Explaining AI

Microsoft researchers announced human parity in key AI perception tasks, featuring ResNet's 96% ImageNet accuracy and Xiaoice's social chatbot that engages users in sessions averaging 23 turns. Beyond performance, the presentation highlighted commercial successes like a 40% coupon conversion rate in Japan while addressing critical challenges such as gender bias in word embeddings and the necessity of "Explainable AI" for high-stakes decision-making. Microsoft is currently employing both bottom-up and top-down methodologies to balance model accuracy with transparency, asserting a societal responsibility to ensure all AI systems remain understandable to humans.