Reproduce “Representation Learning via Composite Subspace Flow Matching” 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
José Afonso, Mário Figueiredo
CSFM separates image information into content and style components and trains a generative flow with different degradation schedules, including a cutoff for style information. This is intended to preserve features useful for recognition while retaining a decoding path. On CIFAR-10, truncated CSFM reaches 75.29% linear-probe accuracy versus 67.11% for OT-Flow, but LeJEPA leads on 1%-label accuracy and k-NN.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.