Temporal-winner-take-all networks for arbitrary selection in connectionist and neural networks

Summary form only given. It is commonplace in connectionist/neural-net models to include winner-take-all (WTA) networks for selecting a single, most active subnetwork out of a set of competing, interconnected subnetworks. However, there is also a need for subnetwork selection, that is on a totally arbitrary basis. To implement arbitrary selection the authors propose temporal winner-take-all (TWTA) networks. A TWTA network includes an arbiter that chooses a winner on the basis of the arrival times of signals from the contending subnetworks. The differences in the arrival times can be due to quasi-random, intuitively meaningless differences between the generation times of the signals and/or the travel times of different signals on their way to the arbiter. The process of arbitrary selection generally involves several repeated attempts at choosing a winner. A probabilistic analysis shows that a winner is almost certain to emerge after a few tries at most, under reasonable settings of parameters. Also, a preliminary version of a developing connection machine simulation of the proposed TWTA network has yielded promising results.<>