Guest
A Google TechTalk, presented by Anastasia Makarova, 2022/08/23 Google BayesOpt Speaker Series - ABSTRACT: Black-box optimization tasks frequently arise in high-stakes applications such as material discovery or hyperparameter tuning of complex systems. In many of these applications, there is often a trade-off between achieving high utility and minimizing risk. Imagine, you use Bayesian optimization for two different problems, first — drug discovery affecting human lives, and second — tuning hyperparameters of an anti-fraud model. The testing setup, the notion of risk, budget constraints, and the price of a mistake are clearly different there, and modern learning methods need to handle these aspects. In this talk, I will discuss the trade-off between exploration (learning about uncertain actions), exploitation (choosing actions that lead to high gains), and risk (avoiding unreliable actions) in heteroscedastic Bayesian optimization, its theoretical and practical aspects. We will further look at the practical question of stopping conditions for BO, and its effect on the solution cost vs quality in the case of hyperparameter optimization. Finally, (if time allows) I am open to discussing ideas on using the domain structure (e.g., dependencies between variables) in such sequential decision-making setups. About the speaker: Anastasiia Makarova is a final-year PhD student at ETH Zurich under the supervision of Andreas Krause. She is also a fellow of the ETH AI Center and NCCR Automation. Her research is on sequential decision-making under uncertainty and representation learning for structured data, such as point clouds or graphs. The results of her PhD were published at NeurIPS / ICCV / IJCAI and contributed to Bayesian optimization, computer vision, and representation learning. She has recently received the best paper award for her work on Bayesian optimization for AutoML.
