Learning state space dynamics in recurrent networks

Fully recurrent (asymmetrical) networks can be used to learn temporal trajectories. The network is unfolded in time, and backpropagation is used to train the weights. The presence of recurrent connections creates internal states in the system which vary as a function of time. The resulting dynamics can provide interesting additional computing power but learning is made more difficult by the existence of internal memories. This study first exhibits the properties of recurrent networks in terms of convergence when the internal states of the system are unknown. A new energy functional is provided to change the weights of the units in order to the control the stability of the fixed points of the network's dynamics. The power of the resultant algorithm is illustrated with the simulation of a content addressable memory. Next, the more general case of time trajectories on a recurrent network is studied. An application is proposed in which trajectories are generated to draw letters as a function of an input. In another application of recurrent systems, a neural network certain temporal properties observed in human callosally sectioned brains. Finally the proposed algorithm for stabilizing dynamics around fixed points is extended to one for stabilizing dynamics around time trajectories. Its effects are illustrated on a network which generates Lisajous curves.