Similarity Transfer for Knowledge Distillation

TL;DR

This paper introduces a novel similarity transfer method for knowledge distillation that leverages instance similarity correlations and mixup techniques to improve the performance of compact neural networks.

Contribution

It proposes a new distillation approach that fully utilizes category similarities and instance correlations through mixup, outperforming existing methods.

Findings

Significant accuracy improvements on CIFAR-10, CIFAR-100, CINIC-10, and Tiny-ImageNet.

Outperforms vanilla knowledge distillation and state-of-the-art methods.

Effective use of virtual samples created by mixup enhances student model performance.

Abstract

Knowledge distillation is a popular paradigm for learning portable neural networks by transferring the knowledge from a large model into a smaller one. Most existing approaches enhance the student model by utilizing the similarity information between the categories of instance level provided by the teacher model. However, these works ignore the similarity correlation between different instances that plays an important role in confidence prediction. To tackle this issue, we propose a novel method in this paper, called similarity transfer for knowledge distillation (STKD), which aims to fully utilize the similarities between categories of multiple samples. Furthermore, we propose to better capture the similarity correlation between different instances by the mixup technique, which creates virtual samples by a weighted linear interpolation. Note that, our distillation loss can fully utilize the incorrect classes similarities by the mixed labels. The proposed approach promotes the performance of student model as the virtual sample created by multiple images produces a similar probability distribution in the teacher and student networks. Experiments and ablation studies on several public classification datasets including CIFAR-10,CIFAR-100,CINIC-10 and Tiny-ImageNet verify that this light-weight method can effectively boost the performance of the compact student model. It shows that STKD substantially has outperformed the vanilla knowledge distillation and has achieved superior accuracy over the state-of-the-art knowledge distillation methods.