Generalization in Learning Multiple Temporal Patterns Using RNNPB

This paper examines the generalization capability in learning multiple temporal patterns by the recurrent neural network with parametric bias (RNNPB). Our simulation experiments indicated that the RNNPB can learn multiple patterns as generalized by extracting relational structures shared among the training patterns. It was, however, shown that such generalizations cannot be achieved when the relational structures are complex. Our analysis clarified that the qualitative differences appear in the self-organized internal structures of the network between generalized cases and not-generalized ones.