We consider the following pervasive scenario in contemporary scientific discovery: we are faced with an enormous space of candidate designs for some task we must explore. However, our ability to explore this space is severely limited due to the significant inherent costs of experimentation and/or simulation. Thus it is critical that we allocate our limited resources effectively. Bayesian experimental design is a statistical paradigm for adaptive experimental design that has established a niche in accelerating discovery in such scenarios. We will provide a high-level, intuitive introduction to Bayesian experimental design from and discuss one particular setting in depth: "active search," where we seek to discover rare, valuable points from a large space of alternatives. We will discuss the surprising difficulty of this problem in theory and introduce efficient, nonmyopic policies to solve it in practice, demonstrating its effectiveness on large-scale drug and materials discovery tasks.
Biography:
Professor Garnett's primary research interest is Bayesian active learning, with a focus on applications in the natural sciences and engineering. A major theme in his research is automating scientific discovery, broadly interpreted to include both theory and practice and both policy design and modeling. He has collaborated in multiple domains across the natural sciences, including astronomy, drug discovery, materials science, surface science, personalized medicine, and animal behavior. His long-term research vision is building fully automated, robust systems for active learning, to democratize machine learning and transform scientific practice in the 21st century.
Location: 373 Cory