Abstract:Deep neural models have achieved remarkable performance on various supervised and unsupervised learning tasks, but it is a challenge to deploy these large-size networks on resource-limited devices. As a representative type of model compression and acceleration methods, knowledge distillation (KD) solves this problem by transferring knowledge from heavy teachers to lightweight students. However, most distillation methods focus on imitating the responses of teacher networks but ignore the information redundancy of student networks. In this article, we propose a novel distillation framework difference-based channel contrastive distillation (DCCD), which introduces channel contrastive knowledge and dynamic difference knowledge into student networks for redundancy reduction. At the feature level, we construct an efficient contrastive objective that broadens student networks' feature expression space and preserves richer information in the feature extraction stage. At the final output level, more detailed knowledge is extracted from teacher networks by making a difference between multiview augmented responses of the same instance. We enhance student networks to be more sensitive to minor dynamic changes. With the improvement of two aspects of DCCD, the student network gains contrastive and difference knowledge and reduces its overfitting and redundancy. Finally, we achieve surprising results that the student approaches and even outperforms the teacher in test accuracy on CIFAR-100. We reduce the top-1 error to 28.16% on ImageNet classification and 24.15% for cross-model transfer with ResNet-18. Empirical experiments and ablation studies on popular datasets show that our proposed method can achieve state-of-the-art accuracy compared with other distillation methods.

Using Less but Important Information for Feature Distillation

DCCD: Reducing Neural Network Redundancy Via Distillation

A Simple and Generic Framework for Feature Distillation via Channel-wise Transformation

Improved Feature Distillation via Projector Ensemble

Self-boosting for Feature Distillation

Exploring Graph-based Knowledge: Multi-Level Feature Distillation via Channels Relational Graph

Fixing the Teacher-Student Knowledge Discrepancy in Distillation

Function-Consistent Feature Distillation

Show, Attend and Distill:Knowledge Distillation via Attention-based Feature Matching

A Comprehensive Overhaul of Feature Distillation

Customizing a Teacher for Feature Distillation

Feature Map Distillation of Thin Nets for Low-Resolution Object Recognition

Multistage feature fusion knowledge distillation

Tree-like Decision Distillation

In Defense of Feature Mimicking for Knowledge Distillation.

Knowledge Distillation with the Reused Teacher Classifier

DFD: Distillng the Feature Disparity Differently for Detectors

Distilling Knowledge by Mimicking Features

Knowledge distillation based on multi-layer fusion features

Differentiable Dynamic Channel Association for Knowledge Distillation

Knowledge Fusion Distillation: Improving Distillation with Multi-scale Attention Mechanisms