Using Adaptive Sequential Neurocontrol for Efficient Learning of Translation and Rotation Invariance

A system is described which uses system realization and `gradient descent through frozen model networks' for learning the sequential generation of fovea trajectories such that the nal position of a moving fovea corresponds to a target in a visual scene. The target may be arbitrarily rotated and translated, and it might even move. No teacher provides the desired activations of `eye-muscles' at various times. The only goal information is the desired nal input corresponding to the target. The task involves a complex temporal credit assignment problem and an attention shifting problem.