Bilinear MLPs enable weight-based mechanistic interpretability.

Michael T. Pearce, Thomas Dooms, Alice Rigg, Jos Oramas, Lee Sharkey

Browse the full ICLR paper archive.