Recurrent Backpropagation Networks
In the stereo problem and in the traveling salesman problem, the dynamical parameters are hand-crafted so as to embed the problem constraints. Such an approach assumes that the constraints are known and that they can be easily represented in the dynamical system. This state of affairs, however, is certainly the exception rather than the rule; thus it is useful to have techniques for extract ing constraints from examples and embedding them in convergent dynamical systems. The Hebb rule is one rule for doing this. The purpose of this chapter is to describe a gradient technique for extracting information from examples and for embedding it in convergent dynamical systems. Before going into the details, however, it is useful to review some basic notions.