Logit Standardization in Knowledge Distillation

Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, Xiaochun Cao

arXiv:2403.01427 · 2026-07-27 공개 · arXiv · PDF

knowledge-distillation model-capacity imagenet cifar-100 teacher-student-models distillation-evaluation logit-standardization softmax-temperature

Abstract

Knowledge distillation involves transferring soft labels from a teacher to a student using a shared temperature-based softmax function. However, the assumption of a shared temperature between teacher and student implies a mandatory exact match between their logits in terms of logit range and variance. This side-effect limits the performance of student, considering the capacity discrepancy between them and the finding that the innate logit relations of teacher are sufficient for student to learn. To address this issue, we propose setting the temperature as the weighted standard deviation of logit and performing a plug-and-play Z-score pre-process of logit standardization before applying softmax and Kullback-Leibler divergence. Our pre-process enables student to focus on essential logit relationsfrom teacher rather than requiring a magnitude match, and can improve the performance of existing logit-based distillation methods. We also show a typical case where the conventional setting of sharing temperature between teacher and student cannot reliably yield the authentic dis-tillation evaluation; nonetheless, this challenge is success-fully alleviated by our Z-score. We extensively evaluate our method for various student and teacher models on CIFAR-100 and ImageNet, showing its significant superiority. The vanilla knowledge distillation powered by our pre-process can achieve favorable performance against state-of-the-art methods, and other distillation variants can obtain considerable gain with the assistance of our pre-process. The codes, pre-trained models and logs are released on Github.

한국어 요약

한 줄 요약

로짓 표준화를 통해 지식 증류 성능을 향상시키는 새로운 전처리 방법을 제안한다.

핵심 기여도

핵심 아이디어

기존 지식 증류(KD)는 교사와 학습자의 로짓에 동일한 온도를 적용하는 방식을 사용하지만, 이는 로짓의 범위와 분산이 일치해야 한다는 암묵적 가정을 내포한다. 이는 특히 학습자의 용량이 작을 경우 성능 저하를 초래한다. 본 연구는 정보 이론의 엔트로피 최대화 원리를 기반으로, 온도가 라그랑주 승수에서 유도된다는 점을 이론적으로 증명하고, 이에 따라 교사와 학습자, 샘플 간에 온도를 다르게 설정할 수 있음을 보인다. 이를 바탕으로 로짓 표준편차를 기반으로 한 가변 온도와 Z-score 전처리를 제안하여, 학습자가 로짓의 절대 크기보다는 상대적 관계에 집중할 수 있도록 유도한다.

기술적 접근법

주요 결과

의의 및 한계

실용적 활용