Research map / Label-efficient and robust learning / Knowledge distillation
Mining Data Impressions from Deep Models as Substitute for the Unavailable Training Data
Gaurav Kumar Nayak, Konda Reddy Mopuri, Saksham Jain, Anirban Chakraborty
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 Dirichlet-modelled Data Impressions instead of data-free distillation from few original samples | ZSKD, LeNet-5-Half student · MNIST | accuracy: 86.70 -> 98.77 | Improved |
| use Dirichlet-modelled Data Impressions instead of 200 original samples | ZSKD, LeNet-5-Half student · Fashion-MNIST | accuracy: 72.50 -> 79.62 | Improved |
| replace original training data by Data Impressions for distillation | AlexNet-Half student · CIFAR-10 (AlexNet) | accuracy: 80.08 -> 69.56 | Hurt |
| replace original training data by Data Impressions for distillation to a smaller ResNet | Resnet-18-half student · CIFAR-10 (Resnet-18 teacher) | accuracy: 86.31 -> 81.10 | Hurt |
| replace original training data by Data Impressions for VGG-11 student | VGG-11 student · CIFAR-10 (VGG-19 teacher) | accuracy: 84.93 -> 74.10 | Hurt |
| use Dirichlet-modelled Data Impressions instead of one-hot Class Impressions (ablation) | ZSKD · MNIST, Fashion-MNIST, CIFAR-10 | accuracy | Improved |
| increase the number of Data Impressions in the transfer set (ablation) | ZSKD · MNIST, Fashion-MNIST, CIFAR-10 | accuracy | Improved |
| distill from an adversarially trained teacher instead of a normal teacher using DIs (PGD attack) (ablation) | ZSKD student · MNIST PGD attack | adversarial accuracy: 3.06 -> 54.38 | Improved |
| distill from an adversarially trained teacher instead of a normal teacher using DIs (PGD attack) (ablation) | ZSKD student · CIFAR-10 PGD attack | adversarial accuracy: 15.24 -> 24.92 | Improved |
| distill from a robust teacher using DIs, natural accuracy cost (ablation) | ZSKD student · CIFAR-10 natural | accuracy: 69.56 -> 54.38 | Hurt |
| use a lower beta mixture {0.01, 0.1} instead of {0.1, 1.0} for DI generation (ablation) | Source-free domain adaptation, SVHN to MNIST · SVHN to MNIST domain adaptation | accuracy: 0.690 -> 0.866 | Improved |
| use a lower beta mixture {0.01, 0.1} instead of {0.1, 1.0} for DI generation (ablation) | Source-free domain adaptation, USPS to MNIST · USPS to MNIST domain adaptation | accuracy: 0.7505 -> 0.8915 | Improved |
Closest papers by meaning
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.86)
- Small Scale Data-Free Knowledge Distillation (CVPR 2024 · similarity 0.85)
- Impartial Adversarial Distillation: Addressing Biased Data-Free Knowledge Distillation via Adaptive Constrained Optimization (AAAI 2024 · similarity 0.85)
- A Bitter Lesson for Data Filtering (NeurIPS 2026 · similarity 0.85)
- Slight Corruption in Pre-training Data Makes Better Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Towards Adversarially Robust Dataset Distillation by Curvature Regularization (AAAI 2025 · similarity 0.84)
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.
