Development of an Approach to Implementation of a Model Based on a Generalized Artificial Neural Network Concept

摘要

An approach is proposed to create a model configuration that generalizes the properties of the pre-trained artificial neural network (ANN) submodels included in it, processing the corresponding input data disjoint arrays. The following options for constructing a generalized neural network are considered: a system of pre-trained neural network submodels connected in series to each other; artificial neural network with forced adjustment of weights based on data obtained by processing each submodel of the corresponding data array; sequential connected system of ANN-submodels with adjustment layers; generalized neural network with ANN submodels generating noise signals at additional inputs of the system on the corresponding layers. The limitations and the resulting disadvantages of each of the considered approaches are determined. The best approach has been identified that guarantees the efficiency of a model that satisfies the concept of a generalized neural network model while preserving the properties of the ANN submodels included in it.