Volume 9, Number 3, p.p. 65–68
Elaboration of neural network learning method using pattern recognition
Otar Verulava, Otar Tavdishvili, Tea Todua and Lasha Verulava
Georgian Technical University, 77 Kostava St, 0175 Tbilisi, Georgia
The basic problem of pattern recognition, notably achieving recognition of unknown realizations without error, is considered. A neural network learning process is used to solve it. At the first stage, training set recognition occurs, and then the synaptic coefficients are changed on the basis of the recognition results; i.e., to form and correct the neurotemplate descriptions. At the first stage using the superposition principle, the neuromini and neuromaxi portraits of each pattern are obtained. On the basis of these portraits, a priori estimation of recognition results is achieved, and especially the conditions (criteria) when recognition of the realizations of the training and testing sets is possible without errors. During the neural network learning process each act of recognition represents a stage of learning, in the course of which, depending on recognition results, correction-change of the synaptic coefficients and thresholds of the neurons are made in order to correct recognition errors. After termination of the learning process, the elements of the set of the synaptic coefficients (simultaneously representing the pattern descriptions—neurotemplates) acquire values that will provide recognition of the training set realizations of a given pattern set without error.
Keywords:
contending IMLS fits, extrapolation of potential energy surfaces, Morse potential