DYNAMICS AND NEURAL NETWORK COMPUTATION

Computation is carried out by the motion of the state of a dynamical system. This description unites digital and "neural network" computation. The use of an energy function to understand and to program batch mode computation on feedback networks is reviewed. Recognizing words in continuous speech illustrates the use of a compound network to solve a real-world problem where data arrive continuously over time. A new tool, a Lagrangian formulation, is introduced for studying networks which oscillate, and as a means for formulating problems which need to optimize functional integrals.