Electrical and Computer Engineering
Dean of Engineering
Faculty Host: Rich Baraniuk
Ohio State University
ECE Seminar Series: Vector Approximate Message Passing, Phil Schniter, Ohio State University (698/699)
Thursday, April 20, 2017
to 5:00 PM
1064 Duncan Hall
6100 Main St
Houston, Texas, USA
The approximate message passing (AMP) algorithm recently proposed by Donoho, Maleki, and Montanari is a computationally efficient iterative approach to sparse reconstruction and related problems.
AMP that has a remarkable property: for large i.i.d. sub-Gaussian measurement matrices, its per-iteration behavior is rigorously characterized by a scalar state-evolution whose fixed points, when unique, are Bayes optimal.
However, AMP is fragile in that even small deviations from the i.i.d. sub-Gaussian model can cause the algorithm to diverge.
In this talk, I will describe a ``vector AMP'' (VAMP) algorithm, which also has a rigorous scalar state-evolution.
VAMP's state-evolution, however, holds under a much broader class of large random measurement matrices, those that are right-rotationally invariant.
I will also describe a non-parametric version of VAMP that can cope with an unknown prior and/or likelihood, connections between VAMP and convex optimization algorithms, plug-and-play extensions of VAMP, and connections between VAMP and interpretable deep neural networks.
Biography of Phil Schniter:
Philip Schniter received the B.S. and M.S. degrees in Electrical Engineering from the University of Illinois at Urbana-Champaign in 1992 and 1993, and the Ph.D. degree in Electrical Engineering from Cornell University in Ithaca NY in 2000. From 1993-1996 he was a systems engineer at Tektronix Inc and since 2000 he has been a professor in the ECE department at The Ohio State University in Columbus, OH. For the 2016-17 academic year he is a visiting professor at Duke University. Dr. Schniter's areas of research include signal processing, wireless communications, and machine learning.