Categorization in a layered neural network

The layered feedforward network of Domany, Meir and Kinzel is extended to investigate the ability to recognize any one of a macroscopic number of concepts (ancestors) when the network is trained through a finite number of examples (descendents) of each concept, by means of a generalized Hebbian learning rule between cells on two consecutive layers. Learning curves describing the generalization error are obtained, as well as phase diagrams and basins of attraction that exhibit the possible coexistence of a generalization phase, a retrieval phase of examples and a spin-glass phase.