Séminaire ICI : Pablo M. Olmos

Titre du séminaire et orateur

Discrete Inference via Expectation Consistency Approximations. An application to symbol detection in massive MIMO.

Pablo M. Olmos, Universidad Carlos III de Madrid.

Date et lieu

Lundi 10 juillet 2017, 11h

ENSEA, Cergy Pontoise, salle 384


Discrete Inference emerges in many scenarios of interest in Signal Processing and Digital Communications. In this talk, we will review an approximate inference framework first presented by Manfred Opper and Ole Winther in 2015, known as Expectation Consistency (EC) approximate inference. EC describes the inference problem as the search of a stationary point of an approximation to the free energy of the probability distribution, where any stationary point satisfies a moment matching condition between two different approximations of the true distribution. We develop the EC approximate inference for the problem of probabilistic symbol detection in massive MIMO scenarios, where the goal is to perform approximate marginal inference in a high-dimensional probability mass function. We discuss several methods to optimize the EC free energy function and show their impact in the final performance. Finally, we discuss possible extensions to other problems arising in communications, such as channel decoding.