Always Be Learning

Derek Hoiem

University of Illinois at Urbana-Champaign, USA

Abstract

We typically think of machine learning as a stage, after which a static model is deployed and applied. How can we make AI systems that continually learn and improve? I will juxtapose machine learning with aspects of human learning and memory, particularly highlighting human capabilities that we cannot yet replicate. I'll talk about multiple strategies for online and continual learning of single models, including regularization, rehearsal, multi-modal representations, prompt tuning, as well as strategies to coordinate communities of models.