English

Signal Recovery from Pooling Representations

Machine Learning 2014-03-03 v3

Abstract

In this work we compute lower Lipschitz bounds of p\ell_p pooling operators for p=1,2,p=1, 2, \infty as well as p\ell_p pooling operators preceded by half-rectification layers. These give sufficient conditions for the design of invertible neural network layers. Numerical experiments on MNIST and image patches confirm that pooling layers can be inverted with phase recovery algorithms. Moreover, the regularity of the inverse pooling, controlled by the lower Lipschitz constant, is empirically verified with a nearest neighbor regression.

Cite

@article{arxiv.1311.4025,
  title  = {Signal Recovery from Pooling Representations},
  author = {Joan Bruna and Arthur Szlam and Yann LeCun},
  journal= {arXiv preprint arXiv:1311.4025},
  year   = {2014}
}

Comments

17 pages, 3 figures

R2 v1 2026-06-22T02:08:42.755Z