Time sequential pattern transformation and attractors of recurrent neural networks

For better understanding the function of recurrent neural networks (RNN), we propose that an external input is considered as one period of an oscillatory input. It follows from this that an external time sequential input corresponds to an attractor in a vector field. We show experimentally that RNN can learn: (1) input-output time sequential patterns as trajectories of attractors, and (2) transition between attractors.