Reproduce “FairDeDup: Detecting and Mitigating Vision-Language Fairness Disparities in Semantic Dataset Deduplication” 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
Eric Slyman, Stefan Lee, Scott Cohen, Kushal Kafle
FairDeDup changes semantic deduplication by selecting, within image clusters, samples that represent sensitive concepts less represented among retained images. It uses text prototypes to guide selection rather than keeping images farthest from cluster centers. For CLIP-style models trained on LAION-400M, pruning 50% preserved ImageNet top-5 accuracy at 89.7%; FACET worst-class gender disparity fell versus SemDeDup, but not every measure improved.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.