The emergence of generalization in networks with constrained representations

The authors introduce a constraint on intermediate representations which reduces the number of allowable solutions and leads to generalization for certain classes of problems. Specifically, they constrain the number of intermediate representations to be minimized during training. This representational constraint also defines a training algorithm for multilayered networks. They describe the class of problems for which the algorithm is well suited and discuss the performance of the algorithm with regard to several problems on two-layer networks.<>