The VC-Dimension vs. the Statistical Capacity for Two Layer Networks with Binary Weights

Abstract

A general relationship is developed between the VC-dimension and the statistical capacity which shows that the VC-dimension can be lower bounded (in order) by the statistical error-tolerant capacity of a network composed of random samples drawn from a specific class of distributions. This relationship is then used to actually find the error-tolerant capacity of a specific network which results in a lower bound for the VC-dimension of two layer (N – 2L – 1) feedforward networks with binary interconnections and integer thresholds. For large N and L, the VC-dimension db of such networks is shown to satisfy the relation , where N and 2L are the number of input units and hidden units respectively, and W is the total number of weights.

Cite

Text

Ji and Psaltis. "The VC-Dimension vs. the Statistical Capacity for Two Layer Networks with Binary Weights." Annual Conference on Computational Learning Theory, 1991. doi:10.1016/B978-1-55860-213-7.50026-2

Markdown

[Ji and Psaltis. "The VC-Dimension vs. the Statistical Capacity for Two Layer Networks with Binary Weights." Annual Conference on Computational Learning Theory, 1991.](https://mlanthology.org/colt/1991/ji1991colt-vc/) doi:10.1016/B978-1-55860-213-7.50026-2

BibTeX

@inproceedings{ji1991colt-vc,
  title     = {{The VC-Dimension vs. the Statistical Capacity for Two Layer Networks with Binary Weights}},
  author    = {Ji, Chuanyi and Psaltis, Demetri},
  booktitle = {Annual Conference on Computational Learning Theory},
  year      = {1991},
  pages     = {250-256},
  doi       = {10.1016/B978-1-55860-213-7.50026-2},
  url       = {https://mlanthology.org/colt/1991/ji1991colt-vc/}
}