Computational Separation Between Convolutional and Fully-Connected Networks

Abstract

Convolutional neural networks (CNN) exhibit unmatched performance in a multitude of computer vision tasks. However, the advantage of using convolutional networks over fully-connected networks is not understood from a theoretical perspective. In this work, we show how convolutional networks can leverage locality in the data, and thus achieve a computational advantage over fully-connected networks. Specifically, we show a class of problems that can be efficiently solved using convolutional networks trained with gradient-descent, but at the same time is hard to learn using a polynomial-size fully-connected network.

Cite

Text

Malach and Shalev-Shwartz. "Computational Separation Between Convolutional and Fully-Connected Networks." International Conference on Learning Representations, 2021.

Markdown

[Malach and Shalev-Shwartz. "Computational Separation Between Convolutional and Fully-Connected Networks." International Conference on Learning Representations, 2021.](https://mlanthology.org/iclr/2021/malach2021iclr-computational/)

BibTeX

@inproceedings{malach2021iclr-computational,
  title     = {{Computational Separation Between Convolutional and Fully-Connected Networks}},
  author    = {Malach, Eran and Shalev-Shwartz, Shai},
  booktitle = {International Conference on Learning Representations},
  year      = {2021},
  url       = {https://mlanthology.org/iclr/2021/malach2021iclr-computational/}
}