A unified neural network model for the self-organization of topographic receptive fields an.pdf

A Unied Neural Network Model for the Self-organization ofTopographic Receptive Fields and Lateral InteractionJoseph Sirosh and Risto MiikkulainenDepartment of Computer SciencesThe University of Texas at Austin, Austin, TX 78712email: sirosh,risto@cs.utexas.eduTechnical Report AI94-213January 1994AbstractA self-organizing neural network model for the simultaneous development of topographic receptive eldsand lateral interactions in cortical maps is presented. Both aerent and lateral connections adapt by thesame Hebbian mechanism in a purely local and unsupervised learning process. Aerent input weights of eachneuron self-organize into hill-shaped proles, receptive elds organize topographically across the network, andunique lateral interaction proles develop for each neuron. The resulting self-organized structure remains ina dynamic and continuously-adapting equilibrium with the input. The model can be seen as a generalizationof previous self-organizing models of the visual cortex, and provides a general computational frameworkfor experiments on receptive eld development and cortical plasticity. The model also serves to point outgeneral limits on activity-dependent self-organization: when multiple inputs are presented simultaneously,the receptive eld centers need to be initially ordered for stable self-organization to occur.1 IntroductionThe response properties of neurons in many sensory cortical areas are ordered topographically, i.e. nearbyneurons respond to nearby areas of the receptor surface. Such topographic maps form by the self-organizationof aerent connections to the cortex, driven by external input (Hubel and Wiesel 1965; Miller et al. 1989;Stryker et al. 1988; von der Malsburg 1973). Several neural network models (Amari 1980; Kohonen 1982;Miikkulainen 1991; Willshaw and von der Malsburg 1976) have demonstrated how the global topographicorder can emerge from local cooperative and competitive lateral interactions within the cortex. Such modelsare base