Training autoassociative recurrent neural network with preprocessed training data
Arun Maskara, Andrew S. Noetzel · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1993
The Auto-Associative Recurrent Network (AARN), a modified version of the Simple Recurrent Network (SRN) can be trained to behave as recognizer of a language generated by a regular grammar. The network is trained successfully on an unbounded number of sequences of the language, generated randomly from the Finite State Automation (FSA) of the language. But the training algorithm fails when training is restricted to a fixed finite set of examples. Here, we present a new algorithm for training the AARN from a finite set of language examples. A tree is constructed by preprocessing the training data. The AARN is trained with sequences generated randomly from the tree. The results of the simulations experiments are discussed.