Reproduce “Complete or Sparse: A Tale of Two Identifiabilities” 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
Junze Zhou, David Klindt
The authors analyze when supervised learning produces linear or feature-aligned representations. They show sufficient activation width makes the classifier-relevant code linear in latent features, while axis alignment requires assumptions about data generation and constraints. In synthetic experiments, nonnegative outputs align axes only with sparse data, and supervised sparse decoding recovers compressed latents better than ordinary least squares; results are theoretical or synthetic.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.