Reproduce “Balancing Representations by Identifying and Generating Under-represented Data” 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
Filipe Laitenberger, Hirokatsu Kataoka, Yuki Asano, Cees Snoek, Mohammadreza Salehi
BRIDGE repairs unlabeled image-pretraining data by adding generated variants around images with sparse neighborhoods in its current representation. After each training cycle, it selects anchors using feature-space distances and applies a frozen image generator, without labels or changes to the self-supervised objective. On deliberately imbalanced ImageNet-100, SimCLR’s seven-task linear-probe average rose from 31.2 to 42.3; the main comparison uses more updates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.