Learning Algorithms for Enclosing Points in Bregmanian Spheres

摘要: We discuss the problem of finding a generalized sphere that encloses points originating from single source. The contained in such are within maximal divergence center point. divergences we study known as Bregman which include special case both Euclidean distance and relative entropy. cast learning task an optimization show it results simple dual form has interesting algebraic properties. then general algorithmic framework to solve problem. Our training algorithm employs auxiliary function bounds dual’s objective can be used with broad class functions. As specific application give detailed derivation for analyze generalization ability by adopting margin-style proof techniques. also describe two schemes online algorithms when radius is set advance.