Teacher's pet: understanding and mitigating biases in distillation

TL;DR

This paper investigates biases in knowledge distillation, revealing that it can harm performance on certain subgroups, and proposes techniques to mitigate this issue while maintaining overall accuracy improvements.

Contribution

The paper identifies subgroup biases in distillation and introduces methods to reduce teacher influence on unreliable subgroups, improving fairness without sacrificing overall performance.

Findings

Distillation can negatively impact performance on data subgroups.

Teacher errors are amplified in the student model during distillation.

Mitigation techniques improve subgroup performance while preserving overall accuracy.

Abstract

Knowledge distillation is widely used as a means of improving the performance of a relatively simple student model using the predictions from a complex teacher model. Several works have shown that distillation significantly boosts the student's overall performance; however, are these gains uniform across all data subgroups? In this paper, we show that distillation can harm performance on certain subgroups, e.g., classes with few associated samples. We trace this behaviour to errors made by the teacher distribution being transferred to and amplified by the student model. To mitigate this problem, we present techniques which soften the teacher influence for subgroups where it is less reliable. Experiments on several image classification benchmarks show that these modifications of distillation maintain boost in overall accuracy, while additionally ensuring improvement in subgroup performance.