0

Normalized Attention Without Probability Cage

Replacing softmax in self-attention with normalization yields a bias-robust architecture that outperforms traditional attention mechanisms in empirical studies.

Year
2020
Venue
normalized-attention-without-probability-cage-1
Authors
2
Hosting
Abstract onlyARXIV-DEFAULT

Cite

Notes

Only stored in your browser.

Attribution

Abstract & full text
arxiv.org/abs/2005.09561ARXIV-DEFAULT
TL;DR
Semantic Scholar
Attribution policy →

Abstract

Attention architectures are widely used; they recently gained renewed popularity with Transformers yielding a streak of state of the art results. Yet, the geometrical implications of softmax-attention remain largely unexplored. In this work we highlight the limitations of constraining attention weights to the probability simplex and the resulting convex hull of value vectors. We show that Transformers are sequence length dependent biased towards token isolation at initialization and contrast Transformers to simple max- and sum-pooling - two strong baselines rarely reported. We propose to replace the softmax in self-attention with normalization, yielding a hyperparameter and data-bias robust, generally applicable architecture. We support our insights with empirical results from more than 25,000 trained models. All results and implementations are made available.

Authors

2