A 26.55TOPS/W Explainable AI Processor with Dynamic Workload Allocation and Heat Map Compression/Pruning

Abstract

Explainable artificial intelligence (XAI) explains the decision of an artificial intelligence (AI) system by generating step-by-step evidence. It is a crucial paradigm shift in machine learning (ML) for more reliable and interoperable usage of inference results. Class activation mapping (CAM), which visualizes the intensity of the features of convolutional neural network (CNN), is one of the first and most popular algorithms in XAI. Grad-CAM [1] is an extended version of CAM which enables the heat map generation of each class for more generic usage. It is widely used for various applications such as feature factorization and weak supervision. As shown in Fig. 1, Grad-CAM requires both forward propagation (FP) for inference (INF) and backward propagation (BP) with the heat map generation for explanation (EXP). Unlike conventional training, Grad-CAM’s BP stage only generates specific gradient maps for target classes by performing element-wise multiplication with feature maps and non-linear functions, not accumulating all the gradients. It is allowed to use a simpler fixed-point (FXP) format rather than a floatingpoint format for the heat maps in EXP. In addition, since the Grad-CAM’s heat map points out specific areas of target classes in the image, it is globally sparse but locally dense. To obtain a detailed explanation with inference, efficient hardware acceleration is required for EXP, which has a different manner against training.