In this talk, we will discuss a family of approximate Bayesian methods, known as Expectation-Propagation (EP) to perform Bayesian inference. In a similar fashion to variational Bayes methods, EP can be used to approximate complex distributions using simpler densities, whose moments can be computed efficiently and exploited more efficiently in high-dimensional settings. First, we will review EP principles through small scale linear inverse problems which allow comparisons with sampling approaches. We will then discuss how EP’s factor graphs can be used for image restoration problems, where the noise may not be Gaussian and where convex and non-convex image priors are used. Finally, we will discuss how EP can be incorporated within more complex inference problems such as blind deconvolution.
Bio: Dr Altmann received the Eng. degree in electrical engineering from the ENSEEIHT, Toulouse, France, and the M.Sc. degree in Signal Processing from the National Polytechnic Institute of Toulouse, both in June 2010. He completed his Ph.D. within the Signal and Communication Group of the IRIT Laboratory and received the Ph.D. degree from the INP Toulouse in 2013. In 2014, he was awarded a postdoctoral Fellowship by the Direction Générale de l’Armement (DGA, French Ministry of Defence) and joined Heriot-Watt University to develop computational methods for hyperspectral imaging and Lidar-based ranging applications. From 2015 until 2016, he was a postdoctoral researcher within the Institute of Sensors, Signals and Systems at Heriot-Watt University. Since February 2017, he has been an Assistant Professor at the School of Engineering and Physical Sciences, Heriot-Watt University. In 2017, he was also awarded a 5-year Research Fellowship by the Royal Academy of Engineering to develop new computational methods for low-illumination imaging and sensing. His current research focuses on sampling and approximate Bayesian methods for 2D/3D imaging problems.