Learning Halfspaces Under Log-Concave Densities: Polynomial Approximations and Moment Matching

We give the first polynomial-time algorithm for agnostically learning any function of a constant number of halfspaces with respect to any log-concave distribution (for any constant accuracy parameter)