Setting the Record Straight on Transformer Oversmoothing
This work addresses a fundamental limitation in Transformer models for ML researchers and practitioners, providing insights to improve model design, though it is incremental in refining existing theoretical understanding.
The paper challenges the claim that Transformers inevitably oversmooth with increasing depth, showing that feature similarity increase is not universal and depends on weight eigenspectrum and layer normalization signs, with a simple parameterization method to control smoothing.
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.