Generalization performance of narrow one-hidden layer networks in the teacher-student setting

Abstract

Understanding the generalization abilities of neural networks for simple input-output distributions is crucial to account for their learning performance on real datasets. The classical teacher-student setting, where a network is trained from data obtained thanks to a label-generating teacher model, serves as a perfect theoretical test bed. In this context, a complete theoretical account of the performance of fully connected one-hidden layer networks in the presence of generic activation functions is lacking. In this work, we develop such a general theory for narrow networks, i.e. with a large number of hidden units, yet much smaller than the input dimension. Using methods from statistical physics, we provide closed-form expressions for the typical performance of both finite temperature (Bayesian) and empirical risk minimization estimators, in terms of a small number of summary statistics. In doing so, we highlight the presence of a transition where hidden neurons specialize when the number of samples is sufficiently large and proportional to the number of parameters of the network. Our theory accurately predicts the generalization error of neural networks trained on regression or classification tasks with either noisy full-batch gradient descent (Langevin dynamics) or full-batch gradient descent.