Transformer is the state of the art model for many language and visual tasks. In this paper, we give a deep analysis of its multi-head self-attention (MHSA) module and find that: 1) Each token is a random variable in high dimensional feature space. 2) After layer normalization, these variables are mapped to points on the hyper-sphere. 3) The update of these tokens is a Brownian motion. The Brownian motion has special properties, its second order item should not be ignored. So we present a new second-order optimizer(an iterative K-FAC algorithm) for the MHSA module. In some short words: All tokens are mapped to high dimension hyper-sphere. The Scaled Dot-Product Attention $softmax(frac{mathbf{Q}mathbf{K}^T}{sqrt{d}})$ is just the Markov transition matrix for the random walking on the sphere. And the deep learning process would learn proper kernel function to get proper positions of these tokens. The training process in the MHSA module corresponds to a Brownian motion worthy of further study.
Download