Teaching network connectivity using simulated annealing on a massively parallel processor

A simulated annealing technique for automatically training a machine vision system to recognize and locate complex objects is described. In this method, the training is used to find an optimum connectivity pattern of a fixed number of inputs that have fixed weights, rather than the usual technique of finding the optimum weights for a fixed connectivity. The recognition model uses a two-layer artificial neural network, where the first layer consists of image edge vectors in four directions. Each neuron in the second layer has a fixed number of connections that connect only to those first layer edges that are best for distinguishing the object from a confusing background. Simulated annealing is used to find the best parameters for defining edges in the first layer, as well as the pattern of connections from the first to the second layer. Weights of the connections are either plus or minus one, so that multiplications are avoided, and the system speed is considerably enhanced. In industrial applications on a low-cost parallel SIMD (single instruction multiple data) architecture, objects can be trained by an unskilled user in less than 1 min, and after training, parts can be located in about 100 ms. This method has been found to work very well on integrated circuit patterns