Proof of the Theory-to-Practice Gap in Deep Learning via Sampling Complexity bounds for Neural Network Approximation Spaces

Philipp Grohs (Corresponding author), Felix Voigtlaender

Publications: Working paper

Abstract

We study the computational complexity of (deterministic or randomized) algorithms based on point samples for approximating or integrating functions that can be well approximated by neural networks. Such algorithms (most prominently stochastic gradient descent and its variants) are used extensively in the field of deep learning. One of the most important problems in this field concerns the question of whether it is possible to realize theoretically provable neural network approximation rates by such algorithms. We answer this question in the negative by proving hardness results for the problems of approximation and integration on a novel class of neural network approximation spaces. In particular, our results confirm a conjectured and empirically observed theory-to-practice gap in deep learning. We complement our hardness results by showing that error bounds of a comparable order of convergence are (at least theoretically) achievable.

Original languageEnglish
Pages1085-1143
Number of pages59
Volume24
DOIs
Publication statusPublished - Aug 2024

Publication series

SeriesFoundations of Computational Mathematics
ISSN1615-3375

Austrian Fields of Science 2012

  • 101032 Functional analysis
  • 101031 Approximation theory
  • 102018 Artificial neural networks

Keywords

  • Approximation spaces
  • Deep neural networks
  • Gelfand numbers
  • Information based complexity
  • Randomized approximation
  • Theory-to-computational gaps
  • 41A25
  • 68T05
  • Secondary 41A65
  • 65Y20
  • 68T07
  • Primary 41A46

Fingerprint

Dive into the research topics of 'Proof of the Theory-to-Practice Gap in Deep Learning via Sampling Complexity bounds for Neural Network Approximation Spaces'. Together they form a unique fingerprint.

Cite this