Verification of generality from weight analysis of Integration Neural Network approximators

Abstract

Multiple regression analysis, neural networks (NN), and other methods are used as approximate computation methods for physical phenomena. However, multiple regression analysis has the problem of overlearning due to factors that cannot be formulated, and NN has the problem of overlearning due to the diversity of its representational capabilities. In response to this problem, we have proposed an integrated neural network (INN) that integrates a simple perceptron and an NN, which mimic multiple regression analysis. In this paper, we show that INN can reduce overlearning of the simple perceptron part (multiple regression analysis) and achieve higher accuracy by compensating for errors caused by NNs through weight analysis. Furthermore, it was revealed that the correction of the NN part is effective under adaptive control according to the approximation accuracy of the simple perceptron part.