Research map / Label-efficient and robust learning / Knowledge distillation
Data-to-Model Distillation: Data-Efficient Learning Framework
Ahmad Sajedi, Samir Khaki, Lucy Z. Liu, Ehsan Amjadian, Yuri A. Lawryshyn, Konstantinos N. Plataniotis
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 |
|---|---|---|---|
| distill into generator parameters with embedding matching loss (ablation) | Pre-trained generator (BigGAN) · CIFAR-10 IPC 10 | accuracy improvement: +10.3% | Improved |
| distill with prediction matching loss (ablation) | Pre-trained generator (BigGAN) · CIFAR-10 IPC 10 | accuracy improvement: +7.9% | Improved |
| combine embedding matching and prediction matching losses (ablation) | Pre-trained generator (BigGAN) · CIFAR-10 IPC 10 | accuracy improvement: +12.7% | Improved |
| distill into generator parameters instead of pixels so no re-distillation across IPC | D2M · CIFAR-10 distillation cost (IPC1, 10, 50) | total GPU hours: 47.4 -> 4.9 | Improved |
| raise softmax temperature T up to 4 (ablation) | Prediction matching · CIFAR-10 IPC 10 | accuracy | Improved |
| set task balance lambda too high (1000) or too low (ablation) | D2M loss · CIFAR-10 IPC 10 | accuracy | Hurt |
| swap BigGAN for StyleGAN-XL as generator (ablation) | D2M · CIFAR-10 IPC 50 | accuracy (%): 74.4 -> 75.2 | Improved |
| swap BigGAN for CGAN as generator (ablation) | D2M · CIFAR-10 IPC 50 | accuracy (%): 74.4 -> 72.2 | Hurt |
| use D2M images on unseen architectures | D2M · CIFAR-10 IPC 50 cross-architecture | average accuracy (%): 59.7 -> 63.6 | Improved |
| use D2M-generated proxy set for architecture search vs Random | D2M · CIFAR-10 NAS | best-model test accuracy (%): 88.9 -> 88.9 | No clear effect |
| D2M for IPC 1 uses more GPU memory than pixel distillation | D2M · CIFAR-10 IPC 1 distillation | GPU hours: 1.4 -> 4.9 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TESLA — TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation (CVPR 2023) · better on 4 of 4 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
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.91)
- Data Distillation Can Be Like Vodka: Distilling More Times For Better Quality (ICLR 2024 · similarity 0.90)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.90)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.90)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.90)
- D^4: Dataset Distillation via Disentangled Diffusion Model (CVPR 2024 · similarity 0.90)
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.
