Convex Two-Layer Modeling

This work develops a convex relaxation of hidden-layer conditional models that admits global training that is able to acquire two-layer models that cannot be represented by any single-layer model over the same features, while improving training quality over local heuristics.