Dataset Distillation via Drifting squeezes ImageNet classes into 10 images (NeurIPS)
giannis_daras · x · 2026-11-14
In a NeurIPS 2026 paper, "Dataset Distillation via Drifting," the authors distill each ImageNet class (1,300 images) down to just 10 synthetic images optimized from noise, enough to teach a model what a class like a sea anemone looks like. More details in the linked thread and project page.
Related event: Dataset Distillation Hits SOTA: 10 Images Per Class(2 posts)→
More from Research
- Claude cracks Gödel's 1940-42 notebooks in a day, surfacing 663 unknown theorems — stevenstrogatz · 2026-11-14
- Scientists simulate an entire living cell in 4D, published in Cell — tomchapin · 2026-11-14
- Skeptic doubts Anthropic achieved large-factor bpb gains at fixed compute — teortaxesTex · 2026-11-14
- Quantizing training to MXFP8 with stochastic rounding: the accuracy-speed sweet spot — snowclipsed · 2026-11-14
- Ant Group's Mara Chain: refine prompts/harness/code, matches GEPA with far fewer rollouts — jiqizhixin · 2026-11-14
- Notion Ran ZeroEntropy's Reranker in Production Long Before the Acquisition — CShorten30 · 2026-11-14