Variational Bayesian inference for stochastic processes
- Datum: 13.10.2017
- Uhrzeit: 14:00 - 15:00
- Vortragender: Manfred Opper (Artificial Intelligence Group, TU Berlin)
- Ort: MPA
- Raum: Large Seminar Room E 0.11
- Gastgeber: MPA
Abstract: Variational methods provide tractable approximations to probabilistic and Bayesian inference for problems where exact inference is not tractable or Monte Carlo sampling approaches would be too time consuming. The method is highly popular in the field of machine learning and is based on replacing the exact posterior distribution by an approximation which belongs to a tractable family of distributions.The approximation is optimised by minimising the Kullback—Leibler divergence between the distributions. In this talk I will discuss applications of this method to inference problems for stochastic processes, where latent variables are very high- or infinite dimensional. I will illustrate this approach on three problems: 1) the estimation of hidden paths of stochastic differential equations (SDE) from discrete time observations, 2) the nonparametric estimation of the drift function of SDE and 3) the analysis of neural spike data using a dynamical Ising model.