Signal based Dilation Convolution CAM for Feature Extraction and Analysis in CNN Model

Abstract We propose a new kind of model interpretation and feature display method for quantitatively analysis of deep learning model’s working mechanism in physiological feature extraction and classification. Deep learning is faced with lack of interpretation and inner mechanism for its results, which has limited its application in some high-risk field, such as medical diagnosis. The method, designated as SDC-CAM, raised in this article was developed to display the features curve and key features during deep learning model process. And SDC-CAM can quantitatively show the tendency of specific data on different classification. A threshold value can be set to visually display the colored region as key features to show the working mechanism of convolutional neural network(CNN) model. We believe this method can provide positive effect for CNN interpretation and understanding.