G. Thimm and E. Fiesler, IDIAP, C.P. 592, CH-1920 Martigny, Switzerland, Georg.Thimm@IDIAP.CH
Proper initialization is one of the most important prerequisites for fast convergence of feed-forward neural networks like high order and multilayer perceptrons. This publication aims at determining the optimal value of the initial weight variance (or range), which is the principal parameter of random weight initialization methods for both types of neural networks.
An overview of random weight initialization methods for multilayer perceptrons is presented. These methods are extensively tested using eight real-world benchmark data sets and a broad range of initial weight variances by means of more than 30,000 simulations, in the aim to find the best weight initialization method for multilayer perceptrons.
For high order networks, a large number of experiments (more than 200,000 simulations) was performed, using three weight distributions, three activation functions, several network orders, and the same eight data sets. The results of these experiments are compared to weight initialization techniques for multilayer perceptrons, which leads to the proposal of a suitable weight initialization method for high order perceptrons.
The conclusions on the weight initialization methods for both types of networks are justified by sufficiently small confidence intervals of the mean convergence times.
Keywords: neural network initialization, random weight initialization, initial weight, weight initialization, interconnection strength, comparison of weight initialization methods, high(er) order neural network, high(er) order perceptron, sigma-pi connection, initial weight distribution, activation function, learning rate, multilayer perceptron, neural computation, neural network, neurocomputing, optimization, connectionism, real-world benchmark