ABSTRACT
A neural network architecture combining an adaptive sensory-motor mapping and an online visual error correction (VEC) mechanism is proposed. The sensory-motor network (SMN) acquires a consistent intermodal mapping through primary circular reactions. During performance, SMN generates feed-forward control commands to the motor networks. This feed-forward command is augmented by VEC which is a feed-back control mechanism. The feedback allows the system to correct immediately during performance, errors that may arise from various sources including intermodal inconsistencies. This dual feed-forward feedback control enables the network to maintain the accuracy of its performance while inconsistencies are resolved through long-term adaptation. We present simulation results that illustrate the long-term learning as well as immediate correction capabilities of the network.
