For many scientific applications the forward model consists of an expensive black-box simulator without easy access to gradients of the forward model. In such a situation, inference can be challenging, with standard gradient-free sampling algorithms requiring an intractably large number of serial model evaluations. In this talk I will introduce the flow annealed Kalman filter for approximate Bayesian inference. This is a generalization of the classical ensemble Kalman filter, that inherits its rapid convergence properties (typically ~10 iterations), whilst relaxing the Gaussian ansatz that limits its applicability to non-Gaussian posterior geometries. I will demonstrate the performance of the method on a number of challenging benchmarks and discuss future integration with pocoMC, a library for normalizing flow accelerated cosmological inference tasks.
BIO
Richard Grumitt is a Shui Mu Fellow at Tsinghua University. He completed his undergraduate and DPhil degrees at the University of Oxford, graduating with his DPhil in 2020. Whilst at Oxford he worked with Prof. Angela Taylor, Prof. Mike Jones and Prof. David Alonso on component separation for cosmic microwave background studies. He joined Tsinghua University in 2021, where he works with Prof. Yi Mao and Prof. Uroš Seljak on the development of novel statistical methodology for scalable Bayesian inference, and the application of these methods to inference problems in cosmology.
