Information Augmentation, Reduction and Compression for Interpreting Multi-layered Neural Networks

Abstract

The present paper aims to propose a new type of learning method for interpreting relations between inputs and outputs in multi-layered neural networks. The method is composed of information augmentation, reduction and compression component. In the information augmentation component, information in inputs is forced to increase for the subsequent learning to choose appropriate information among many options. In the information reduction component, information is reduced by selectively choosing strong and active connection weights. Finally, in the information compression component, information contained in multi-layered neural networks is compressed by multiplying all connection weights in all layers for summarizing the main characteristics of connection weights. The method was applied to the improvement of an EC (electric commerce) web site for better profitability. The method could clarify relations between inputs and outputs and its interpretation was more natural than that by the conventional logistic regression analysis. The results suggest that multi-layered neural networks can be used to improve generalization and in addition to interpret final results, which is more important in many applications fields.