Combining distributed and localist computations in real-time neural networks

In order to benefit from the advantages of localist coding, neural models t hat feature winner-take-all representations at the top level of a network h ierarchy must still solve the computational problems inherent in distribute d representations at the lower levels.