Bridging the Gap Between Approximation and Learning via Optimal Approximation by ReLU MLPs of Maximal Regularity

Abstract

The foundations of deep learning are supported by the seemingly opposing perspectives of approximation or learning theory. The former advocates for large/expressive models that need not generalize, while the latter considers classes that generalize but may be too small/constrained to be universal approximators. Motivated by real−world deep learning implementations that are both expressive and statistically reliable, we ask: "Is there a class of neural networks that is both large enough to be universal but structured enough to generalize?" This paper constructively provides a positive answer to this question by identifying a highly structured class of ReLU multilayer perceptions (MLPs), which are optimal function approximators and are statistically well−behaved. We show that any L−Lipschitz function from [0,1]ᵈ to [−n,n] can be approximated to a uniform Ld/(2n) error on [0,1]ᵈ with a sparsely connected L−Lipschitz ReLU MLP of width 𝒪(dnᵈ), depth 𝒪(łog(d)), with 𝒪(dnᵈ) nonzero parameters, and whose weights and biases take values in {0,± 1/2} except in the first and last layers which instead have magnitude at−most $n$. Unlike previously known "large" classes of universal ReLU MLPs, the empirical Rademacher complexity of our class remains bounded even when its depth and width become arbitrarily large.

Further, our class of MLPs achieves a near−optimal sample complexity of 𝒪(łog(N)/√{N}) when given N i.i.d. normalized sub−Gaussian training samples. We achieve this by avoiding the standard approach to constructing optimal ReLU approximators, which sacrifices regularity by relying on small spikes. Instead, we introduce a new construction that perfectly fits together linear pieces using Kuhn triangulations and avoids these small spikes.