Anisotropy Is Inherent to Self-Attention in Transformers
This work addresses a fundamental issue in self-supervised learning for researchers and practitioners using Transformers, as it highlights a pervasive problem affecting model performance across domains, though it is incremental in building on prior observations.
The paper demonstrates that anisotropy, a representation degeneration problem where hidden representations become unexpectedly close in angular distance, is empirically observed in Transformers across various objectives and modalities, suggesting it is inherent to these models rather than solely a consequence of cross-entropy loss on long-tailed token distributions.
The representation degeneration problem is a phenomenon that is widely observed among self-supervised learning methods based on Transformers. In NLP, it takes the form of anisotropy, a singular property of hidden representations which makes them unexpectedly close to each other in terms of angular distance (cosine-similarity). Some recent works tend to show that anisotropy is a consequence of optimizing the cross-entropy loss on long-tailed distributions of tokens. We show in this paper that anisotropy can also be observed empirically in language models with specific objectives that should not suffer directly from the same consequences. We also show that the anisotropy problem extends to Transformers trained on other modalities. Our observations suggest that anisotropy is actually inherent to Transformers-based models.