Difference-Masking: Choosing What to Mask in Continued Pretraining
This work addresses the challenge of efficiently adapting pretrained models to new domains for researchers and practitioners, though it is incremental as it builds on existing masking-and-predicting methods.
The paper tackled the problem of improving continued pretraining by selecting which tokens to mask, rather than using random masking, and found that their Difference-Masking strategy outperformed baselines across four diverse language-only and multimodal video tasks.
The self-supervised objective of masking-and-predicting has led to promising performance gains on a variety of downstream tasks. However, while most approaches randomly mask tokens, there is strong intuition that deciding what to mask can substantially improve learning outcomes. We investigate this in continued pretraining setting in which pretrained models continue to pretrain on domain-specific data before performing some downstream task. We introduce Difference-Masking, a masking strategy that automatically chooses what to mask during continued pretraining by considering what makes a task domain different from the pretraining domain. Empirically, we find that Difference-Masking outperforms baselines on continued pretraining settings across four diverse language-only and multimodal video tasks.