This paper highlights new opportunities for designing large-scale machine learning systems as a consequence of blurring traditional boundaries that have allowed algorithm designers and application-level practitioners to stay -- for the most part -- oblivious to the details of the underlying hardware-level implementations. The hardware/software co-design methodology advocated here hinges on the deployment of compute-intensive machine learning kernels onto compute platforms that trade-off determinism in the computation for improvement in speed and/or energy efficiency. To achieve this, we revisit digital stochastic circuits for approximating matrix computations that are ubiquitous in machine learning algorithms. Theoretical and empirical evaluation is undertaken to assess the impact of the hardware-induced computational noise on algorithm performance. As a proof-of-concept, a stochastic hardware simulator is employed for training deep neural networks for image recognition problems.

In this paper, we propose a novel neural network model called RNN Encoder-Decoder that consists of two recurrent neural networks (RNN). One RNN encodes a sequence of symbols into a fixed-length vector representation, and the other decodes the representation into another sequence of symbols. The encoder and decoder of the proposed model are jointly trained to maximize the conditional probability of a target sequence given a source sequence. The performance of a statistical machine translation system is empirically found to improve by using the conditional probabilities of phrase pairs computed by the RNN Encoder-Decoder as an additional feature in the existing log-linear model. Qualitatively, we show that the proposed model learns a semantically and syntactically meaningful representation of linguistic phrases.

Neural Autoregressive Distribution Estimators (NADEs) have recently been shown as successful alternatives for modeling high dimensional multimodal distributions. One issue associated with NADEs is that they rely on a particular order of factorization for $P(\mathbf{x})$. This issue has been recently addressed by a variant of NADE called Orderless NADEs and its deeper version, Deep Orderless NADE. Orderless NADEs are trained based on a criterion that stochastically maximizes $P(\mathbf{x})$ with all possible orders of factorizations. Unfortunately, ancestral sampling from deep NADE is very expensive, corresponding to running through a neural net separately predicting each of the visible variables given some others. This work makes a connection between this criterion and the training criterion for Generative Stochastic Networks (GSNs). It shows that training NADEs in this way also trains a GSN, which defines a Markov chain associated with the NADE model. Based on this connection, we show an alternative way to sample from a trained Orderless NADE that allows to trade-off computing time and quality of the samples: a 3 to 10-fold speedup (taking into account the waste due to correlations between consecutive samples of the chain) can be obtained without noticeably reducing the quality of the samples. This is achieved using a novel sampling procedure for GSNs called annealed GSN sampling, similar to tempering methods that combines fast mixing (obtained thanks to steps at high noise levels) with accurate samples (obtained thanks to steps at low noise levels).

In this paper we propose and investigate a novel nonlinear unit, called $L_p$ unit, for deep neural networks. The proposed $L_p$ unit receives signals from several projections of a subset of units in the layer below and computes a normalized $L_p$ norm. We notice two interesting interpretations of the $L_p$ unit. First, the proposed unit can be understood as a generalization of a number of conventional pooling operators such as average, root-mean-square and max pooling widely used in, for instance, convolutional neural networks (CNN), HMAX models and neocognitrons. Furthermore, the $L_p$ unit is, to a certain degree, similar to the recently proposed maxout unit (Goodfellow et al., 2013) which achieved the state-of-the-art object recognition results on a number of benchmark datasets. Secondly, we provide a geometrical interpretation of the activation function based on which we argue that the $L_p$ unit is more efficient at representing complex, nonlinear separating boundaries. Each $L_p$ unit defines a superelliptic boundary, with its exact shape defined by the order $p$. We claim that this makes it possible to model arbitrarily shaped, curved boundaries more efficiently by combining a few $L_p$ units of different orders. This insight justifies the need for learning different orders for each unit in the model. We empirically evaluate the proposed $L_p$ units on a number of datasets and show that multilayer perceptrons (MLP) consisting of the $L_p$ units achieve the state-of-the-art results on a number of benchmark datasets. Furthermore, we evaluate the proposed $L_p$ unit on the recently proposed deep recurrent neural networks (RNN).

What do auto-encoders learn about the underlying data generating distribution? Recent work suggests that some auto-encoder variants do a good job of capturing the local manifold structure of data. This paper clarifies some of these previous observations by showing that minimizing a particular form of regularized reconstruction error yields a reconstruction function that locally characterizes the shape of the data generating density. We show that the auto-encoder captures the score (derivative of the log-density with respect to the input). It contradicts previous interpretations of reconstruction error as an energy function. Unlike previous results, the theorems provided here are completely generic and do not depend on the parametrization of the auto-encoder: they show what the auto-encoder would tend to if given enough capacity and examples. These results are for a contractive training criterion we show to be similar to the denoising auto-encoder training criterion with small corruption noise, but with contraction applied on the whole reconstruction function rather than just encoder. Similarly to score matching, one can consider the proposed training criterion as a convenient alternative to maximum likelihood because it does not involve a partition function. Finally, we show how an approximate Metropolis-Hastings MCMC can be setup to recover samples from the estimated distribution, and this is confirmed in sampling experiments.

Deep architecture such as hierarchical semi-Markov models is an important class of models for nested sequential data. Current exact inference schemes either cost cubic time in sequence length, or exponential time in model depth. These costs are prohibitive for large-scale problems with arbitrary length and depth. In this contribution, we propose a new approximation technique that may have the potential to achieve sub-cubic time complexity in length and linear time depth, at the cost of some loss of quality. The idea is based on two well-known methods: Gibbs sampling and Rao-Blackwellisation. We provide some simulation-based evaluation of the quality of the RGBS with respect to run time and sequence length.

Restricted Boltzmann Machines (RBM) [9, 5] have recently attracted an increasing attention for their rich capacity in a variety of learning tasks, including multivariate distribution modelling, feature extraction, classification, and construction of deep architectures [8, 19]. An RBM is a two-layer Markov random field in which the visible layer represents observed variables and the hidden layer represents latent aspects of the data. Pairwise interactions are only permitted for units between layers. As a result, the posterior distribution over the hidden variables and the probability of the data generative model are easy to evaluate, allowing fast feature extraction and efficient sampling-based inference [7]. Nonetheless, most existing work in RBMs implicitly assumes that the visible layer contains variables of the same modality. By far the most popular input types are binary [5] and Gaussian [8]. Recent extension includes categorical [21], ordinal [25], Poisson [6] and Beta [13] data. To the best of our knowledge, none has been considered for multicategorical and category-ranking data, nor for a mixed combination of these data types. In this paper, we investigate a generalisation of the RBM for variables of multiple modalities and types.

Restricted Boltzman Machines (RBMs) have been successfully used in recommender systems. However, as with most of other collaborative filtering techniques, it cannot solve cold start problems for there is no rating for a new item. In this paper, we first apply conditional RBM (CRBM) which could take extra information into account and show that CRBM could solve cold start problem very well, especially for rating prediction task. CRBM naturally combine the content and collaborative data under a single framework which could be fitted effectively. Experiments show that CRBM can be compared favourably with matrix factorization models, while hidden features learned from the former models are more easy to be interpreted.

Restricted Boltzmann machines (RBMs) have proved to be a versatile tool for a wide variety of machine learning tasks and as a building block for deep architectures [12, 24, 28]. The original proposals mainly handle binary visible and hidden units. Whilst binary hidden units are broadly applicable as feature detectors, non-binary visible data requires different designs. Recent extensions to other data types result in type-dependent models: the Gaussian for continuous inputs [12], Beta for bounded continuous inputs [16], Poisson for count data [9], multinomial for unordered categories [25], and ordinal models for ordered categories [37, 35]. The Boltzmann distribution permits several types to be jointly modelled, thus making the RBM a good tool for multimodal and complex social survey analysis. The work of [20, 29, 40] combines continuous (e.g., visual and audio) and discrete modalities (e.g., words). The work of [34] extends the idea further to incorporate ordinal and rank data. However, there are conceptual drawbacks: First, conditioned on the hidden layer, they are still separate type-specific models; second, handling ordered categories and ranks is not natural; and third, specifying direct correlation between these types remains difficult. The main thesis of this paper is that many data types can be captured in one unified model.

Restricted Boltzmann machines (RBMs) [36, 9, 20] have recently attracted significant interest due to their versatility in a variety of unsupervised and supervised learning tasks [35, 18, 25], and in building deep architectures [14, 31]. A RBM is a bipartite undirected model that captures the generative process in which a data vector is generated from a binary hidden vector. The bipartite architecture enables very fast data encoding and sampling-based inference; and together with recent advances in learning procedures, we can now process massive data with large models [13, 37, 2]. This paper presents our contributions in developing RBM specifications as well as learning and inference procedures for multivariate ordinal data. This extends and consolidates the reach of RBMs to a wide range of user-generated domains - social responses, recommender systems, product/paper reviews, and expert assessments of health and ecosystems indicators.