Extreme learning machine with diversified neurons
Jacek Kabziński · 2016
It was demonstrated that the standard selection of input weights and biases for Extreme Learning Machine (ELM) may lead to ill-conditioning of the output weights calculation and result in great values of the output weights. Two slight modifications of the standard approach were proposed: i) addition of some distance-based neurons (like Radial Basis Functions - RBF) with centers situated at numerically identified local extrema of the training data, ii) modification of random generation of input weights and biases for sigmoid neurons to enable enhanced variation of activation functions. The first modification improves the modeling accuracy for a moderate number of hidden neurons and the second allows to improve the conditioning of output weights calculation and reduce the optimal output weights significantly.