Reproduce “Forward Compatible Training for Large-Scale Embedding Retrieval Systems” 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
Vivek Ramanujan, Pavan Kumar Anasosalu Vasu, Ali Farhadi, Oncel Tuzel, Hadi Pouransari
The paper introduces Forward-Compatible Training (FCT) for image retrieval, storing side-information with each old gallery embedding and learning a transformation that combines them into a feature compatible with an independently trained new model. On ImageNet-1k, FCT reaches 65.0% CMC top-1 versus 46.9% for BCT, but requires extra storage and computation and is not established across arbitrary domains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.