Research map / Label-efficient and robust learning / Knowledge distillation
IDEAL: Query-Efficient Data-Free Learning from Black-Box Models
Jie Zhang, Chen Chen, Lingjuan Lyu
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 |
|---|---|---|---|
| Use the student model, rather than the teacher, to train the generator without querying the teacher | IDEAL · CIFAR10 | top-1 test accuracy (%): 68.82 vs 37.91 | Improved |
| Use the student model, rather than the teacher, to train the generator without querying the teacher | IDEAL · CIFAR100 | top-1 test accuracy (%): 36.96 vs 20.01 | Improved |
| Add the information loss L_infor to generator training (ablation) | IDEAL · CIFAR10 | top-1 test accuracy (%): 62.68 -> 68.82 | Improved |
| Add the information loss L_infor to generator training (ablation) | IDEAL · SVHN | top-1 test accuracy (%): 83.21 -> 86.91 | Improved |
| Add the cross-entropy loss L_ce to generator training (ablation) | IDEAL · CIFAR10 | top-1 test accuracy (%): 10.68 -> 68.82 | Improved |
| Add the cross-entropy loss L_ce to generator training (ablation) | IDEAL · SVHN | top-1 test accuracy (%): 12.48 -> 86.91 | Improved |
| Reinitialize the generator at each epoch (ablation) | IDEAL · CIFAR10 | top-1 test accuracy (%): 55.32 -> 68.82 | Improved |
| Reinitialize the generator at each epoch (ablation) | IDEAL · SVHN | top-1 test accuracy (%): 80.53 -> 86.91 | Improved |
| Set generator-training iterations E_G to 3 instead of the reported best value 5 (ablation) | IDEAL · MNIST | top-1 test accuracy (%): 65.36 vs 96.51 | Hurt |
| Set generator-training iterations E_G to 50 instead of the reported best value 10 (ablation) | IDEAL · CIFAR10 | top-1 test accuracy (%): 68.82 -> 58.69 | Hurt |
| Use larger generator-training iteration counts (ablation) | IDEAL · CIFAR10 | top-1 test accuracy (%): 30.25, 57.56, 68.82, 62.45, 58.69 | Mixed |
| Use generator re-initialization and balanced per-class sample generation rather than class-imbalance reweighting (ablation) | IDEAL · CIFAR10 | top-1 test accuracy (%): 62.14, 63.31 -> 68.82 | Improved |
Later papers that compared with this method
2 later accepted papers report IDEAL as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Unsupervised Process-Aware Coreset Selection for In-Context Learning (ICML 2026 · better on 4 of 4 shared results)
- Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks (AAAI 2024 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Teacher as a Lenient Expert: Teacher-Agnostic Data-Free Knowledge Distillation (AAAI 2024 · similarity 0.88)
- Small Scale Data-Free Knowledge Distillation (CVPR 2024 · similarity 0.87)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.86)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.86)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.86)
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 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.
