Simple Recurrence Improves Masked Language Models

Lei Tao, Tian Ran, Bastings Jasmijn, Parikh Ankur P.. Arxiv 2022

[Paper]    
BERT Efficiency And Optimization Fine Tuning Masked Language Model Model Architecture Pretraining Methods RAG Training Techniques Transformer

In this work, we explore whether modeling recurrence into the Transformer architecture can both be beneficial and efficient, by building an extremely simple recurrent module into the Transformer. We compare our model to baselines following the training and evaluation recipe of BERT. Our results confirm that recurrence can indeed improve Transformer models by a consistent margin, without requiring low-level performance optimizations, and while keeping the number of parameters constant. For example, our base model achieves an absolute improvement of 2.1 points averaged across 10 tasks and also demonstrates increased stability in fine-tuning over a range of learning rates.

Similar Work