Autoeoncoders and Information Augmentation for Improved Generalization and Interpretation in Multi-Layered Neural Networks

The present paper aims to propose a new type of learning method for multi-layered neural network to solve the vanishing information problem. The vanishing information problem means that multi-layered neural networks tend to lose their error information as well as input information by going through many hidden layers. To overcome this problem, the new method tries to capture information in inputs as much as possible by increasing the number of inputs, producing composite input variables. The new method was applied to the symmetric data set and wine data sets. For the symmetric data set, the new method could capture the symmetric property of input data set with better generalization performance. For the wine data set, the new method could capture combined characteristics detected by the bagging method and logistic regression analysis with better generalization performance.