F-neuron in the clustering problems

Abstract

We propose a new type of neuron based on the use of Fourier transform properties. This new type of neuron, called Fourier neuron (F-neuron), simplifies solving of a range of problems belonging to the class of problems of creating self-organizing networks using teacherless learning. The application of such F-neuron improves the quality and efficiency of automatic clustering of objects. We described the basic principles and approaches that allow to consider the properties vector as a parametric piecewise linear function, which provides the possibility to switch to Fourier-images operation both for input objects and for learning weights. The reasons for transferring information processing to Fourier space are justified, automatic orthogonalization and ranking of the Fourier image of the feature vector is explained. The advantages of the statistical approach to neuron training and construction of the refined neuron state function based on the parameters of the normal distribution are analyzed. We describe the procedure of training and pre-training the F-neuron that uses a statistical model based on the use of parameters of a normal distribution to calculate the confidence interval. We described an algorithm for recalculating normal distribution parameters when a new sample is added to the cluster. We reviewed some results of F-neuron technology and compared it with a traditional perceptron. A list of references and citations to the author’s previous works are given below.