RealFormer: Transformer Likes Residual Attention

TL;DR

RealFormer introduces a residual attention mechanism in Transformer models, significantly improving performance and stability across various NLP tasks compared to standard Transformers.

Contribution

It proposes a simple residual attention layer for Transformers, enhancing performance and training stability over existing models like BERT and ETC.

Findings

Outperforms canonical Transformer on multiple NLP benchmarks

Stabilizes training and results in sparser attention patterns

Achieves superior results on tasks like GLUE, SQuAD, and translation

Abstract

Transformer is the backbone of modern NLP models. In this paper, we propose RealFormer, a simple and generic technique to create Residual Attention Layer Transformer networks that significantly outperform the canonical Transformer and its variants (BERT, ETC, etc.) on a wide spectrum of tasks including Masked Language Modeling, GLUE, SQuAD, Neural Machine Translation, WikiHop, HotpotQA, Natural Questions, and OpenKP. We also observe empirically that RealFormer stabilizes training and leads to models with sparser attention. Source code and pre-trained checkpoints for RealFormer can be found at https://github.com/google-research/google-research/tree/master/realformer.