Learning in systematically designed networks

The authors describe a network design methodology that is capable of specifying the structure of neural networks that are predisposed to satisfy the syntactic constraints of problems. This methodology is observed to produce networks that might also be vulnerable to certain unfeasible equilibria. However, a training strategy that is developed allows these systematically designed networks to learn from solutions that they autonomously develop. The result is an adaptation of the Hopfield model that can learn to generate only solutions to the original problem. As a consequence of the fact that the network has been predisposed to find solutions, learning can take place without producing a training set prior to the training session. Furthermore, training only occurs when a solution is found, and since the set of solutions is typically much smaller than the total number of possible states, training time is reduced.<>