Research map / Learning theory and algorithms / Kernels and regression theory
Efficient Dataset Distillation using Random Feature Approximation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Approximate the NNGP kernel with random neural-network features instead of computing the exact kernel | RFAD versus KIP · CIFAR-10 training iterations | speedup: over two-orders-of-magnitude | Helped |
| Use RFAD rather than KIP for fixed-label distillation | CIFAR-10, 10 images/class · CIFAR-10, 10 images/class | accuracy: 67.0 -> 73.1 | Helped |
| Use RFAD rather than KIP with learned labels | MNIST, 1 image/class · MNIST, 1 image/class | accuracy: 97.3\pm0.1 -> 97.2\pm0.2 | No clear effect |
| Use RFAD rather than KIP with learned labels | Fashion-MNIST, 1 image/class · Fashion-MNIST, 1 image/class | accuracy: 82.9\pm0.2 -> 84.6\pm0.2 | Helped |
| Use RFAD rather than KIP with learned labels | SVHN, 1 image/class · SVHN, 1 image/class | accuracy: 64.3\pm0.4 -> 57.4\pm.8 | Hurt |
| Use RFAD rather than KIP with learned labels | CIFAR-100, 1 image/class · CIFAR-100, 1 image/class | accuracy: 34.9\pm0.1 -> 44.1\pm0.1 | Helped |
| Replace MSE with cross-entropy loss and jointly learn temperature scaling (ablation) | RFAD · Fashion-MNIST, 1 image/class | accuracy: +8\% | Helped |
| Remove label scaling for finite-network transfer (ablation) | RFAD distilled set · CIFAR-10, 1 image/class | accuracy: 48.7 \pm 1.6 -> 53.2 \pm 1.2 | Hurt |
| Remove centering for finite-network transfer (ablation) | RFAD distilled set · CIFAR-10, 1 image/class | accuracy: 48.0 \pm 1.7 -> 53.2 \pm 1.2 | Hurt |
| Remove label scaling for finite-network transfer (ablation) | RFAD distilled set · Fashion-MNIST, 1 image/class | accuracy: 51.0 \pm 19.2 -> 76.7 \pm 1.7 | Hurt |
| Remove centering for finite-network transfer (ablation) | RFAD distilled set · Fashion-MNIST, 1 image/class | accuracy: 69.8 \pm 3.4 -> 76.7 \pm 1.7 | Hurt |
| Evaluate distilled sets on networks with InstanceNorm when trained without it (ablation) | RFAD · CIFAR-10, 1 image/class | accuracy: 61.1 \pm 0.7 -> 35.3 \pm 0.9 | Hurt |
Closest papers by meaning
- Provable and Efficient Dataset Distillation for Kernel Ridge Regression (NeurIPS 2024 · similarity 0.89)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.88)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.86)
- Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets? (NeurIPS 2026 · similarity 0.85)
- Data Distillation Can Be Like Vodka: Distilling More Times For Better Quality (ICLR 2024 · similarity 0.85)
- Embarrassingly Simple Dataset Distillation (ICLR 2024 · 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.