Toy model

The dataset used in the paper is a toy model, an MLP with a single linear hidden unit (1MLP) f (x) = w2w1x, which allows us to compare BP and PC exactly.

BibTex: