Does RoBERTa Perform Better than BERT in Continual Learning: An Attention Sink Perspective

TL;DR

This paper investigates how attention sink tokens in pre-trained models like RoBERTa and BERT affect continual learning performance, proposing a pre-scaling method to improve attention diversity and CL outcomes.

Contribution

It introduces a novel pre-scaling mechanism that reduces attention sink effects, enhancing continual learning performance of pre-trained models.

Findings

Pre-scaling improves CL performance without experience replay.

Attention sinks like [SEP] tokens impact model interference in sequential tasks.

Pre-trained models with attention sinks may underperform despite high capacity.

Abstract

Continual learning (CL) aims to train models that can sequentially learn new tasks without forgetting previous tasks' knowledge. Although previous works observed that pre-training can benefit CL, it remains unclear whether a pre-trained model with higher downstream capacity also performs better in CL. In this paper, we observe that pre-trained models may allocate high attention scores to some 'sink' tokens, such as [SEP] tokens, which are ubiquitous across various tasks. Such attention sinks may lead to models' over-smoothing in single-task learning and interference in sequential tasks' learning, which may compromise the models' CL performance despite their high pre-trained capabilities. To reduce these effects, we propose a pre-scaling mechanism that encourages attention diversity across all tokens. Specifically, it first scales the task's attention to the non-sink tokens in a probing stage, and then fine-tunes the model with scaling. Experiments show that pre-scaling yields substantial improvements in CL without experience replay, or progressively storing parameters from previous tasks.