An Efficient Optimization Algorithm for Group Sparsity
By Mark Schmidt
This talk will introduce the 'group' variable selection problem, and motivate
looking at this problem in the context of simultaneously learning the parameters
and graph structure of conditional random fields for early detection of coronary
heart disease from multi-view ultrasound video.
We propose that the `spectral projected gradient' algorithm is a promising
approach for efficiently solving the group variable selection problem. This
constrained optimization method is extremely simple and very efficient, but
requires calculation of the projection operator on the constraint set. This talk
will outline how to efficiently compute this projection, leading to an efficient
optimization method for large-scale problems.