Reproduce “Distilling Datasets Into Less Than One Image” 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
Asaf Shul, Eliahu Horwitz, Yedid Hoshen
Researchers at The Hebrew University of Jerusalem developed Poster Dataset Distillation (PoDD), a method that distills entire datasets into a single 'poster' image, enabling extreme compression by using overlapping patches and shared pixels. PoDD achieves meaningful classification performance with less than one image per class, outperforming previous methods and setting new benchmarks at higher compression rates.
Key insights
- Pixel sharing among multiple classes within a single 'poster' image significantly reduces the total pixel budget required for dataset representation, enabling compression ratios below 1 image per class.
- Semantically similar classes should be placed adjacently in the poster image, as this spatial arrangement improves the quality of shared pixel information and overall training performance.
- Using learnable soft labels for overlapping patches, rather than hard labels for discrete images, is crucial for effectively representing multi-class information within shared regions of the poster.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.