On the localization of feedforward networks

Interference in neural networks occurs when learning in one area of the input space causes unlearning in another area. Networks that are less susceptible to interference are called spatially local networks. These networks are often used in neurocontrol, in online applications, where, because of the real time nature of the task, interference is often a problem. Although there are heuristics as to what makes a network local, there is no theoretical framework for measuring localization. This paper provides a formal definition of interference and localization that will allow measurement of a network's local properties. These definitions will be useful in developing learning algorithms that make networks more local. This may lead to faster learning over the entire input domain.