Research map / Label-efficient and robust learning / Knowledge distillation
Mosaicking to Distill: Knowledge Distillation from Out-of-Domain Data
Gongfan Fang, Yifan Bao, Jie Song, Xinchao Wang, Donglin Xie, Chengchao Shen, Mingli Song
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 |
|---|---|---|---|
| replace direct OOD distillation with MosaicKD patch-based synthesis | CIFAR-100 students using CIFAR-10 OOD data · CIFAR-100 (CIFAR-10 OOD) | accuracy: 63.48 -> 69.55 | Improved |
| replace direct OOD distillation with MosaicKD patch-based synthesis | CIFAR-100 students using ImageNet OOD subset · CIFAR-100 (ImageNet OOD subset) | accuracy: 50.75 -> 67.60 | Improved |
| replace direct OOD distillation with MosaicKD patch-based synthesis | CIFAR-100 students using Places365 OOD subset · CIFAR-100 (Places365 OOD subset) | accuracy: 45.50 -> 66.23 | Improved |
| replace direct OOD distillation with MosaicKD patch-based synthesis | CIFAR-100 students using SVHN OOD data · CIFAR-100 (SVHN OOD) | accuracy: 27.81 -> 41.38 | Improved |
| replace KD on OOD images with MosaicKD | NYUv2 segmentation student · NYUv2 | mIoU: 0.406 -> 0.454 | Improved |
| replace direct OOD distillation with MosaicKD | CUB-200 student using Places365 · CUB-200 | accuracy: 18.12 -> 26.11 | Improved |
| replace direct OOD distillation with MosaicKD | Stanford Dogs student using Places365 · Stanford Dogs | accuracy: 19.13 -> 28.02 | Improved |
| replace KD on Places365 OOD data with MosaicKD | ImageNet student · ImageNet (Places365 OOD) | accuracy: 21.76 -> 26.51 | Improved |
| increase patch size from 8 to 18 (ablation) | MosaicKD · CIFAR-100 (CIFAR-10 OOD) | accuracy: 61.01 -> 61.03 | No clear effect |
| use patch size 8 instead of patch size 18 (ablation) | MosaicKD · CIFAR-100 (Places365 OOD) | accuracy: 53.99 -> 56.70 | Improved |
| use patch size 4 instead of patch size 8 (ablation) | MosaicKD · CIFAR-100 (SVHN OOD) | accuracy: 31.87 -> 32.09 | Improved |
| increase discriminator prediction stride from 1 to 2 (ablation) | MosaicKD · CIFAR-100 (CIFAR-10 OOD) | accuracy: 61.01 -> 59.56 | Hurt |
Closest papers by meaning
- Toward Effective Knowledge Distillation: Navigating Beyond Small-Data Pitfall (TPAMI 2025 · similarity 0.90)
- Small Scale Data-Free Knowledge Distillation (CVPR 2024 · similarity 0.90)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · similarity 0.89)
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 2025 · similarity 0.88)
- AuG-KD: Anchor-Based Mixup Generation for Out-of-Domain Knowledge Distillation (ICLR 2024 · similarity 0.87)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 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.
