175

Generalization Bounds for Neural Networks via Approximate Description Length

Neural Information Processing Systems (NeurIPS), 2019
Abstract

We investigate the sample complexity of networks with bounds on the magnitude of its weights. In particular, we consider the class \[ H=\left\{W_t\circ\rho\circ \ldots\circ\rho\circ W_{1} :W_1,\ldots,W_{t-1}\in M_{d, d}, W_t\in M_{1,d}\right\} \] where the spectral norm of each WiW_i is bounded by O(1)O(1), the Frobenius norm is bounded by RR, and ρ\rho is the sigmoid function ex1+ex\frac{e^x}{1+e^x} or the smoothened ReLU function $ \ln (1+e^x)$. We show that for any depth tt, if the inputs are in [1,1]d[-1,1]^d, the sample complexity of HH is O~(dR2ϵ2)\tilde O\left(\frac{dR^2}{\epsilon^2}\right). This bound is optimal up to log-factors, and substantially improves over the previous state of the art of O~(d2R2ϵ2)\tilde O\left(\frac{d^2R^2}{\epsilon^2}\right). We furthermore show that this bound remains valid if instead of considering the magnitude of the WiW_i's, we consider the magnitude of WiWi0W_i - W_i^0, where Wi0W_i^0 are some reference matrices, with spectral norm of O(1)O(1). By taking the Wi0W_i^0 to be the matrices at the onset of the training process, we get sample complexity bounds that are sub-linear in the number of parameters, in many typical regimes of parameters. To establish our results we develop a new technique to analyze the sample complexity of families HH of predictors. We start by defining a new notion of a randomized approximate description of functions f:XRdf:X\to\mathbb{R}^d. We then show that if there is a way to approximately describe functions in a class HH using dd bits, then d/ϵ2d/\epsilon^2 examples suffices to guarantee uniform convergence. Namely, that the empirical loss of all the functions in the class is ϵ\epsilon-close to the true loss. Finally, we develop a set of tools for calculating the approximate description length of classes of functions that can be presented as a composition of linear function classes and non-linear functions.

View on arXiv
Comments on this paper