On student-teacher deviations in distillation: does it pay to disobey?

TL;DR

This paper investigates why student networks in knowledge distillation sometimes outperform teachers despite deviations from teacher probabilities, revealing that exaggeration of confidence and bias can enhance generalization.

Contribution

It characterizes student-teacher deviations in KD, linking confidence exaggeration and implicit bias to improved student performance and generalization.

Findings

Student exaggerates teacher's confidence levels.

KD accelerates convergence along top data eigendirections.

Exaggerated bias and confidence contribute to better generalization.

Abstract

Knowledge distillation (KD) has been widely used to improve the test accuracy of a "student" network, by training it to mimic the soft probabilities of a trained "teacher" network. Yet, it has been shown in recent work that, despite being trained to fit the teacher's probabilities, the student may not only significantly deviate from the teacher probabilities, but may also outdo than the teacher in performance. Our work aims to reconcile this seemingly paradoxical observation. Specifically, we characterize the precise nature of the student-teacher deviations, and argue how they can co-occur with better generalization. First, through experiments on image and language data, we identify that these probability deviations correspond to the student systematically exaggerating the confidence levels of the teacher. Next, we theoretically and empirically establish another form of exaggeration in some simple settings: KD exaggerates the implicit bias of gradient descent in converging faster along the top eigendirections of the data. Finally, we tie these two observations together: we demonstrate that the exaggerated bias of KD can simultaneously result in both (a) the exaggeration of confidence and (b) the improved generalization of the student, thus offering a resolution to the apparent paradox. Our analysis brings existing theory and practice closer by considering the role of gradient descent in KD and by demonstrating the exaggerated bias effect in both theoretical and empirical settings.