Reproduce “SAS: Semantic-aware Sampling for Generative Dataset Distillation” 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

Mingzhuo Li, Guang Li, Linfeng Ye, Jiafeng Mao, Takahiro Ogawa, Konstantinos N. Plataniotis, Miki Haseyama

SAS is a post-sampling selection step for generative dataset distillation. It uses CLIP’s shared image-text space to filter for class relevance and separation, then selects examples sequentially using updated within-class diversity scores. On ImageWoof, with 10 images per class, SAS raises ResNetAP-10’s validation accuracy on a Minimax-generated pool from 35.7 ± 0.3% to 40.3 ± 0.5%; this comparison is not evidence of universal gains.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.