Race networks: a theory of competitive recognition networks based on the rate of reactivation of neurons in cortical columns

A theory of noniterative neural processing is presented for use in recognition networks performing competitive pattern classification. This scheme has advantages over the more iterative approaches in terms of processing speed, simplicity, and stability while learning. The theory derives from an examination of the operation of the simple cells of the primary visual cortex. A competitive pattern classification system consists of a set of feature detectors, each tuned to respond to a specific type of pattern. The collection of feature detectors classifies a pattern by finding the detectors that respond most strongly to the pattern. In race networks this is achieved by setting up a race between feature detectors to overcome their inhibition, as opposed to the more conventional method of the competitive feedback of activation strengths. At the start of each processing cycle all recognition cells receive a potent dose of inhibition. Stimulus signals are applied to the cells to reactivate them. A cell responds best to the pattern it is tuned to detect. The first cell to overcome its inhibition signals a response and triggers a wave of inhibition that restarts the cycle.<>