Mixed Precision of Quantization of Transformer Language Models for Speech Recognition

State-of-the-art neural language models represented by Transformers are becoming increasingly complex and expensive for practical applications. Low-bit deep neural network quantization techniques provides a powerful solution to dramatically reduce their model size. Current low-bit quantization methods are based on uniform precision and fail to account for the varying performance sensitivity at different parts of the system to quantization errors. To this end, novel mixed precision DNN quantization methods are proposed in this paper. The optimal local precision settings are automatically learned using two techniques. The first is based on a quantization sensitivity metric in the form of Hessian trace weighted quantization perturbation. The second is based on mixed precision Transformer architecture search. Alternating direction methods of multipliers (ADMM) are used to efficiently train mixed precision quantized DNN systems. Experiments conducted on Penn Treebank (PTB) and a Switchboard corpus trained LF-MMI TDNN system suggest the proposed mixed precision Transformer quantization techniques achieved model size compression ratios of up to 16 times over the full precision baseline with no recognition performance degradation. When being used to compress a larger full precision Transformer LM with more layers, overall word error rate (WER) reductions up to 1.7% absolute (18% relative) were obtained.

Low-bit Quantization of Recurrent Neural Network Language Models Using Alternating Direction Methods of Multipliers

The high memory consumption and computational costs of Recurrent neural network language models (RNNLMs) limit their wider application on resource constrained devices. In recent years, neural network quantization techniques that are capable of producing extremely low-bit compression, for example, binarized RNNLMs, are gaining increasing research interests. Directly training of quantized neural networks is difficult. By formulating quantized RNNLMs training as an optimization problem, this paper presents a novel method to train quantized RNNLMs from scratch using alternating direction methods of multipliers (ADMM). This method can also flexibly adjust the trade-off between the compression rate and model performance using tied low-bit quantization tables. Experiments on two tasks: Penn Treebank (PTB), and Switchboard (SWBD) suggest the proposed ADMM quantization achieved a model size compression factor of up to 31 times over the full precision baseline RNNLMs. Faster convergence of 5 times in model training over the baseline binarized RNNLM quantization was also obtained. Index Terms: Language models, Recurrent neural networks, Quantization, Alternating direction methods of multipliers.

DHA: End-to-End Joint Optimization of Data Augmentation Policy, Hyper-parameter and Architecture

Automated machine learning (AutoML) usually involves several crucial components, such as Data Augmentation (DA) policy, Hyper-Parameter Optimization (HPO), and Neural Architecture Search (NAS). Although many strategies have been developed for automating these components in separation, joint optimization of these components remains challenging due to the largely increased search dimension and the variant input types of each component. Meanwhile, conducting these components in a sequence often requires careful coordination by human experts and may lead to sub-optimal results. In parallel to this, the common practice of searching for the optimal architecture first and then retraining it before deployment in NAS often suffers from low performance correlation between the search and retraining stages. An end-to-end solution that integrates the AutoML components and returns a ready-to-use model at the end of the search is desirable. In view of these, we propose DHA, which achieves joint optimization of Data augmentation policy, Hyper-parameter and Architecture. Specifically, end-to-end NAS is achieved in a differentiable manner by optimizing a compressed lower-dimensional feature space, while DA policy and HPO are updated dynamically at the same time. Experiments show that DHA achieves state-of-the-art (SOTA) results on various datasets, especially 77.4\% accuracy on ImageNet with cell based search space, which is higher than current SOTA by 0.5\%. To the best of our knowledge, we are the first to efficiently and jointly optimize DA policy, NAS, and HPO in an end-to-end manner without retraining.

Bayesian Transformer Language Models for Speech Recognition

State-of-the-art neural language models (LMs) represented by Transformers are highly complex. Their use of fixed, deterministic parameter estimates fail to account for model uncertainty and lead to over-fitting and poor generalization when given limited training data. In order to address these issues, this paper proposes a full Bayesian learning framework for Transformer LM estimation. Efficient variational inference based approaches are used to estimate the latent parameter posterior distributions associated with different parts of the Transformer model architecture including multi-head self-attention, feed forward and embedding layers. Statistically significant word error rate (WER) reductions up to 0.5\% absolute (3.18\% relative) and consistent perplexity gains were obtained over the baseline Transformer LMs on state-of-the-art Switchboard corpus trained LF-MMI factored TDNN systems with i-Vector speaker adaptation. Performance improvements were also obtained on a cross domain LM adaptation task requiring porting a Transformer LM trained on the Switchboard and Fisher data to a low-resource DementiaBank elderly speech corpus.

Understanding the wiring evolution in differentiable neural architecture search

Controversy exists on whether differentiable neural architecture search methods discover wiring topology effectively. To understand how wiring topology evolves, we study the underlying mechanism of several existing differentiable NAS frameworks. Our investigation is motivated by three observed searching patterns of differentiable NAS: 1) they search by growing instead of pruning; 2) wider networks are more preferred than deeper ones; 3) no edges are selected in bi-level optimization. To anatomize these phenomena, we propose a unified view on searching algorithms of existing frameworks, transferring the global optimization to local cost minimization. Based on this reformulation, we conduct empirical and theoretical analyses, revealing implicit inductive biases in the cost's assignment mechanism and evolution dynamics that cause the observed phenomena. These biases indicate strong discrimination towards certain topologies. To this end, we pose questions that future differentiable methods for neural wiring discovery need to confront, hoping to evoke a discussion and rethinking on how much bias has been enforced implicitly in existing NAS methods.

Neural Architecture Search for Speech Recognition

Deep neural networks (DNNs) based automatic speech recognition (ASR) systems are often designed using expert knowledge and empirical evaluation. In this paper, a range of neural architecture search (NAS) techniques are used to automatically learn two hyper-parameters that heavily affect the performance and model complexity of state-of-the-art factored time delay neural network (TDNN-F) acoustic models: i) the left and right splicing context offsets; and ii) the dimensionality of the bottleneck linear projection at each hidden layer. These include the standard DARTS method fully integrating the estimation of architecture weights and TDNN parameters in lattice-free MMI (LF-MMI) training; Gumbel-Softmax DARTS that reduces the confusion between candidate architectures; Pipelined DARTS that circumvents the overfitting of architecture weights using held-out data; and Penalized DARTS that further incorporates resource constraints to adjust the trade-off between performance and system complexity. Parameter sharing among candidate architectures was also used to facilitate efficient search over up to $7^{28}$ different TDNN systems. Experiments conducted on a 300-hour Switchboard conversational telephone speech recognition task suggest the NAS auto-configured TDNN-F systems consistently outperform the baseline LF-MMI trained TDNN-F systems using manual expert configurations. Absolute word error rate reductions up to 1.0% and relative model size reduction of 28% were obtained.

Bayesian x-vector: Bayesian Neural Network based x-vector System for Speaker Verification

Speaker verification systems usually suffer from the mismatch problem between training and evaluation data, such as speaker population mismatch, the channel and environment variations. In order to address this issue, it requires the system to have good generalization ability on unseen data. In this work, we incorporate Bayesian neural networks (BNNs) into the deep neural network (DNN) x-vector speaker verification system to improve the system's generalization ability. With the weight uncertainty modeling provided by BNNs, we expect the system could generalize better on the evaluation data and make verification decisions more accurately. Our experiment results indicate that the DNN x-vector system could benefit from BNNs especially when the mismatch problem is severe for evaluations using out-of-domain data. Specifically, results show that the system could benefit from BNNs by a relative EER decrease of 2.66% and 2.32% respectively for short- and long-utterance in-domain evaluations. Additionally, the fusion of DNN x-vector and Bayesian x-vector systems could achieve further improvement. Moreover, experiments conducted by out-of-domain evaluations, e.g. models trained on Voxceleb1 while evaluated on NIST SRE10 core test, suggest that BNNs could bring a larger relative EER decrease of around 4.69%.

DSNAS: Direct Neural Architecture Search without Parameter Retraining

If NAS methods are solutions, what is the problem? Most existing NAS methods require two-stage parameter optimization. However, performance of the same architecture in the two stages correlates poorly. In this work, we propose a new problem definition for NAS, task-specific end-to-end, based on this observation. We argue that given a computer vision task for which a NAS method is expected, this definition can reduce the vaguely-defined NAS evaluation to i) accuracy of this task and ii) the total computation consumed to finally obtain a model with satisfying accuracy. Seeing that most existing methods do not solve this problem directly, we propose DSNAS, an efficient differentiable NAS framework that simultaneously optimizes architecture and parameters with a low-biased Monte Carlo estimate. Child networks derived from DSNAS can be deployed directly without parameter retraining. Comparing with two-stage methods, DSNAS successfully discovers networks with comparable accuracy (74.4%) on ImageNet in 420 GPU hours, reducing the total time by more than 34%.