Setting the Record Straight on Transformer Oversmoothing

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Dovonon, Gbètondji J-S, Bronstein, Michael M., Kusner, Matt J.
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914889230974976
author Dovonon, Gbètondji J-S
Bronstein, Michael M.
Kusner, Matt J.
author_facet Dovonon, Gbètondji J-S
Bronstein, Michael M.
Kusner, Matt J.
contents Transformer-based models have recently become wildly successful across a diverse set of domains. At the same time, recent work has shown empirically and theoretically that Transformers are inherently limited. Specifically, they argue that as model depth increases, Transformers oversmooth, i.e., inputs become more and more similar. A natural question is: How can Transformers achieve these successes given this shortcoming? In this work we test these observations empirically and theoretically and uncover a number of surprising findings. We find that there are cases where feature similarity increases but, contrary to prior results, this is not inevitable, even for existing pre-trained models. Theoretically, we show that smoothing behavior depends on the eigenspectrum of the value and projection weights. We verify this empirically and observe that the sign of layer normalization weights can influence this effect. Our analysis reveals a simple way to parameterize the weights of the Transformer update equations to influence smoothing behavior. We hope that our findings give ML researchers and practitioners additional insight into how to develop future Transformer-based models.
format Preprint
id arxiv_https___arxiv_org_abs_2401_04301
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Setting the Record Straight on Transformer Oversmoothing
Dovonon, Gbètondji J-S
Bronstein, Michael M.
Kusner, Matt J.
Machine Learning
Transformer-based models have recently become wildly successful across a diverse set of domains. At the same time, recent work has shown empirically and theoretically that Transformers are inherently limited. Specifically, they argue that as model depth increases, Transformers oversmooth, i.e., inputs become more and more similar. A natural question is: How can Transformers achieve these successes given this shortcoming? In this work we test these observations empirically and theoretically and uncover a number of surprising findings. We find that there are cases where feature similarity increases but, contrary to prior results, this is not inevitable, even for existing pre-trained models. Theoretically, we show that smoothing behavior depends on the eigenspectrum of the value and projection weights. We verify this empirically and observe that the sign of layer normalization weights can influence this effect. Our analysis reveals a simple way to parameterize the weights of the Transformer update equations to influence smoothing behavior. We hope that our findings give ML researchers and practitioners additional insight into how to develop future Transformer-based models.
title Setting the Record Straight on Transformer Oversmoothing
topic Machine Learning
url https://arxiv.org/abs/2401.04301