A common problem with massively parallel networks is crosstalk: unwanted mutual interference associated with distributed representations in connectionist networks. Many connectionist models have been constructed, but it has turned out to be very difficult to avoid crosstalk, especially in large scale models. In this paper, several current techniques to eliminate crosstalk are surveyed and a new sequential filtering approach is described which eliminates crosstalk in parallel connectionist networks. The new method combines the feature-integration theory of attention and the theory of connectionist systems. Two kinds of filtering techniques, attention focusing and WTA (winner-take-all) subnetworks, are applied to index features in a set of individuals. The system is implemented on the Rochester Connectionist Simulator.
