Setting the Record Straight on Transformer Oversmoothing
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: transformers, oversmoothing, rank collapse
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Transformer-based models have recently become wildly successful across a diverse set of domains. At the same time, recent work has shown that Transformers are inherently low-pass filters that can oversmooth the input. This causes their performance to quickly saturate as model depth increases. A natural question is: How can Transformers achieve success given this shortcoming? In this work we show that in fact Transformers are not inherently low-pass filters. Instead, whether Transformers oversmooth or not depends on the eigenspectrum of their update equations. Our analysis extends prior work in oversmoothing and in the closely-related phenomenon of rank collapse. We show that many successful Transformer models have attention and weights which satisfy conditions that avoid oversmoothing. Finally, we describe a simple way to reparameterize the weights of the Transformer update equations to ensure that oversmoothing does not occur. Compared to other solutions for oversmoothing, our approach does not require a new architecture, or any additional hyperparameters.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8455
Loading