Visualizando la Toma de Decisiones de una Red Convolucional Mediante Mapeo de Activación de Clases Ponderado por Gradiente

Resumen

The success of deep learning in many kinds of applications has been observed; in the case of computer vision, there are applications such as classification, object detection, face recognition, etc; however, these kinds of artificial intelligence systems are not transparent to users, which means that we don ́t know how they make their decisions to support their answers. The lack of interpretability and explainability in black-box models represents a significant challenge in modern artificial intelligence systems, such as convolutional neural networks (CNNs). This paper explores the behavior of Gradient-weighted Class Activation Mapping (Grad-CAM) on the MNIST dataset, whose content is manuscript digit images in gray scale. The objective is to identify which input image regions have the greatest influence when the CNN classifies that input image, providing a visual interpretation that facilitates understanding of the classifier’s behavior. Experimental results show that Grad-CAM highlights the relevant stroke areas in each digit image, which helps to detect ambiguity and allows correct misclassification.