Laterally connected lobe component analysis: Precision and topography

Abstract

Due to the pressure of evolution, the brains of organisms need to self-organize at different scales during different developmental stages. In early stages, the brain must organize globally (e.g., large cortical areas) to form “smooth” topographic representation that is critical for superior generalization with its limited connections. At later stages, the brain must fine tune its microstructures of representation for “precision” - high-level performance and specialization. But smoothness and precision are two conflicting criteria. The self-organizing map (SOM) mechanisms of self-organization through isotropic updating and other published computational variants have dealt with global to local smoothing and lateral adaptation, but we show in our work that they are insufficient to deliver superior performance. In this paper, we introduce a combination of several mechanisms that, together, address these two conflicting criteria: lateral excitation through adaptive connections, explicit adaptive top-down connections (attention), dually-optimal lobe component analysis (LCA) for synaptic updating, simulated lateral inhibition through winners-take-all, and a developmental schedule that sets the number of winners, which decreases over time. Major performance improvements due to the combination of these mechanisms are shown in the reported experiments.