Reproduce “Feature Superposition in Neural Networks: From Theory to Practice” 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

Dai Shi, Xiaoyu Li, Andi Han, José Miguel Hernández-Lobato

The authors survey overlapping feature representation and feature-recovery evidence. They define superposition through feature directions and decoder recovery, and distinguish six evidence levels, from representation to faithfulness. In a two-feature Bernoulli example, a ReLU decoder’s R² approaches 1 as activity becomes sparse, while the linear decoder remains at 1/2; this is not a guarantee for trained models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.