Reproduce “Does RoBERTa Perform Better than BERT in Continual Learning: An Attention Sink Perspective” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xueying Bai, Yifan Sun, Niranjan Balasubramanian

This research identifies that "attention sinks," where non-semantic tokens receive high, undifferentiated attention, cause over-smoothing and cross-task interference in pre-trained language models during continual learning. A proposed "pre-scaling" mechanism encourages diverse attention and leads to substantial improvements in continual learning performance, enabling models like RoBERTa to better utilize their pre-trained capacity without relying on memory-intensive techniques.

Key insights

  • Attention sinks are ubiquitous in bidirectional PLMs, particularly in upper layers, and are a primary driver of over-smoothing by making token representations overly similar, which then causes cross-task interference.
  • RoBERTa's observed underperformance in continual learning, compared to BERT, is linked to its greater susceptibility to attention sink-induced over-smoothing due to lower initial sink attention deviations.
  • Encouraging diverse attention distribution, particularly on sink tokens, effectively mitigates over-smoothing and cross-task interference, allowing PLMs to retain knowledge better and leverage their pre-trained capabilities in continual learning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.