AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

Function-Consistent Feature Distillation

Dongyang Liu, Meina Kan, Shiguang Shan, Xilin CHEN

ICLR 2023 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives · 3 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add function-consistency matching to intermediate features (ablation)feature distillation · CIFAR100, ResNet32x4 → ShuffleV1top-1 accuracy: 76.77 -> 78.12Improved
add function-consistency matching to intermediate features (ablation)feature distillation · CIFAR100, ResNet32x4 → ShuffleV2top-1 accuracy: 76.49 -> 78.18Improved
add appearance matching alongside function matching (ablation)feature distillation · CIFAR100, ResNet32x4 → ShuffleV1top-1 accuracy: 76.73 -> 78.12Improved
add appearance matching alongside function matching (ablation)feature distillation · CIFAR100, ResNet32x4 → ShuffleV2top-1 accuracy: 77.35 -> 78.18Improved
add the student-later-part function loss L_func' (ablation)FCFD · CIFAR100, ResNet32x4 → ShuffleV1top-1 accuracy: 77.92 -> 78.12Improved
add the student-later-part function loss L_func' (ablation)FCFD · CIFAR100, ResNet32x4 → ShuffleV2top-1 accuracy: 78.02 -> 78.18Improved
add the teacher-later-part function loss L_func (ablation)FCFD · CIFAR100, ResNet32x4 → ShuffleV1top-1 accuracy: 77.19 -> 78.12Improved
add the teacher-later-part function loss L_func (ablation)FCFD · CIFAR100, ResNet32x4 → ResNet8x4top-1 accuracy: 76.62 -> 76.62No clear effect
add L_func'-L2 feature matching (ablation)FCFD function-loss design · CIFAR100, ResNet32x4 → ResNet8x4top-1 accuracy: 76.62 -> 76.57Hurt
add L_func'-L2 feature matching (ablation)FCFD function-loss design · CIFAR100, ResNet32x4 → ShuffleV1top-1 accuracy: 78.12 -> 77.82Hurt
add L_func'-L2 feature matching (ablation)FCFD function-loss design · CIFAR100, ResNet32x4 → ShuffleV2top-1 accuracy: 78.18 -> 78.14No clear effect
use FCFD rather than the strongest listed baselineResNet32x4 → ResNet8x4 distillation · CIFAR100top-1 accuracy: 76.32 -> 76.62Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy