Extreme learning machine adapted to noise based on optimization algorithms

Abstract

The extreme learning machine for neural networks of feedforward of a single hidden layer randomly assigns the weights of entry and analytically determines the weights the output by means the Moore-Penrose inverse, this algorithm tends to provide an extremely fast learning speed preserving the adjustment levels achieved by classifiers such as multilayer perception and support vector machine. However, the Moore-Penrose inverse loses precision when using data with additive noise in training. That is why in this paper a method to robustness of extreme learning machine to additive noise proposed. The method consists in computing the weights of the output layer using non-linear optimization algorithms without restrictions. Tests are performed with the gradient descent optimization algorithm and with the Levenberg-Marquardt algorithm. From the implementation it is observed that through the use of these algorithms, smaller errors are achieved than those obtained with the Moore-Penrose inverse.