Nonparametric Weight Initialization of Neural Networks via Integral Representation
Machine Learning
2014-02-20 v3 Neural and Evolutionary Computing
Abstract
A new initialization method for hidden parameters in a neural network is proposed. Derived from the integral representation of the neural network, a nonparametric probability distribution of hidden parameters is introduced. In this proposal, hidden parameters are initialized by samples drawn from this distribution, and output parameters are fitted by ordinary linear regression. Numerical experiments show that backpropagation with proposed initialization converges faster than uniformly random initialization. Also it is shown that the proposed method achieves enough accuracy by itself without backpropagation in some cases.
Cite
@article{arxiv.1312.6461,
title = {Nonparametric Weight Initialization of Neural Networks via Integral Representation},
author = {Sho Sonoda and Noboru Murata},
journal= {arXiv preprint arXiv:1312.6461},
year = {2014}
}
Comments
For ICLR2014, revised into 9 pages; revised into 12 pages (with supplements)