Reproduce “Dataset Distillation via Drifting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
George Cazenavette, Angelina Quan, Giannis Daras, Antonio Torralba, Vincent Sitzmann
The authors introduce dataset distillation that optimizes class distributions in frozen feature spaces rather than classifier losses. Synthetic images are pulled toward real examples and separated from one another across augmented views. With ten noise-initialized images per ImageNet-1k class, a ResNet-18 reaches 27.0 ± 0.2% top-1 accuracy, versus 11.7 ± 0.3% for the strongest noise-initialized baseline; the results do not establish lower overall compute.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.