English

A jamming transition from under- to over-parametrization affects loss landscape and generalization

Machine Learning 2020-01-08 v5 Disordered Systems and Neural Networks Machine Learning

Abstract

We argue that in fully-connected networks a phase transition delimits the over- and under-parametrized regimes where fitting can or cannot be achieved. Under some general conditions, we show that this transition is sharp for the hinge loss. In the whole over-parametrized regime, poor minima of the loss are not encountered during training since the number of constraints to satisfy is too small to hamper minimization. Our findings support a link between this transition and the generalization properties of the network: as we increase the number of parameters of a given model, starting from an under-parametrized network, we observe that the generalization error displays three phases: (i) initial decay, (ii) increase until the transition point --- where it displays a cusp --- and (iii) slow decay toward a constant for the rest of the over-parametrized regime. Thereby we identify the region where the classical phenomenon of over-fitting takes place, and the region where the model keeps improving, in line with previous empirical observations for modern neural networks.

Keywords

Cite

@article{arxiv.1810.09665,
  title  = {A jamming transition from under- to over-parametrization affects loss landscape and generalization},
  author = {Stefano Spigler and Mario Geiger and Stéphane d'Ascoli and Levent Sagun and Giulio Biroli and Matthieu Wyart},
  journal= {arXiv preprint arXiv:1810.09665},
  year   = {2020}
}

Comments

arXiv admin note: text overlap with arXiv:1809.09349

R2 v1 2026-06-23T04:49:20.324Z