CVAIMar 16, 2022

Decoupled Knowledge Distillation

arXiv:2203.08679v2845 citationsh-index: 14Has Code
Originality Incremental advance
AI Analysis

This work addresses the problem of inefficient knowledge transfer in machine learning for researchers and practitioners, offering a more flexible and effective distillation method, though it is incremental as it builds on existing logit distillation concepts.

The paper tackles the overlooked potential of logit distillation in knowledge distillation by reformulating it into target and non-target class components, revealing limitations in classical methods and proposing Decoupled Knowledge Distillation (DKD), which achieves comparable or better results with higher efficiency on datasets like CIFAR-100 and ImageNet.

State-of-the-art distillation methods are mainly based on distilling deep features from intermediate layers, while the significance of logit distillation is greatly overlooked. To provide a novel viewpoint to study logit distillation, we reformulate the classical KD loss into two parts, i.e., target class knowledge distillation (TCKD) and non-target class knowledge distillation (NCKD). We empirically investigate and prove the effects of the two parts: TCKD transfers knowledge concerning the "difficulty" of training samples, while NCKD is the prominent reason why logit distillation works. More importantly, we reveal that the classical KD loss is a coupled formulation, which (1) suppresses the effectiveness of NCKD and (2) limits the flexibility to balance these two parts. To address these issues, we present Decoupled Knowledge Distillation (DKD), enabling TCKD and NCKD to play their roles more efficiently and flexibly. Compared with complex feature-based methods, our DKD achieves comparable or even better results and has better training efficiency on CIFAR-100, ImageNet, and MS-COCO datasets for image classification and object detection tasks. This paper proves the great potential of logit distillation, and we hope it will be helpful for future research. The code is available at https://github.com/megvii-research/mdistiller.

Code Implementations1 repo
Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes