Task decomposition through competition in a modular connectionist architecture - the what and where vision tasks (01 Jan 1991)
This article studies the question of how different components of a computational architecture can develop different functional specializations. It describes a novel architecture consisting of multiple neural networks that compete for the right to learn each data item. Given an input pattern, the network whose output is closest to the target output is allowed to do the most learning on that item. Other networks learn little or nothing about the item. The tendency of the architecture is to partition the set of data items so that different networks learn different items and, thus, acquire different functions. Sample simulations apply the architecture to the identification and localization of an object depicted in a visual image. A lesson of this work is that modularity or at least functional specializations, need not be determined soley by genetic factors. Instead, learning may also play a role in the functional organization of a modular system.
Article URL: http://cognitrn.psych.indiana.edu/rgoldsto/cogsci/Jacobs.pdf