Educating a neural network for image processing

A preliminary series of experiments is reported in which a neural-network-based window filter was trained to perform complicated image processing tasks. It is shown that the analogy to training a human to perform a complex task is a useful one in developing a training strategy. In particular, four methods of improving learning were investigated. These are: pre-training on simpler but similar tasks, learning to perform useful sub-tasks, use of heuristic rules and structuring of training examples to make the required operation explicit