What Explainable AI Cannot Explain And What Can Be Done

The effectiveness of a machine learning model is often marred with its inability to explain its decisions to the users. To address this problem, a whole new branch of explainable AI (XAI) has emerged, and the researchers are actively pursuing different methodologies to establish a user-friendly AI.  But what about the existing XAI approaches? Are they any good? Where do they fail? To answer these questions, a team of researchers from UC Berkeley and Boston University have investigated the challenges and possible solutions. Their exploration led to a novel technique that will be discussed in the last section of this article. Explaining The Inexplicable To illustrate the inexplicability, one of the authors, Alvin Wan, in a blog, has used the example of saliency maps and dec