Understanding the decision-making process of cnn in modulation recognition via iterative channel relevance

Deep learning techniques, with their excellent feature extraction and representation, are being employed more and more in signal modulation recognition. However, most modulation identification systems lack transparency and are referred to as black boxes. The field has a restricted number of interpretable ways, and the currently available interpretable techniques lack considerable localization features. To address this issue, we present a method called Iterative Channel Relevance that enhances the Class Activation Mapping methodology(ICR-CAM). This proposed method is to provide a clearer understanding of how modulation recognition networks make judgments. By introducing the concept of channel scores, which assess the level of influence that shallow layers have on deep layers during the process of forward propagation. Coupled with the relevance score produced from backpropagation, this complicated score as the initial weight for every channel. Furthermore, drawing influence from the fundamentals of mask iterative optimization, This method provides for a thorough evaluation of channel significance in both the forward and backward directions, can create saliency map which strongly connect to the model’s internal decision-making mechanism. Extensive investigations have revealed that ICR-CAM offers a more accurate visual localization capability compared to earlier approaches, as evidenced in both visual and quantitative assessments.