Neural-network learning and Mark Twain's cat

摘要

In current practice in the engineering community, neural networks are used as only one useful class of adaptive pattern recognizer. Neural networks, however, are far more than devices that can learn accurate input-output transformation or form good category boundaries for pattern classifiers. They are a new form of computer, good at some unfamiliar problems, but quite poor at some familiar ones. An application involving a neural network learning some elementary arithmetic is discussed. It is shown that a simple network program can be implemented by differential weighting of the input data vector. In favorable cases the programming vector can be estimated by seeing relatively few examples of the output, if the task and the structure of the data allow it. Therefore, easy programming is allowed in only a limited domain, controlled by the data representation.