Philemon Brakel

Towards End-to-End Speech Recognition with Deep Convolutional Neural Networks.

Convolutional Neural Networks (CNNs) are effective models for reducing spectral variations and modeling spectral correlations in acoustic features for automatic speech recognition (ASR). Hybrid speech recognition systems incorporating CNNs with Hidden Markov Models/Gaussian Mixture Models (HMMs/GMMs) have achieved the state-of-the-art in various benchmarks. Meanwhile, Connectionist Temporal Classification (CTC) with Recurrent Neural Networks (RNNs), which is proposed for labeling unsegmented sequences, makes it feasible to train an end-to-end speech recognition system instead of hybrid settings. However, RNNs are computationally expensive and sometimes difficult to train. In this paper, inspired by the advantages of both CNNs and the CTC approach, we propose an end-to-end speech framework for sequence labeling, by combining hierarchical CNNs with CTC directly without recurrent connections. By evaluating the approach on the TIMIT phoneme recognition task, we show that the proposed model is not only computationally efficient, but also competitive with the existing baseline systems. Moreover, we argue that CNNs have the capability to model temporal correlations with appropriate context information.

Training restricted boltzmann machines with multi-tempering: harnessing parallelization

Restricted Boltzmann Machines (RBMs) are unsupervised probabilistic neural networks that can be stacked to form Deep Belief Networks. Given the recent popularity of RBMs and the increasing availability of parallel computing architectures, it becomes interesting to investigate learning algorithms for RBMs that benefit from parallel computations. In this paper, we look at two extensions of the parallel tempering algorithm, which is a Markov Chain Monte Carlo method to approximate the likelihood gradient. The first extension is directed at a more effective exchange of information among the parallel sampling chains. The second extension estimates gradients by averaging over chains from different temperatures. We investigate the efficiency of the proposed methods and demonstrate their usefulness on the MNIST dataset. Especially the weighted averaging seems to benefit Maximum Likelihood learning.

Energy-based temporal neural networks for imputing missing values

Imputing missing values in high dimensional time series is a difficult problem. There have been some approaches to the problem [11,8] where neural architectures were trained as probabilistic models of the data. However, we argue that this approach is not optimal. We propose to view temporal neural networks with latent variables as energy-based models and train them for missing value recovery directly. In this paper we introduce two energy-based models. The first model is based on a one dimensional convolution and the second model utilizes a recurrent neural network. We demonstrate how ideas from the energy-based learning framework can be used to train these models to recover missing values. The models are evaluated on a motion capture dataset.