A Probabilistic Interpretation of Transformers
Machine Learning
2022-05-03 v1
Abstract
We propose a probabilistic interpretation of exponential dot product attention of transformers and contrastive learning based off of exponential families. The attention sublayer of transformers is equivalent to a gradient ascent step of the log normalizer, which is the log-sum-exp term in the Hopfield theory of attention. This ascent step induces a parallel expansion of points, which is counterbalanced by a contraction from layer normalization. We also state theoretical limitations of our theory and the Hopfield theory and suggest directions for resolution.
Cite
@article{arxiv.2205.01080,
title = {A Probabilistic Interpretation of Transformers},
author = {Alexander Shim},
journal= {arXiv preprint arXiv:2205.01080},
year = {2022}
}
Comments
Accepted in ICML 2021 Workshop: Self-Supervised Learning for Reasoning and Perception