Detector neural network vs connectionist ANNs

Abstract

Most widely used modern artificial neural networks are based on the connectionist paradigm of building and learning. The authors propose an alternative detector approach. The basis of this approach is the original architecture of the neural network, as well as a new procedure for its learning. The developed neural network is called the detector neural network. This network consists of two layers of neurons. The neurons of the first layer are called neurons-pre-detectors and they do not learn. They are designed to highlight the structural elements of recognizable images, as well as to determine their measured parameters. The types of structural elements and their parameters are set a priori and depend on the type and complexity of recognizable images. Neurons of the second layer can be trained. They recognize individual complex images. These neurons are called neurons-detectors (ND). The model of the ND is significantly different from all known models of neurons and has important features: the presence of a dendritic tree model, a new looks at the role and value of synaptic coupling coefficients, an original approach to the formation of a neuron reaction. The training procedure for the ND is called counter learning. In the process of counter learning, an information model – template of the most simplified image structure of a particular class is formed and remembered by the ND. This template is called a concept. The main role in the formation of the concept is played by the model of the dendritic tree of the neuron. During the classification process, the ND tries to simplify the recognizable image until it coincides with the concept. The article provides a comparative analysis of the detector approach and the connectionist paradigm. The advantages of the detector approach, according to the authors, open up new possibilities in the study of the problem of constructing Artificial General Intelligence.