Structure Discovery by Visualization of Deep Neural Network for EEG

Approach utilizing Between-Model Variance

Abstract

Recently, research using deep learning has been conducted in various fields. However, deep neural network (DNN) has a problem that it is unclear how the model extract features internally. Therefore, "Explainable AI" is required, and research such as visualization has been conducted. In this research, we assume that a learned model with higher classification accuracy pays attention to more essential structure, and compared the visualization results between models which has different classification accuracy. Using this, we examined the method to discover the essential structure of the target phenomenon by the visualization of learned model for the same phenomenon. We conducted an experiment to learn Event Related Potential (ERP) from EEG using DNN. As a result, characteristic structures known in ERP research was obtained from a visualization analysis of the learned model.