Error bounds for approximations with deep ReLU networks

Abstract

We study expressive power of shallow and deep neural networks with piece-wise linear activation functions. We establish new rigorous upper and lower bounds for the network complexity in the setting of approximations in Sobolev spaces. In particular, we prove that deep ReLU networks more efficiently approximate smooth functions than shallow networks. In the case of approximations of 1D Lipschitz functions we describe adaptive depth-6 network architectures more efficient than the standard shallow architecture.

31 pages; major revision in v3; submitted to Neural Networks

Cited by in corpus (6)

• Neural networks and rational functions
• The universal approximation power of finite-width deep ReLU networks
• Neural tangent kernels, transportation mappings, and universal approximation
• A gradual, semi-discrete approach to generative network training via explicit Wasserstein minimization
• Hierarchically Compositional Tasks and Deep Convolutional Networks
• Algorithmic Complexities in Backpropagation and Tropical Neural Networks