advantage_nonlinearity

Repository for the paper "A solvable high-dimensional model where nonlinear autoencoders learn structure invisible to PCA while test loss misaligns with generalization"