Research map / Label-efficient and robust learning / Knowledge distillation
Aligning Logits Generatively for Principled Black-Box Knowledge Distillation in the Wild
Xiang Xiang, Jing Ma, Dongrui Wu, Zhigang Zeng, Xilin Chen
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 information maximization loss (alpha 0.0 -> 0.5), soft responses (ablation) | MEKD · CIFAR-100 (ResNet56->MobileNet) | top-1 accuracy (%): 60.87 -> 67.07 | Improved |
| add information maximization loss (alpha 0.0 -> 0.5), hard responses (ablation) | MEKD · CIFAR-100 (ResNet56->MobileNet) | top-1 accuracy (%): 61.31 -> 67.36 | Improved |
| add KL logit-level loss (beta 0.0 -> 1.0), soft responses (ablation) | MEKD · CIFAR-100 (ResNet56->MobileNet) | top-1 accuracy (%): 56.28 -> 67.07 | Improved |
| add KL logit-level loss (beta 0.0 -> 1.0), hard responses (ablation) | MEKD · CIFAR-100 (ResNet56->MobileNet) | top-1 accuracy (%): 56.23 -> 67.36 | Improved |
| use L1 instead of L2 norm for image-level loss (ablation) | MEKD (soft) · CIFAR-10 (ResNet56->MobileNet) | top-1 accuracy (%): 86.85 / 86.63 | No clear effect |
| use L1 instead of L2 norm for image-level loss (ablation) | MEKD (soft) · CIFAR-100 (ResNet56->MobileNet) | top-1 accuracy (%): 67.07 / 66.95 | No clear effect |
| use hard instead of soft teacher responses | MEKD · Table 1 benchmarks | top-1 accuracy (%): difference of less than 1% | No clear effect |
| increase query data size from 0.1K to 50K (full) on CIFAR-10 (ablation) | MEKD (soft) · CIFAR-10 (ResNet56->MobileNet) | top-1 accuracy (%): 49.04 -> 93.48 | Improved |
| use MEKD instead of DB3KD with only 0.1K samples | MEKD (soft) · CIFAR-10, data size 0.1K | top-1 accuracy (%): 43.05 -> 49.04 | Improved |
| use MEKD instead of DB3KD on ImageNet-1K (RN50 to RN34) | MEKD (soft) · ImageNet-1K | top-1 accuracy (%): 58.61 -> 59.89 | Improved |
| use MEKD instead of DB3KD on out-of-domain SVHN (ResNet56 teacher) | MEKD (soft) · SVHN (teacher trained on Syn. Digits) | top-1 accuracy (%): 90.15 -> 86.45 | Hurt |
| add information maximization loss without KL loss (ablation) | MEKD · CIFAR-10/100 limited queries (10K to 50K) | top-1 accuracy (%) | No clear effect |
Closest papers by meaning
- Aligning Logits Generatively for Principled Black-Box Knowledge Distillation (CVPR 2024 · similarity 0.99)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.86)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.86)
- Gap Preserving Distillation by Building Bidirectional Mappings with A Dynamic Teacher (ICLR 2025 · similarity 0.86)
- Small Scale Data-Free Knowledge Distillation (CVPR 2024 · similarity 0.85)
- CustomKD: Customizing Large Vision Foundation for Edge Model Improvement via Knowledge Distillation (CVPR 2025 · similarity 0.85)
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.
