Mimetic Initialization of Self-Attention Layers

TL;DR

This paper introduces a simple, learning-free initialization method for self-attention layers in Transformers, inspired by pre-trained models, which improves training speed and accuracy on vision tasks like CIFAR-10 and ImageNet.

Contribution

The paper proposes a novel mimetic initialization scheme for self-attention layers that mimics pre-trained weights, enhancing training efficiency and performance without additional learning.

Findings

Over 5% accuracy improvement on CIFAR-10

Over 4% accuracy improvement on ImageNet

Faster training convergence with better final accuracy

Abstract

It is notoriously difficult to train Transformers on small datasets; typically, large pre-trained models are instead used as the starting point. We explore the weights of such pre-trained Transformers (particularly for vision) to attempt to find reasons for this discrepancy. Surprisingly, we find that simply initializing the weights of self-attention layers so that they "look" more like their pre-trained counterparts allows us to train vanilla Transformers faster and to higher final accuracies, particularly on vision tasks such as CIFAR-10 and ImageNet classification, where we see gains in accuracy of over 5% and 4%, respectively. Our initialization scheme is closed form, learning-free, and very simple: we set the product of the query and key weights to be approximately the identity, and the product of the value and projection weights to approximately the negative identity. As this mimics the patterns we saw in pre-trained Transformers, we call the technique "mimetic initialization".