Faster learning through a probabilistic approximation algorithm

The author proves that the learning problem in connections of networks is NP-complete, i.e. no polynomial-time algorithm exists which will correctly modify connection weights of a neural network. Although no perfect algorithm exists, a method called the probabilistic approximation algorithm is presented. This method, which can be used with any learning rule, would allow network designers to build networks with a predetermined probability of certain kind of error. He shows that for any learning rule that does not utilize probabilistic approximation, the probability of convergence will increase when the approximation method is employed.<>