Friday, May, 3, 2019, Invited lecture for the Austrian Council on Robotics and Artificial Intelligence (ACRAI).
Title: From Machine Learning to Explainable AI
Abstract: After a short introduction into the concepts of human-centered AI a brief introduction to probabilisitic machine learning – the workhorse of AI – is given, illustrated on a few examples from the medical domain. In narrowly defined tasks (automomous driving, game playing, image classification, etc.) deep learning even exceeds human performance. However, in certain cases it is necessary to disentangle the explanatory factors of the data, i.e. to understand the data in the context of a problem and to be able to explain why a certain machine decision has been made. Current state-of-the-art methods on explaining predictors include for example optimization and decomposition, i.e. relevance propagation and concept activation. Future human-AI interfaces need a mapping between explainability, which is the property of a computer, and causability, which is the property of a human.