Reproduce “BERT Busters: Outlier Dimensions that Disrupt Transformers” 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
Olga Kovaleva, Saurabh Kulshreshtha, Anna Rogers, Anna Rumshisky
Pre-trained Transformer models exhibit extreme fragility to the targeted removal of a minuscule fraction of specific LayerNormalization parameters, less than 0.0001% of total weights, which leads to disproportionate performance degradation across various tasks. The study from the University of Massachusetts Lowell and the University of Copenhagen shows these critical "outlier dimensions" are consistent across layers and emerge early during pre-training.
Key insights
- Transformer models are profoundly fragile to the targeted removal of a minute fraction of specific LayerNorm parameters, challenging the prevailing notion of their broad robustness to weight pruning.
- These critical "outlier dimensions" are not just high in magnitude but also distinguished by their consistent dimensional position across layers and appear to anchor the embedding space.
- Outlier dimensions emerge relatively early in the pre-training process, with their development strongly correlating with a rapid decrease in language modeling loss and perplexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.