Structured in Space, Randomized in Time: Leveraging Dropout in RNNs for Efficient Training

TL;DR

This paper introduces a structured dropout method for LSTM RNNs that induces sparsity in hidden states, enabling significant training speedups without loss of accuracy across multiple NLP tasks.

Contribution

The authors propose a novel dropout pattern that creates structured sparsity in LSTM hidden states, facilitating efficient computation reduction during training.

Findings

Achieved 1.23x to 1.64x faster training times

Maintained comparable performance metrics

Validated across language modeling, translation, and NER tasks

Abstract

Recurrent Neural Networks (RNNs), more specifically their Long Short-Term Memory (LSTM) variants, have been widely used as a deep learning tool for tackling sequence-based learning tasks in text and speech. Training of such LSTM applications is computationally intensive due to the recurrent nature of hidden state computation that repeats for each time step. While sparsity in Deep Neural Nets has been widely seen as an opportunity for reducing computation time in both training and inference phases, the usage of non-ReLU activation in LSTM RNNs renders the opportunities for such dynamic sparsity associated with neuron activation and gradient values to be limited or non-existent. In this work, we identify dropout induced sparsity for LSTMs as a suitable mode of computation reduction. Dropout is a widely used regularization mechanism, which randomly drops computed neuron values during each iteration of training. We propose to structure dropout patterns, by dropping out the same set of physical neurons within a batch, resulting in column (row) level hidden state sparsity, which are well amenable to computation reduction at run-time in general-purpose SIMD hardware as well as systolic arrays. We conduct our experiments for three representative NLP tasks: language modelling on the PTB dataset, OpenNMT based machine translation using the IWSLT De-En and En-Vi datasets, and named entity recognition sequence labelling using the CoNLL-2003 shared task. We demonstrate that our proposed approach can be used to translate dropout-based computation reduction into reduced training time, with improvement ranging from 1.23x to 1.64x, without sacrificing the target metric.