Using constitutive information in neural networks for control

It is pointed out that knowledge of the task to be performed by a neural network can often be used to design the network. As illustrated by elementary examples, constitutive information can be used to construct a network from several subnetworks. Because the subnetworks are simpler, they reduce training time. The structuring also reveals where nonneural calculations can be used to make more effective hybrid systems. Network structure also may be a first step toward validation