Quantized neuronal modeling: Quantum gate structure in Elman networks

This paper investigates the model of Elman network with quantum gate architecture and it´s online learning algorithm. The new neural structure, compared with the conventional Elman network (CEN), contains a quantum map layer which can be used for solving the pattern mismatch between the context layer and the input layer. A corresponding training algorithm for this new neural architecture is also presented, as opposed to the standard back-propagation (BP) learning algorithm for ENs. According to the new learning laws, the rotation parameter and the reversal parameter of the quantum gate are updated based on gradient-descent methods. The numerical experiment shows that the proposed network has better generalization performance and faster convergence speed than conventional Elman networks.