Estimation of the relevance of the neural network parameters

The paper investigates a method for optimizing the structure of a neural network. It is assumed that the number of neural network parameters can be reduced without significant loss of quality and without significant increase in the variance of the loss function. The paper proposes a method for automatic estimation of the relevance of parameters to prune a neural network. This method analyzes the covariance matrix of the posteriori distribution of the model parameters and removes the least relevant and multicorrelate parameters. It uses the Belsly method to search for multicorrelation in the neural network. The proposed method was tested on the Boston Housing data set, the Wine data set, and synthetic data.