carlo method is outlined to compute the time evolution of the probability distribution. The behavior of DE-LSTM is investigated by using the Ornstein-Uhlenbeck process and noisy observations of Mackey-Glass equation and forced Van der Pol oscillators. While the probability distribution computed by the conventional maximum log likelihood method makes a good prediction of the first and second moments, the Kullback-Leibler divergence shows that the penalized maximum log likelihood method results in a probability distribution closer to the ground truth. It is shown that DE-LSTM makes a good prediction of the probability distribution without assuming any distributional properties of the noise. For a multiple-step forecast, it is found that the prediction uncertainty, denoted by the 95% confidence interval, does not grow monotonically in time. For a chaotic system, Mackey-Glass time series, the 95% confidence interval first grows, then exhibits an oscillatory behavior, instead of growing indefinitely, while for the forced Van der Pol oscillator, the prediction uncertainty does not grow in time even for 3,000-step forecast.

Vanishing long-term gradients are a major issue in training standard recurrent neural networks (RNNs), which can be alleviated by long short-term memory (LSTM) models with memory cells. However, the extra parameters associated with the memory cells mean an LSTM layer has four times as many parameters as an RNN with the same hidden vector size. This paper addresses the vanishing gradient problem using a high order RNN (HORNN) which has additional connections from multiple previous time steps. Speech recognition experiments using British English multi-genre broadcast (MGB3) data showed that the proposed HORNN architectures for rectified linear unit and sigmoid activation functions reduced word error rates (WER) by 4.2% and 6.3% over the corresponding RNNs, and gave similar WERs to a (projected) LSTM while using only 20%--50% of the recurrent layer parameters and computation.

Deep learning is a sub-field of machine learning which uses several learning algorithms to solve real-world tasks as image recognition, facial detection, signal processing, on supervised, unsupervised and reinforcement learning of feature representation at successively higher, more abstract layers. Those algorithms are artificial models such as Convolution Neural Networks (CNN), Deep Belief Networks (DBNs), Recurrent Neural Networks (RNNs) and Auto-encoders (AE). Even with the growth and success on many application of deep learning, some issues still remain unsolved in general. One of these issues is the catastrophic forgetting problem [1]. This issue can be seen as an handicap to develop truly intelligent systems. Catastrophic forgetting arises when a neural network is not capable of preserving the past learned task when learning new task. There are some approaches that benefit from previously learned information to improve performance of learning new information, for example fine-tuning [2] where the parameters of the old task are adjusted for adapting to a new task. Other approach well known is feature extraction [3] where the parameters of the old network are unchanged and the parameters of the outputs of one or more layers are used to extract feature for the new task.

We consider the problem of learning fair decision systems in complex scenarios in which a sensitive attribute might affect the decision along both fair and unfair pathways. We introduce a causal approach to disregard effects along unfair pathways that simplifies and generalizes previous literature. Our method corrects observations adversely affected by the sensitive attribute, and uses these to form a decision. This avoids disregarding fair information, and does not require an often intractable computation of the path-specific effect. We leverage recent developments in deep learning and approximate inference to achieve a solution that is widely applicable to complex, non-linear scenarios.

One of the main drivers behind the rapid recent advances in machine learning has been the availability of efficient system support. This comes both through hardware acceleration, but also in the form of efficient software frameworks and programming models. Despite significant progress, scaling compute-intensive machine learning workloads to a large number of compute nodes is still a challenging task, with significant latency and bandwidth demands. In this paper, we address this challenge, by proposing SPARCML, a general, scalable communication layer for machine learning applications. SPARCML is built on the observation that many distributed machine learning algorithms either have naturally sparse communication patters, or have updates which can be sparsified in a structured way for improved performance, without any convergence or accuracy loss. To exploit this insight, we design and implement a set of communication efficient protocols for sparse input data, in conjunction with efficient machine learning algorithms which can leverage these primitives. Our communication protocols generalize standard collective operations, by allowing processes to contribute sparse input data vectors, of heterogeneous sizes. We call these operations sparse-input collectives, and present efficient practical algorithms with strong theoretical bounds on their running time and communication cost. Our generic communication layer is enriched with additional features, such support for non-blocking (asynchronous) operations, and support for low-precision data representations. We validate our algorithmic results experimentally on a range of large-scale machine learning applications and target architectures, showing that we can leverage sparsity for order- of-magnitude runtime savings, compared to state-of-the art methods and frameworks.