Scaling description of generalization with number of parameters in deep learning

Supervised deep learning involves the training of neural networks with a large number N of parameters. For large enough N, in the so-called over-parametrized regime, one can essentially fit the training data points. Sparsitybased arguments would suggest that the generalization error increases as N grows past a certain threshold N*. Instead, empirical studies have shown that in the over-parametrized regime, generalization error keeps decreasing with N. We resolve this paradox through a new framework. We rely on the so-called Neural Tangent Kernel, which connects large neural nets to kernel methods, to show that the initialization causes finite-size random fluctuations parallel to f(N) - < f(N)>parallel to similar to N-1/4 of the neural net output function f(N) around its expectation < f(N)>. These affect the generalization error epsilon(f(N)) for classification: under natural assumptions, it decays to a plateau value epsilon(f(infinity)) in a power-law fashion similar to N-1/2. This description breaks down at a so-called jamming transition N = N*. At this threshold, we argue that parallel to f(N)parallel to diverges. This result leads to a plausible explanation for the cusp in test error known to occur at N*. Our results are confirmed by extensive empirical observations on the MNIST and CIFAR image datasets. Our analysis finally suggests that, given a computational envelope, the smallest generalization error is obtained using several networks of intermediate sizes, just beyond N*, and averaging their outputs.