Addressing Multi-class Classification Tasks by means of RBFNN-like Models using Modular Indistinguishability Operators

Abstract

[eng] As a machine learning model, a Radial Basis Function Neural Network (RBFNN) is an artificial neural network that comprises a single hidden layer whose neurons implement a set of radial-basis functions that are linearly combined at the output layer. The most popular form of RBFNN makes use of the Gaussian RBF combined with the Euclidean distance. In this work, we explore the use of Modular Indistinguishability Operators (MIO) in the hidden layer of RBFNN-like structures in replacement of Gaussian RBFs. Moreover, we introduce a new modular metric m(t, p, q) and next use it to derive a specific MIO IL(t, a, b) from the Luckasiewicz t-norm. The resulting MIO is next used as the activation function of the hidden neurons, whose parameters are determined on the basis of the samples of the underlying dataset. The performance of the new machine learning model is evaluated through a set of publicly available datasets in the context of multi-class classification tasks.