Spline-Orthogonal Extended Neo-Fuzzy Neuron
Yevgeniy V. Bodyanskiy, Yuriy P. Zaychenko, Nonna Ye. Kulishova, Galib Hamidov · 2019
The development of successful human-computer interaction means depends on the possibilities of automatic recognition of signals that are natural or familiar to people. These signals include speech, handwritten texts, gestures, facial expressions, and so on. To develop interfaces taking into account such signals, artificial intelligence methods, in particular, artificial neural networks, are being widely used now. One of the vast classes of neural networks are hybrid neuro-fuzzy systems, which are characterized by universal approximation capabilities and considerable computational simplicity. Such systems allow the implementation of speed-optimal learning algorithms also. These qualities make hybrid neuro-fuzzy systems very useful in solving classification problems, especially for online mode. This paper proposes an architecture and learning algorithm for spline-orthogonal extended neo-fuzzy neuron, which allows simplifying the network learning process when new data is received. For this purpose, the usage of nonlinear synapses based on Chebyshev orthogonal polynomials for neo-fuzzy neuron is proposed. B-splines are used as fuzzy membership functions. The effectiveness of the developed architecture and learning algorithm is studied on the example of the problem of basic person emotional states recognition in real time.