Research map / Detection and segmentation / CNN features
Refine Myself by Teaching Myself: Feature Refinement via Self-Knowledge Distillation
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 self-teacher soft-label distillation (ablation) | FRSKD classifier · CIFAR-100, WRN-16-2 | accuracy: 70.42$\pm$ -> 73.12$\pm$ | Helped |
| add feature-map distillation to self-teacher soft-label distillation (ablation) | FRSKD · CIFAR-100, WRN-16-2 | accuracy: 73.12$\pm$ -> 73.27$\pm$ | Helped |
| combine FRSKD with SLA-SD (ablation) | FRSKD · CIFAR-100, ResNet18 | accuracy: 77.71$\pm$ -> 82.04$\pm$ | Helped |
| add self-teacher soft-label distillation (ablation) | FRSKD classifier · CUB200 | accuracy: 51.72$\pm$ -> 62.29$\pm$ | Helped |
| add feature-map distillation to self-teacher soft-label distillation (ablation) | FRSKD · CUB200 | accuracy: 62.29$\pm$ -> 65.39$\pm$ | Helped |
| combine FRSKD with SLA-SD (ablation) | FRSKD · CUB200 | accuracy: 65.39$\pm$ -> 67.80$\pm$ | Helped |
| use attention transfer rather than FitNet feature distillation (ablation) | FRSKD · CUB200 | accuracy: 58.97$\pm$ -> 65.39$\pm$ | Helped |
| use attention transfer rather than overhaul feature distillation (ablation) | FRSKD · CUB200 | accuracy: 58.82$\pm$ -> 65.39$\pm$ | Helped |
| use depth-dependent channels (BiFPNc) rather than fixed-width BiFPN at 256 channels (ablation) | self-teacher network · CIFAR-100 | FLOPs: $\times$ 2.38 -> $\times$ 0.68 | Helped |
| use depth-dependent channels (BiFPNc) rather than fixed-width BiFPN at 256 channels (ablation) | self-teacher network · CIFAR-100 | accuracy: 73.54$\pm$ -> 73.27$\pm$ | Hurt |
| add Mixup alongside FRSKD (ablation) | FRSKD · CUB200 | accuracy: 65.39$\pm$ -> 67.98$\pm$ | Helped |
| add CutMix alongside FRSKD (ablation) | FRSKD · MIT67 | accuracy: 61.74$\pm$ -> 66.19$\pm$ | Helped |
Closest papers by meaning
- Self-Decoupling and Ensemble Distillation for Efficient Segmentation (AAAI 2023 · similarity 0.91)
- SDPGO: Efficient Self-Distillation Training Meets Proximal Gradient Optimization (NeurIPS 2025 · similarity 0.90)
- Knowledge Distillation from A Stronger Teacher (NeurIPS 2022 · similarity 0.89)
- Distilling Knowledge from Heterogeneous Architectures for Semantic Segmentation (AAAI 2025 · similarity 0.89)
- Structured Knowledge Distillation for Semantic Segmentation (CVPR 2019 · similarity 0.88)
- Multi-label Self Knowledge Distillation (AAAI 2025 · similarity 0.87)
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.