Research map / Label-efficient and robust learning / Knowledge distillation
IDeaL: Data-Free Multi-Teacher Distillation via Improved Dead Leaves
Feyza Yavuz, Mert Bulent SARIYILDIZ, Diane Larlus
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 |
|---|---|---|---|
| optimize Dead Leaves pixels with decorrelation losses (IDeaL) | Dead Leaves distillation data · ImageNet val (1M samples) | top-1 accuracy: 66.9 -> 78.2 | Improved |
| optimize Dead Leaves pixels with decorrelation losses (IDeaL) | Dead Leaves distillation data · ImageNet val (1K samples) | top-1 accuracy: 64.7 -> 74.1 | Improved |
| optimize Dead Leaves pixels with decorrelation losses (IDeaL) | Dead Leaves distillation data · ADE20K (1M samples) | mIoU: 29.1 -> 34.5 | Improved |
| optimize Dead Leaves pixels with decorrelation losses (IDeaL) | Dead Leaves distillation data · NYUd (1M samples) | RMSE: 0.632 -> 0.573 | Improved |
| optimize Dead Leaves pixels with decorrelation losses (IDeaL) | Dead Leaves distillation data · Transfer (15 tasks, 1M samples) | top-1 accuracy: 65.5 -> 70.6 | Improved |
| use 1K IDeaL samples instead of 1K ImageNet images | multi-teacher distillation, 1K samples · Transfer (15 tasks) | top-1 accuracy: 65.3 -> 68.6 | Improved |
| use 1M IDeaL samples instead of 1M ImageNet images | multi-teacher distillation, 1M samples · ImageNet val | top-1 accuracy: 83.2 -> 78.2 | Hurt |
| replace Dead Leaves with Gaussian noise as initialization | distillation data · ImageNet val (1M samples) | top-1 accuracy: 66.9 -> 22.6 | Hurt |
| replace Dead Leaves with Fractals | distillation data · ImageNet val (1M samples) | top-1 accuracy: 66.9 -> 32.7 | Hurt |
| increase Dead Leaves size from 1K to 1M | Dead Leaves distillation data · ImageNet val | top-1 accuracy: 64.7 -> 66.9 | No clear effect |
| optimize samples with all 4 teachers jointly instead of 1 teacher (DINO) (ablation) | IDeaL · ImageNet val (10K samples) | top-1 accuracy: 74.4 -> 77.0 | Improved |
| optimize jointly across 4 teachers instead of combining per-teacher sets (ablation) | IDeaL 10K samples · ImageNet val | top-1 accuracy: 75.1 -> 77.0 | Improved |
Closest papers by meaning
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.88)
- Scaling Multi-Teacher Distillation for Digital Pathology (NeurIPS 2026 · similarity 0.88)
- Distillation Scaling Laws (ICML 2025 · similarity 0.88)
- Your Student is Better Than Expected: Adaptive Teacher-Student Collaboration for Text-Conditional Diffusion Models (CVPR 2024 · similarity 0.88)
- How to Trade Off the Quantity and Capacity of Teacher Ensemble: Learning Categorical Distribution to Stochastically Employ a Teacher for Distillation (AAAI 2024 · similarity 0.87)
- Hybrid Data-Free 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.
