I am a senior researcher at Microsoft Research AI4Science where I work on machine learning methods for chemistry with Frank Noé and Jan Hermann, focusing on QMC. I also have a strong interest in Bayesian experimental design and active learning. I am driven by the desire to understand how machine learning can help us to solve critical problems in the sciences and to build new, sustainable technology. Previously, I did my PhD in Statistical Machine Learning at the University of Oxford, supervised by Yee Whye Teh and Tom Rainforth in the Computational Stats and Machine Learning Group in the Department of Statistics. Before starting my PhD, I studied mathematics at Cambridge where my Director of Studies was Julia Gog.
A large part of my PhD work was on Bayesian experimental design: how do we design experiments that will be most informative about the process being investigated? One approach is to optimize the Expected Information Gain (EIG), which can be seen as a mutual information, over the space of possible designs. In my work, I have developed variational methods to estimate the EIG, stochastic gradient methods to optimize over designs, and how to obtain unbiased gradient estimators of EIG. In more recent work, we have studied policies that can choose a sequence of designs automatically. This talk offers a 30 minute introduction to experimental design and my research in this area. To use Bayesian experimental design in practice, we have developed a range of tools in deep probabilistic programming language Pyro: our aim is to allow automatic experimental design for any Pyro model.
Since EIG is a mutual information, I am also interested in the intersection between information theory and machine learning. This led me to study contrastive representation learning and the role of invariance in these methods, as well as reproducing SimCLR in PyTorch.