Research map / Label-efficient and robust learning / Knowledge distillation
Correlation Congruence for Knowledge Distillation
Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, Zhaoning Zhang
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 correlation congruence to instance-level KD (ablation) | KD · CIFAR-100 (ResNet-20) | top-1 accuracy: 70.8 -> 72.4 | Improved |
| add correlation congruence to instance-level KD (ablation) | KD · CIFAR-100 (ResNet-14) | top-1 accuracy: 68.3 -> 70.2 | Improved |
| add correlation congruence to instance-level KD (ablation) | KD · ImageNet-1K (MobileNetV2 0.5x) | top-1 accuracy: 66.7 -> 67.7 | Improved |
| add correlation congruence to instance-level KD (ablation) | KD · MSMT17 (ResNet-18) | rank-1 / mAP: 56.8 -> 59.7 rank-1; 28.3 -> 30.7 mAP | Improved |
| add correlation congruence to L2-mimic loss (ablation) | L2 mimic · MegaFace (1M distractors) | rank-1 identification rate: 83.01 -> 86.29 | Improved |
| use Gaussian RBF instead of MMD as the correlation metric (ablation) | CCKD · MSMT17 | mAP: 29.4 -> 30.4 | Improved |
| use Gaussian RBF instead of Bilinear Pool as the correlation metric (ablation) | CCKD · MSMT17 | mAP: 30.2 -> 30.4 | Improved |
| raise Taylor expansion order of Gaussian RBF from 1 to 3 (ablation) | CCKD · MSMT17 | rank-1 / mAP: 59.2 -> 60.5 rank-1; 30.1 -> 30.7 mAP | Improved |
| use CUR-sampler (k=4) instead of uniform random sampler (ablation) | CCKD · MSMT17 | rank-1 / mAP: 57.2 -> 59.7 rank-1; 28.6 -> 30.2 mAP | Improved |
| use SUR-sampler (k=4) instead of CUR-sampler (k=4) (ablation) | CCKD · MSMT17 | mAP: 30.2 -> 31.1 | Improved |
| increase images per class k from 4 to 20 in CUR-sampler (ablation) | CCKD · MSMT17 | mAP: 30.2 -> 10.7 | Hurt |
| increase images per class k from 4 to 20 in SUR-sampler (ablation) | CCKD · MSMT17 | mAP: 31.1 -> 13.7 | Hurt |
Closest papers by meaning
- Local Correlation Consistency for Knowledge Distillation (ECCV 2020 · similarity 0.90)
- Student Customized Knowledge Distillation: Bridging the Gap Between Student and Teacher (ICCV 2021 · similarity 0.89)
- Distribution Shift Matters for Knowledge Distillation with Webly Collected Images (ICCV 2023 · similarity 0.88)
- Scaled Decoupled Distillation (CVPR 2024 · similarity 0.87)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · similarity 0.86)
- Cooperative Knowledge Distillation: A Learner Agnostic Approach (AAAI 2024 · similarity 0.86)
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.
