Research map / Label-efficient and robust learning / Knowledge distillation
Exploring Inter-Channel Correlation for Diversity-Preserved Knowledge Distillation
Li Liu, Qingle Huang, Sihao Lin, Hongwei Xie, Bing Wang, Xiaojun Chang, Xiaodan Liang
Read the paperCheck your idea against 137,217 papers
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add 1x1 linear transformation layer on student feature (ablation) | ICKD-C · CIFAR-100 test (ResNet110 to ResNet20) | top-1 accuracy: 70.96 -> 71.68 | Improved |
| add 1x1 linear transformation layer on student feature (ablation) | ICKD-C · CIFAR-100 test (VGG13 to VGG8) | top-1 accuracy: 73.85 -> 73.42 | Hurt |
| add 1x1 linear transformation layer on student feature (ablation) | ICKD-C · CIFAR-100 test (ResNet110 to ResNet32) | top-1 accuracy: 73.90 -> 73.89 | No clear effect |
| add 1x1 linear transformation layer on student feature (ablation) | ICKD-C · CIFAR-100 test (ResNet32x4 to ResNet8x4) | top-1 accuracy: 74.40 -> 75.25 | Improved |
| add KL loss to ICC loss (ablation) | ICKD-C · CIFAR-100 test (WRN-40-2 to WRN-40-1) | top-1 accuracy: 74.33 -> 74.63 | Improved |
| add KL loss to ICC loss (ablation) | ICKD-C · CIFAR-100 test (WRN-40-2 to WRN-16-2) | top-1 accuracy: 75.64 -> 75.57 | No clear effect |
| distill ICC at stage 4 only versus stage 1 only (ablation) | ICKD-C · ImageNet (ResNet34 to ResNet18) | top-1 accuracy: 70.49 -> 72.16 | Improved |
| add ICC transfer at stages 3 and 4 instead of stage 4 only (ablation) | ICKD-C · ImageNet (ResNet34 to ResNet18) | top-1 accuracy: 72.16 -> 72.33 | Improved |
| use Gaussian kernel instead of the inner-product ICC (L2) form (ablation) | ICKD-C · ImageNet (ResNet34 to ResNet18) | top-1 accuracy: 72.16 -> 70.63 | Hurt |
| use smooth L1 loss instead of L2 (ablation) | ICKD-C · ImageNet (ResNet34 to ResNet18) | top-1 accuracy: 72.16 -> 72.29 | No clear effect |
| partition the feature map into 32x32 grid for ICC (ablation) | ICKD-S · Pascal VOC val (ResNet18 student) | mIoU: 74.14 -> 75.01 | Improved |
| partition the feature map into 16x16 grid instead of 32x32 (ablation) | ICKD-S · Pascal VOC val (MobileNetV2 student) | mIoU: 72.58 -> 72.79 | Improved |
Closest papers by meaning
- Channel-Wise Knowledge Distillation for Dense Prediction (ICCV 2021 · similarity 0.90)
- Heterogeneous Complementary Distillation (AAAI 2026 · similarity 0.90)
- You Don’t Need Aligned Representations: Knowledge Distillation via Random Prototype Spaces (NeurIPS 2026 · similarity 0.90)
- Wasserstein Distance Rivals Kullback-Leibler Divergence for Knowledge Distillation (NeurIPS 2024 · similarity 0.89)
- Debiased Distillation for Consistency Regularization (AAAI 2025 · similarity 0.88)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.88)
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.
