Estimating and Maximizing Mutual Information for Knowledge Distillation

TL;DR

This paper introduces a mutual information maximization approach for knowledge distillation that enhances low-capacity models by transferring knowledge from more complex models, improving accuracy across various architectures.

Contribution

The paper presents a flexible contrastive-based method to estimate and maximize mutual information for knowledge transfer between arbitrary teacher and student networks.

Findings

Outperforms existing methods across diverse network pairs

Achieves 74.55% accuracy on CIFAR100 with ShufflenetV2

Improves ImageNet ResNet-18 accuracy from 68.88% to 70.32%

Abstract

In this work, we propose Mutual Information Maximization Knowledge Distillation (MIMKD). Our method uses a contrastive objective to simultaneously estimate and maximize a lower bound on the mutual information of local and global feature representations between a teacher and a student network. We demonstrate through extensive experiments that this can be used to improve the performance of low capacity models by transferring knowledge from more performant but computationally expensive models. This can be used to produce better models that can be run on devices with low computational resources. Our method is flexible, we can distill knowledge from teachers with arbitrary network architectures to arbitrary student networks. Our empirical results show that MIMKD outperforms competing approaches across a wide range of student-teacher pairs with different capacities, with different architectures, and when student networks are with extremely low capacity. We are able to obtain 74.55% accuracy on CIFAR100 with a ShufflenetV2 from a baseline accuracy of 69.8% by distilling knowledge from ResNet-50. On Imagenet we improve a ResNet-18 network from 68.88% to 70.32% accuracy (1.44%+) using a ResNet-34 teacher network.