ReMem: Mutual Information-Aware Fine-tuning of Pretrained Vision Transformers for Effective Knowledge Distillation

TL;DR

This paper proposes a mutual information-aware fine-tuning method for pretrained Vision Transformers to enhance knowledge distillation, especially from strong models, by reweighting MLP blocks to improve transfer effectiveness.

Contribution

It introduces a novel mutual information-aware optimization technique and a heuristic reweighting of MLP blocks to improve knowledge transfer from pretrained ViTs.

Findings

Enhanced knowledge distillation from strong pretrained models.

Reweighting MLP blocks improves transfer on small or imbalanced datasets.

Method benefits small student models significantly.

Abstract

Knowledge distillation from pretrained visual representation models offers an effective approach to improve small, task-specific production models. However, the effectiveness of such knowledge transfer drops significantly when distilling from strong models that are pretrained in a large scale. In this paper, we address this challenge for pretrained Vision Transformers (ViTs) by exploring methods to fine-tune them for more effective knowledge transfer. Motivated by the connection between mutual information and distillation effectiveness, we propose to employ mutual information-aware optimization during finetuning. For small or highly-imbalanced downstream datasets where such optimization becomes less effective, we introduce a simple yet effective heuristic of reweighting MLP blocks. This approach is inspired by our observation that top MLP blocks are primarily responsible for mutual information loss. Our method enables small student models to benefit from those pretrained models among the strongest.