A DISCRETE VERSION OF THE DYNAMIC LINK NETWORK

Learning in a dynamic link network (DLN) is a composition of two dynam ics: neural dynamics inside layers and link dynamics between layers. B ased upon a rigorous analysis of the neural dynamics, we find an algor ithm for selecting the parameters of the DLN in such a way that the ne ural dynamics preferentially converges to any chosen attractor. This c ontrol is important because the attractors of the neural dynamics dete rmine the link dynamics which is the main tool for pattern retrieval. Thus in terms of our constructive algorithm it is possible to explore the link dynamics using all kinds of attractors of the neural dynamics , In particular, we show how to get on-center activity patterns which have been extensively used in the application of the DLN to image reco gnition tasks as well as having an important role in the image process ing of the retina. We propose also a Hopfield-like discretized version of the neural dynamics which converges to the attractors much faster than the original DLN.