Accelerating Neural Transformer via an Average Attention Network

TL;DR

This paper introduces an average attention network to replace self-attention in Transformer decoders, significantly speeding up decoding without sacrificing translation quality, demonstrated on multiple language pairs.

Contribution

Proposes an average attention network as a faster alternative to self-attention in Transformer decoders, enabling over four times faster decoding with minimal performance loss.

Findings

Decoding speed increased over four times.

Maintained translation quality comparable to original Transformer.

Effective across six language pairs in WMT17 tasks.

Abstract

With parallelizable attention networks, the neural Transformer is very fast to train. However, due to the auto-regressive architecture and self-attention in the decoder, the decoding procedure becomes slow. To alleviate this issue, we propose an average attention network as an alternative to the self-attention network in the decoder of the neural Transformer. The average attention network consists of two layers, with an average layer that models dependencies on previous positions and a gating layer that is stacked over the average layer to enhance the expressiveness of the proposed attention network. We apply this network on the decoder part of the neural Transformer to replace the original target-side self-attention model. With masking tricks and dynamic programming, our model enables the neural Transformer to decode sentences over four times faster than its original version with almost no loss in training time and translation performance. We conduct a series of experiments on WMT17 translation tasks, where on 6 different language pairs, we obtain robust and consistent speed-ups in decoding.