Synthesis of neural networks with linear programs

The two critical steps of choosing an initial neural network configuration and then choosing a collection of network weights so that the network best approximates a given training set can be formulated as a linear program. The inequalities necessary to construct the linear program are subsets of Boolean symmetric functions, naturally implementable with threshold logic devices. The construction bypasses problems with local minima with current training algorithms. The training process becomes a linear programming problem whose solution provides the sought approximation. The construction also provides clear methods of trading the final approximation precision provided by the network with the computing times needed to obtain and use it