PoWER-BERT: Accelerating BERT Inference via Progressive Word-vector Elimination

TL;DR

PoWER-BERT is a method that accelerates BERT inference by eliminating redundant word-vectors based on self-attention significance, achieving up to 4.5x faster inference with minimal accuracy loss.

Contribution

It introduces a novel strategy for reducing BERT inference time by selectively removing redundant word-vectors using a learned significance measure.

Findings

Up to 4.5x reduction in BERT inference time with <1% accuracy loss

Achieves up to 6.8x speedup on ALBERT with minimal accuracy impact

Outperforms prior methods in accuracy-time trade-off

Abstract

We develop a novel method, called PoWER-BERT, for improving the inference time of the popular BERT model, while maintaining the accuracy. It works by: a) exploiting redundancy pertaining to word-vectors (intermediate encoder outputs) and eliminating the redundant vectors. b) determining which word-vectors to eliminate by developing a strategy for measuring their significance, based on the self-attention mechanism. c) learning how many word-vectors to eliminate by augmenting the BERT model and the loss function. Experiments on the standard GLUE benchmark shows that PoWER-BERT achieves up to 4.5x reduction in inference time over BERT with <1% loss in accuracy. We show that PoWER-BERT offers significantly better trade-off between accuracy and inference time compared to prior methods. We demonstrate that our method attains up to 6.8x reduction in inference time with <1% loss in accuracy when applied over ALBERT, a highly compressed version of BERT. The code for PoWER-BERT is publicly available at https://github.com/IBM/PoWER-BERT.