Mathias Müller: Efficient Online Variational Estimation via Monte Carlo Sampling

Abstract: This talk addresses online variational estimation in parametric state-space models. We present a recursive algorithm for efficiently computing the evidence lower bound and its gradients as observations arrive sequentially. Based on i.i.d. Monte Carlo sampling and a structured variational family, the method jointly learns model parameters and approximates the posterior distribution of the latent states. We discuss its theoretical motivation through an asymptotic contrast function and Markov chain ergodicity, and illustrate its performance with numerical experiments.