The architecture of a neural network with a sequential division of images into pairs

The known classical architectures of neural networks have many weak properties and high limitations such as great difficulties in choosing proper parameters (the numbers of neurons, connections, layers), in learning and in expanding a learned network and some others. In this paper, to overcome these shortcomings, we consider the architecture of a neural network in which the recognition is performed in pairs. The neural network is created on the basis of the neural network architecture that implements the method of the nearest neighbor, without using analytical expressions and a set of selected samples. The paper shows that such an architecture can be applied to the recognition problem with a very large number of images (classes). At the same time, the proposed neural network has a simple architecture, with the possibility of simpler learning of the neural network. The neural network architecture allows to add new recognizable images to the neural network without changing the previous network settings.