Andrew Jacobsen

 

 

Abstract: As machine learning systems become increasingly intertwined with our day-to-day lives, it becomes increasingly important that machine learning algorithms satisfy strong guarantees on their performance. Yet despite a wealth of sophisticated machine learning theory, performance guarantees often fail to meaningfully transfer to real-world problems. In this talk, I will present some recent perspectives from Online Convex Optimization that aim to address some of these gaps between theory and practice, by providing algorithms that require little-to-no prior knowledge or assumptions to achieve their performance guarantees (hyperparameter-free algorithms). A key focus of the talk will to develop an intuitive understanding of a natural connection between hyperparameter-free learning and the problem of adapting to non-stationary data.