Modulating surrogates for bayesian optimization

Abstract

Probabilistic numerics provides a narrative to extend our traditional approach of uncertainty about data to uncertainty about computations. In this talk I will provide a brief background to Probabilistic numerics and show why it is beneficial to think about Bayesian optimisation and other active learning techniques within this framework. In specific I will show how these guiding principles have allowed us to formulate a set of surrogate models that allows us to focus on details that are informative for search while ignoring detrimental structures that are challenging to model from few observations. We will show how this leads to a significantly more robust and efficient active learning loop.

Notes

• The talk is primarily based on Bodin, E., Kaiser, M., Kazlauskaite, I., Dai, Z., Campbell, N. D. F., & Ek, C. H., Modulating surrogates for bayesian optimization., In , Proceedings of the 37th International Conference on Machine Learning, {ICML} 2019, 12-18 July 2020, Virtual (pp. ) (2020).

• Dr. Carl Henrik Ek is a senior lecturer at the University of
  Cambridge. Together with Prof. Neil Lawrence, Jessica
  Montgomery and Dr. Ferenc Huzar he leads the newly formed machine learning research group in the Cambridge computer lab. He is interested in building models that allows for principled treatment of uncertainty that provides interpretable handles to introduce strong prior knowledge. His website can be found here.