Abstract: A fully recurrent neural network with a nonmonotonic activation function that treats temporal sequences without expanding them into spatial patterns is described. This network associates a complex spatiotemporal pattern with a simple one using trajectory attractors formed by simple learning. Computer simulations show that the model not only has high recognition and generation abilities but can also perform advanced processing using bidirectional interactions.
Keywords: Temporal sequence, Bidirectional processing, Recognition and generation, Nonmonotone neural network