Self-Organizing Neural Architectures for Vision, Learning, and Robotic Control

Abstract

Self-organizing neural architectures are described for the visual perception of static and moving forms, for autonomous real-time learning of multidimensional associative maps, and for adaptive control of variable-speed multi joint motor trajectories. Motion filtering and segmentation are carried out by a motion Boundary Contour System. Associative learning is accomplished by a Vector Associative Map, which provides an on-line alternative to the off-line properties of Back Propagation for error-based learning. Trajectory formation is carried out by a Vector Integration to Endpoint model.