[2Q3-J-2] Machine learning: explainability, knowledge acquisition

[2Q3-J-2-01] Investigation of the influence of over training on CNN focusing on change of heat map which brings interpretability

〇Yoshihisa Furusawa¹, Yoshimasa Tawatsuji^3,2, Tatsunori Matsui² (1. Waseda University, 2. Faculty of Human Sciences, Waseda University, 3. Graduate School of Human Sciences, Waseda University)

Keywords:deep learning, interpretability, visualization

Although Convolutional Neural Network (CNN) is used in many studies, the interpretability of CNN have been considered problematic and various methods have already been proposed from related research. All of these methods have been verified with selected models based on Early Stopping and are evaluated with only one model. However, in the classification problem, it is reported that the accuracy increases due to over training, so it is not known whether model selection by Early Stopping is appropriate. Therefore, in this research, the influence of this over training is considered from the viewpoint of the explanability of CNN. As a result, for each learning period, we found a dataset in which the value of the correct class and entropy of the output oscillate in the learning process. And we found that it is possible to create a heatmap with different similarity from the heatmap created by Early Stopping.