Feed Forward Neural Network entities

Abstract

Feed Forward Neural Networks (FFNNs) are computational techniques inspired by the physiology of the brain and used in the approximation of general mappings from one finite dimensional space to another. They present a practical application of the theoretical resolution of Hilbert's 13^{th problem} by Kolmogorov and Lorenz, and have been used with success in a variety of applications. However, as the training data grows in both dimension and size, larger network implementations are required. As a consequence, in most cases scaling problems arise; the existing training algorithms can not handle the vast search space, saturation occurs at the output of the hidden layer nodes and, in general, the network becomes inflexible, slow and inefficient. Considering all of the above, we are proposing a methodology for breaking down the traditionally single and rigid FFNN into an entity of simpler units, in line with the connectionist view of the distributive representation of knowledge and using Kolmogorov's paradigm of approximating functions of many variables by compositions of one-variable functions. Although the entities' concept is still developing, some preliminary results indicate superiority over the single FFNN model when applicable to problems involving high-dimensional data (e.g. financial/meteorological data analysis, etc.).