Pay Attention when Required

TL;DR

This paper introduces the PAR Transformer, which reduces computational cost by replacing a significant portion of self-attention blocks with feed-forward blocks, maintaining performance across multiple language modeling benchmarks.

Contribution

The paper proposes the PAR Transformer architecture that optimizes block ordering and reduces compute time by replacing self-attention with feed-forward blocks, without sacrificing accuracy.

Findings

35% lower compute time compared to Transformer-XL

Retains perplexity on WikiText-103 benchmark

Validated on text8, enwiki8, and BERT models

Abstract

Transformer-based models consist of interleaved feed-forward blocks - that capture content meaning, and relatively more expensive self-attention blocks - that capture context meaning. In this paper, we explored trade-offs and ordering of the blocks to improve upon the current Transformer architecture and proposed PAR Transformer. It needs 35% lower compute time than Transformer-XL achieved by replacing ~63% of the self-attention blocks with feed-forward blocks, and retains the perplexity on WikiText-103 language modelling benchmark. We further validated our results on text8 and enwiki8 datasets, as well as on the BERT model.