Reading list and daily arXiv feed

Awesome Superposition

Neural networks represent more features than they have dimensions. Here is the work that explains how they do it and how the features can be recovered, with the new papers added every day.

Curated worksLoading papers

Star on GitHub

Training

A toy autoencoder, trained live in your browser, stores 5 features in 2 dimensions. Each arrow is one feature's direction; darker arrows are more important features. Make the features sparser and watch them share the plane.
Papers per submission day

Most central this week

Selected works, ordered as in the survey: first the theory, then the practice. Each entry has one sentence on what the work shows. Suggestions are welcome as pull requests to config/reading_list.toml.