TY  - RPRT
TI  - A solvable high-dimensional model where nonlinear autoencoders learn structure invisible to PCA while test loss misaligns with generalization
AU  - Vicente Conde Mendes
AU  - Lorenzo Bardone
AU  - Cédric Koller
AU  - Jorge Medina Moreira
AU  - Vittorio Erba
AU  - Emanuele Troiani
AU  - Lenka Zdeborová
PY  - 2026
UR  - https://arxiv.org/abs/2602.10680
ID  - 2602.10680
ER  -