Extending class activation mapping using Gaussian receptive field

Abstract

Understanding the inner behaviors of deep neural networks is an important issue, and various visualization methods have been suggested. However, recent studies showed that most existing saliency methods have low visualization accuracy, even compared to a random importance assignment. In this study, we seek an improved saliency algorithm. Focusing on class activation mapping (CAM)-based saliency methods, we discuss two problems with the existing studies. First, we introduce conservativeness, a property that prevents redundancy and deficiency in saliency map and ensures that the saliency map is on the same scale as the prediction score. We identify that existing CAM studies do not satisfy the conservativeness and derive a new CAM equation with the improved theoretical property. Second, we discuss the common practice of using bilinear upsampling as problematic. We propose Gaussian upsampling, an improved upsampling method that reflects deep neural networks’ properties. Based on these two options, we propose Extended-CAM, an advanced CAM-based visualization method. In various visualization benchmarks, datasets, and architectures, our Extended-CAM presents a more accurate visualization than is obtainable with other existing methods.