Reproduce “SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability” 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
Ali Nasiri-Sarvi, Hassan Rivaz, Mahdi S. Hosseini
SPARC trains a shared sparse feature space across CLIP’s image and text encoders and DINO’s image encoder, so corresponding latent indices can be compared across streams. Global TopK selects shared indices, while cross-reconstruction encourages information transfer between streams. On Open Images label-profile Jaccard, it scores 0.8018 ± 0.0039 versus USAE’s 0.2166 ± 0.0024; this is a label-based proxy, not proof of universally shared concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.