AI We Can Trust: Controlling Generative AI to Ensure Reliable Creation

Generative artificial intelligence (AI) is now starting to play a central role in human and computer creativity, with practical applications ranging from the generation of art, language, and image caption, to the discovery of new drugs and next-generation materials. However, reservations about safety and reliability have limited widespread and autonomous deployment of these AI systems. This talk presents a framework of end-to-end learning for control in a way that is data-efficient and reliable, illustrated by a suite of controllable generative models. We demonstrate the effectiveness of our models by subjecting them to the synthesis of creative images, sentiment-consistent sentences, and novel antibiotics/SARS-CoV-2 inhibitors.