Picture for Vicente Conde Mendes

Vicente Conde Mendes

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

Add code
Feb 11, 2026
Viaarxiv icon