A Bayesian framework for extreme learning machine with application for automated cancer detection

Abstract

Fairly recently, extreme learning machine (ELM) has been proposed as a single-hidden layer feedforward neural network (SLFN), where the input weights are randomly initiated and never updated, and the output weights are analytically computed. Setting the parameters of the hidden layer randomly may not be always effective if the function that is learned is not simple and the amount of labeled data is not small, even if theoretical studies have shown that ELM maintains the universal approximation capability. To address this issue, we propose a new approach inspired by the Bayesian paradigm as an alternative to the random initiation of the hidden node parameters. The idea behind this model is that we can use the information (prior knowledge) about a certain labeled data through the correlation between attributes and decision classes. The prior knowledge is acquired through the Goodman-Kruskal Gamma rank correlation between attributes and labels, assuming that the input weights should be related to the influence of attributes upon labels. Five publicly available high-dimensional datasets regarding cancer (breast, lung, colon, and ovarian) related to cDNA arrays, DNA microarray, and mass spectroscopy are used for experimentation and model assessment. We compared the performance of this classifier with that of three 'neighboring' algorithms, such as a basic ELM, a SLFN trained by backpropagation (BP) algorithm, and a radial basis function network (RBF). The experimental results undoubtedly indicated that the proposed variant of ELM is very effective and its performance is superior to that of the comparison models.