Expert-Informed, User-Centric Explanations for Machine Learning

We argue that the dominant approach to explainable AI for explaining image classification, annotating images with heatmaps, provides little value for users unfamiliar with deep learning. We argue that explainable AI for images should produce output like experts produce when communicating with one another, with apprentices, and with novices. We provide an expanded set of goals of explainable AI systems and propose a Turing Test for explainable AI. Explaining Image Classification Explaining the decisions of AI has emerged as an important research topic. Considerable progress in image classification using deep learning (Krizhevsky, et al., 2012; LeCun, et al., 2015) has created significant interest in explaining the results of image classification. Although there are many applications for explainable AI (XAI), this paper first focuses on learning to classify images. We then discuss broader implications for explainable AI. Recent conferences include tutorials and workshops on explainable AI. There are several good surveys of XAI (Chakraborty et al., 2017 & Došilović et al., 2018) . This is not one of them. Instead, after working on problems with experts in radiology and ophthalmology and on bird identification, we have concluded that existing techniques leave much room for improvement. The field needs additional directions and methodology, including clarifying XAI's goals, particularly with respect to users, experts, and image classification. Although some of XAI's original goals were to "explain their decisions and actions to human users" (Gunning & Aha, 2018) the current state-of-the-art is developer-centric rather than user-centric. The dominant method for explaining image classification is assigning an importance score to pixels or regions on a saliency map or heatmap superimposed on an image, visualizing a region's importance with color scales (red, orange, yellow…). Methods developed for creating heatmaps include occlusion sensitivity (Zeiler &