Research map / Label-efficient and robust learning / Knowledge distillation
One-for-All: Bridge the Gap Between Heterogeneous Architectures in Knowledge Distillation
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 exit branches at all four stages (vs last layer only) (ablation) | OFA-KD, DeiT-T teacher, ResNet18 student · ImageNet-1K | top-1 accuracy (%): 70.62 -> 71.34 | Helped |
| add exit branches at all four stages (vs last layer only) (ablation) | OFA-KD, ResNet50 teacher, DeiT-T student · ImageNet-1K | top-1 accuracy (%): 75.18 -> 75.73 | Helped |
| add single exit branch at stage 2 or 3 (ablation) | OFA-KD, ResNet50 teacher, DeiT-T student · ImageNet-1K | top-1 accuracy (%): 75.18 -> 75.21 / 75.17 | No clear effect |
| use more branches {1,2,4} vs {1,4} (ablation) | OFA-KD, DeiT-T teacher, ResNet18 student · ImageNet-1K | top-1 accuracy (%): 71.04 -> 71.08 | Helped |
| scale OFA loss by 0.5 instead of 1.0 (ablation) | OFA-KD · ImageNet-1K ResNet50->DeiT-T | top-1 accuracy (%): 75.73 -> 75.29 | Hurt |
| change gradient clipping max norm from 5 to 7 (ablation) | OFA-KD · ImageNet-1K ResNet50->DeiT-T | top-1 accuracy (%): 75.73 -> 75.36 | Hurt |
| use OFA modulating parameter gamma tuned (1.4 for DeiT-T teacher, 1.1 for ResNet50) (ablation) | OFA loss · ImageNet-1K | top-1 accuracy (%) | Helped |
| replace FitNet feature matching with OFA-KD for ConvNeXt-T teacher, ResMLP-S12 student | hint-based KD · CIFAR-100 | top-1 accuracy (%): 45.47 -> 81.22 | Helped |
| OFA-KD vs second-best baseline (DIST) for heterogeneous teachers | KD baselines · ImageNet-1K | top-1 accuracy (%): 70.64 -> 71.34 (DeiT-T->ResNet18) | Helped |
| OFA-KD vs DIST on homogeneous ResNet34->ResNet18 | KD baselines · ImageNet-1K | top-1 accuracy (%): 72.07 -> 72.10 | No clear effect |
Closest papers by meaning
- Distilling Knowledge from Heterogeneous Architectures for Semantic Segmentation (AAAI 2025 · similarity 0.90)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.90)
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · similarity 0.89)
- Over-parameterized Student Model via Tensor Decomposition Boosted Knowledge Distillation (NeurIPS 2024 · similarity 0.89)
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.89)
- CIFD: Controlled Information Flow to Enhance Knowledge Distillation (NeurIPS 2024 · similarity 0.89)
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.