Toward Student-Oriented Teacher Network Training For Knowledge Distillation
This work addresses a fundamental discrepancy in knowledge distillation for machine learning practitioners, offering an incremental improvement by optimizing teacher training for student performance.
The paper tackles the problem that high-performing teacher networks do not always produce the best students in knowledge distillation, proposing a teacher training method called SoTeacher that incorporates Lipschitz and consistency regularization into empirical risk minimization to better approximate true label distributions. Experiments on benchmark datasets show that SoTeacher consistently improves student accuracy.
How to conduct teacher training for knowledge distillation is still an open problem. It has been widely observed that a best-performing teacher does not necessarily yield the best-performing student, suggesting a fundamental discrepancy between the current teacher training practice and the ideal teacher training strategy. To fill this gap, we explore the feasibility of training a teacher that is oriented toward student performance with empirical risk minimization (ERM). Our analyses are inspired by the recent findings that the effectiveness of knowledge distillation hinges on the teacher's capability to approximate the true label distribution of training inputs. We theoretically establish that the ERM minimizer can approximate the true label distribution of training data as long as the feature extractor of the learner network is Lipschitz continuous and is robust to feature transformations. In light of our theory, we propose a teacher training method SoTeacher which incorporates Lipschitz regularization and consistency regularization into ERM. Experiments on benchmark datasets using various knowledge distillation algorithms and teacher-student pairs confirm that SoTeacher can improve student accuracy consistently.