Explicit and Implicit Knowledge Distillation via Unlabeled Data

TL;DR

This paper introduces an efficient data-free knowledge distillation approach that uses unlabeled sample selection and a class-dropping mechanism to improve training efficiency and accuracy without relying on high-cost generators.

Contribution

It proposes a novel unlabeled sample selection method combined with class-dropping and a structured relation-based distillation technique, reducing computational costs and enhancing distillation effectiveness.

Findings

Faster convergence compared to existing methods

Higher accuracy achieved in experiments

Reduced computational costs without sacrificing performance

Abstract

Data-free knowledge distillation is a challenging model lightweight task for scenarios in which the original dataset is not available. Previous methods require a lot of extra computational costs to update one or more generators and their naive imitate-learning lead to lower distillation efficiency. Based on these observations, we first propose an efficient unlabeled sample selection method to replace high computational generators and focus on improving the training efficiency of the selected samples. Then, a class-dropping mechanism is designed to suppress the label noise caused by the data domain shifts. Finally, we propose a distillation method that incorporates explicit features and implicit structured relations to improve the effect of distillation. Experimental results show that our method can quickly converge and obtain higher accuracy than other state-of-the-art methods.