RealFormer: Transformer Likes Residual Attention
This work addresses the performance limitations of existing Transformer models for a broad range of NLP applications.
This paper proposes RealFormer, a technique to create Residual Attention Layer Transformer networks. It significantly outperforms canonical Transformers and their variants across a wide range of NLP tasks, including Masked Language Modeling, GLUE, SQuAD, and Neural Machine Translation.
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.