Machine Learning Seminar
Title:
Robust Compressive Sensing using Vector Approximate Message Passing
Abstract:
We consider the problem of estimating a structured coefficient vector "x" from noisy linear observations "y=Ax+w," often referred to as compressive sensing in ECE, linear regression in statistics, or linear inverse problems in medical imaging.
Recently, approximate message passing (AMP) has been proposed as a fast an accurate solution when the measurement matrix A is a typical realization of iid Gaussian random matrix. But AMP is not applicable to generic matrices. We thus propose a "vector AMP" (VAMP) algorithm that applies to a much larger class of matrices while maintaining the key benefits of AMP: fast and accurate inference with a rigorous state-evolution formalism that agrees with the replica prediction. We also describe VAMP extensions covering generalized linear models and non-parametric regression (i.e., model-free inference).
This talk represents joint work with Sundeep Rangan (NYU) and Alyson Fletcher (UCLA).