Speaker: Dr. Dan Zhang（Bosch Center for Artificial Intelligence）
When: 2018-7-2 10:00am-11:00am
Where: Conference Room 1319, first Floor of A1 Building in CNV
Abstract:
Many signal processing problems can be formulated as the following: One aims at an estimate of a latent random variable given the realization of a statistically related observable random variable. It is a statistical inference task. If one has the access to the likelihood function and the prior density of the latent variable, then the inference task can be theoretically performed under the Bayesian framework. In many applications, exact inference however can be too complex to achieve. In such cases, we have to resort to some form of approximations. They generally fall into two classes, i.e., deterministic and stochastic approximations. In this talk, we focus on a family of deterministic approximations termed variational Bayesian inference. Our aim is to unify message passing algorithms for variational Bayesian inference under the framework of constrained Bethe free energy minimization, which originates from physics. In doing so, we obtain a mathematical framework to permit systematic designs and optimizations of algorithms for achieving a good comprise between the fidelity and tractability of variational Bayesian inference.
Short bio:
Dr. Dan Zhang is a research scientist at the Bosch Center for Artificial Intelligence, working on deep learning in particular from the Bayesian viewpoint. She received my PhD in electrical engineering from RWTH Aachen University. After that, she spent three years at TU Dresden, working as a postdoctoral researcher and also leading a research team at Vodafone chair mobile communications systems. During my PhD as well as her postdoctoral phase, my research focus was on investigation of approximate Bayesian inference techniques from theory to practice. The main application area was wireless communications. Combining deep learning with Bayesian techniques is an exciting research field. Since joining Bosch, she has been aiming at assessing the uncertainty in deep learning with the aid of Bayesian techniques. Her research interests include Deep Learning, Bayesian Inference, and Statistical Signal Processing.
