Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Cross-Level Distillation and Feature Denoising for Cross-Domain Few-Shot Classification
Read the paperCheck 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 cross-level distillation (CLD) to SimCLR (ablation) | SimCLR student · EuroSAT 5-way 1-shot | accuracy: +4.28 (75.10 to 79.38) | Helped |
| replace cross-level KD with same-level multi-block KD (ablation) | SimCLR student · EuroSAT 5-way 1-shot | accuracy: +0.38 over SimCLR only; CLD is +3.90 higher | Helped |
| add old student network to CLD (ablation) | CLD · EuroSAT 5-way 1-shot | accuracy: 76.85 without old student vs 79.38 with (tau=1) | Helped |
| add feature denoising (FD, h=64, beta=0.4) to SimCLR (ablation) | SimCLR · EuroSAT 5-way 1-shot | accuracy: +2.92 (75.10 to 78.02) | Helped |
| add feature denoising (FD) to SimCLR+CLD (ablation) | SimCLR+CLD · EuroSAT 5-way 1-shot | accuracy: +3.20 (79.38 to 82.52) | Helped |
| add feature denoising (FD) to BYOL+CLD (ablation) | BYOL+CLD · EuroSAT 5-way 1-shot | accuracy: +3.31 (69.47 to 72.78) | Helped |
| add feature denoising (FD) to Transfer+SimCLR (ablation) | Transfer+SimCLR · EuroSAT 5-way 1-shot | accuracy: -1.95 (65.92 to 63.97) | Hurt |
| add feature denoising (FD) to STARTUP (ablation) | STARTUP · EuroSAT 5-way 1-shot | accuracy: -4.37 (65.24 to 60.87) | Hurt |
| add cross-level distillation plus FD on SimCLR, versus Dynamic-Distillation baseline | SimCLR+CLD+FD · EuroSAT 5-way 1-shot | accuracy: +9.38 (73.14 to 82.52) | Helped |
| SimCLR+CLD+FD average across four BSCD-FSL datasets versus Dynamic-Distillation | SimCLR+CLD+FD · BSCD-FSL average (4 datasets) | 1-shot average accuracy: 58.77 vs 53.33 (+5.44) | Helped |
| use a self-supervised rather than supervised pre-trained teacher (ablation) | SimCLR+CLD+FD · EuroSAT 5-way 1-shot | accuracy: 80.30 (self-supervised teacher) vs 82.52 (supervised teacher) | Hurt |
Closest papers by meaning
- Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.90)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.88)
- Flatten Long-Range Loss Landscapes for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.88)
- Leveraging Normalization Layer in Adapters with Progressive Learning and Adaptive Distillation for Cross-Domain Few-Shot Learning (AAAI 2024 · similarity 0.87)
- The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.87)
- Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensator (ICLR 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.