Using localizing learning to improve supervised learning algorithms

Slow learning of neural-network function approximators can frequently be at tributed to interference, which occurs when learning in one area of the inp ut space causes unlearning in another area. To mitigate the effect of unlea rning, this paper develops an algorithm that adjusts the weights of an arbi trary, nonlinearly parameterized network such that the potential for future interference during learning is reduced. This is accomplished by the reduc tion of a biobjective cost function that combines the approximation error a nd a term that measures interference. Analysis of the algorithm's convergen ce properties shows that learning with this algorithm reduces future unlear ning. The algorithm can be used either during on-line learning or can be us ed to condition a network to have immunity from interference during a futur e learning stage. A simple example demonstrates how interference manifests itself in a network and how less interference can lead to more efficient le arning. Simulations demonstrate how this new learning algorithm speeds trai ning in various situations due to the extra cost function term.