Reproduce “Learning a Unified Cross-Model Semantic Dictionary via Gated Bottleneck Sparse Autoencoders” 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
Jinchi Zhu, Tie Luo
GB-SAE learns a shared concept dictionary across vision models, while separate private dictionaries preserve model-specific residuals. It uses a shared encoder and gated shared and private decoders to reconstruct model features. On ImageNet-1K with ViT, SigLIP, and DINOv2, shared mean R² is 0.54 versus SPARC’s 0.41, though these metrics are not exactly comparable; retrieval also improves within tested vision settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.