English

Bounds on learning in polynomial time

Disordered Systems and Neural Networks 2017-02-08 v1

Abstract

The performance of large neural networks can be judged not only by their storage capacity but also by the time required for learning. A polynomial learning algorithm with learning time N2\sim N^2 in a network with NN units might be practical whereas a learning time eN\sim e^N would allow rather small networks only. The question of absolute storage capacity αc\alpha_c and capacity for polynomial learning rules αp\alpha_p is discussed for several feed-forward architectures, the perceptron, the binary perceptron, the committee machine and a perceptron with fixed weights in the first layer and adaptive weights in the second layer. The analysis is based partially on dynamic mean field theory which is valid for NN\to\infty. Especially for the committee machine a value αp\alpha_p considerably lower than the capacity predicted by replica theory or simulations is found. This discrepancy is resolved by new simulations investigating the learning time dependence and revealing subtleties in the definition of the capacity.

Keywords

Cite

@article{arxiv.cond-mat/9705259,
  title  = {Bounds on learning in polynomial time},
  author = {Heinz Horner and Anthea Bethge},
  journal= {arXiv preprint arXiv:cond-mat/9705259},
  year   = {2017}
}

Comments

12 pages Latex with 8 eps figures