Structural Features of the Neural Network Classifier

Abstract

One of the promising areas in signal/image analysis systems and pattern recognition systems is the use of neural network technologies. This approach has been widely used in technical and medical diagnosis with their hardware and software implementation, in particular in medical express diagnostics. A feature of this approach is the possibility of implementing a dialogue mode, simultaneous processing of alternative versions and processing of symbol variables by recognizing information of various nature. This article analyzes the structural, functional and training features of two classical neural networks: the Hopfield network and the Hamming network, which is the simplest classifier of binary vectors. Taking into account the advantages of both of these neural networks it has been allowed to develop the structure and functioning principle of the proposed neural network classifier. The presented structure of the neural network classifier is an improvement in the structure of the Hamming neural network. The difference is the removal in the neural network classifier of positive lateral connections in the neurons of the competitive layer, which implements the well-known WTA paradigm (winner takes all). And this causes attenuation of weak output signals to a level below the sensitivity threshold. Thus, the WTA strategy is implemented, which stops the iterative process in case of victory of one of the neurons in the competitive layer. Such an approach allowed not only to simplify the structure of the neural network classifier, but also to expand the scope of its application for classification by the maximum of discriminant functions. Simulation of the classification process in the proposed neural network classifier confirmed the acceleration of this process by almost 2 times. Structural modeling of the hidden layer of the neural network classifier demonstrated the correct answers at its outputs when specifying specific input combinations.