Parallel Training of GRU Networks with a Multi-Grid Solver for Long Sequences

TL;DR

This paper introduces a novel parallel training method for GRU networks using a multigrid solver, significantly speeding up training on long sequences by hierarchical correction of hidden states.

Contribution

The paper presents a new parallel-in-time training scheme for GRUs based on multigrid reduction, enabling efficient training of very long sequences.

Findings

Achieves up to 6.5x speedup over serial training.

Performance improves with increasing sequence length.

Effective hierarchical correction accelerates end-to-end communication.

Abstract

Parallelizing Gated Recurrent Unit (GRU) networks is a challenging task, as the training procedure of GRU is inherently sequential. Prior efforts to parallelize GRU have largely focused on conventional parallelization strategies such as data-parallel and model-parallel training algorithms. However, when the given sequences are very long, existing approaches are still inevitably performance limited in terms of training time. In this paper, we present a novel parallel training scheme (called parallel-in-time) for GRU based on a multigrid reduction in time (MGRIT) solver. MGRIT partitions a sequence into multiple shorter sub-sequences and trains the sub-sequences on different processors in parallel. The key to achieving speedup is a hierarchical correction of the hidden state to accelerate end-to-end communication in both the forward and backward propagation phases of gradient descent. Experimental results on the HMDB51 dataset, where each video is an image sequence, demonstrate that the new parallel training scheme achieves up to 6.5$\times$ speedup over a serial approach. As efficiency of our new parallelization strategy is associated with the sequence length, our parallel GRU algorithm achieves significant performance improvement as the sequence length increases.