ABSTRACT
Winner-takes-all (WTA) behavior in a competitive system has been central to theories of computational map formation in the brain because WTA behavior makes it possible for a distributed system to make choices. Competitive models have generally relied on direct lateral inhibition, which requires a very high proportion of inhibitory connections, whereas the cerebral cortex is dominated by excitatory connections. This article introduces new theoretical analysis of a competitive model proposed by [Reggia, 1985] as an alternative to direct lateral inhibition. This model, based on competitive activation with virtual lateral inhibition (CAVLI), is seen to be based on iterative Bayesian updating, with architectural outlines that closely resemble ART1, but with no bottom-up weights and with rational (multiplicative) error adjustment. CAVLI and ART1 are compared with respect to inhibitory mechanisms, biological plausibility, and scaling.
