P. E. Khrustalyov

Considering a direct connection of productivity of procedure of recognition with level of generalising ability of a neural network which represents itself as the qualifier, the purpose of the given work is increase of probability of correct classification of object by productive development of generalising ability of multilayer perceptron. For definition of productivity of procedure of classification of object we will enter criterion on which the estimation of results of recognition will be made. The standard error will be an indicative measure of difference of the demanded signal formed by the qualifier, from actual. It appears, that generalising ability of a neural network probably to improve having some knowledge of points of a local minimum of a surface of function of errors. For maintenance of qualitative classification of objects by multilayer perceptron the heuristic technique of training of a network which consists in the following is offered: such combination of parametres of training of a network at which the network comes into a local minimum on a significant amount of epoch is searched. If the network on achievement of certain quantity of epoch of training "leaves" a local minimum in which it has been placed - the error of classification of images continues to aspire to the minimum value. When the classification error will reach the value established by the designer, network training comes to an end. According to the received results the technique of the accelerated training of a neural network which consists in avoiding of cases of hit of a network in local minima, negatively affects generalising ability of a network, in comparison with the described heuristic method. From here follows, that the statement about dependence of generalising ability of a neural network only from construction of training sample, architecture of a network and complexity of a solved problem is expedient for calling into question. Also probably to assume, that the local minimum is a barrier, passing which multilayered perceptron develops generalising ability most productively.