Knowledge Distillation By Sparse Representation Matching

TL;DR

This paper introduces Sparse Representation Matching (SRM), a novel knowledge distillation method that transfers intermediate features from teacher to student CNNs using sparse representations, improving robustness and performance.

Contribution

SRM is a new plug-and-play neural processing block that effectively transfers intermediate knowledge via sparse representations, outperforming existing KD methods.

Findings

SRM outperforms other knowledge distillation techniques on multiple datasets.

SRM is robust to architectural differences between teacher and student networks.

SRM can be integrated efficiently into any CNN using stochastic gradient descent.

Abstract

Knowledge Distillation refers to a class of methods that transfers the knowledge from a teacher network to a student network. In this paper, we propose Sparse Representation Matching (SRM), a method to transfer intermediate knowledge obtained from one Convolutional Neural Network (CNN) to another by utilizing sparse representation learning. SRM first extracts sparse representations of the hidden features of the teacher CNN, which are then used to generate both pixel-level and image-level labels for training intermediate feature maps of the student network. We formulate SRM as a neural processing block, which can be efficiently optimized using stochastic gradient descent and integrated into any CNN in a plug-and-play manner. Our experiments demonstrate that SRM is robust to architectural differences between the teacher and student networks, and outperforms other KD techniques across several datasets.