Evolving Large-Scale Modular Neural Networks

Modularity is a major feature of biological central nervous systems. For ex-ample, the human/primate cerebral cortex is composed of dozens of structurally and functionally identiable regions that are interconnected in a hierarchical network [1]. Motivated by this, our research group is studying the evolution and self-organization of modular neural networks. We are interested both in ex-plaining the origin of modularity in biological systems, and in understanding the fundamental principles that govern the emergence of modular neural networks in general. There has been a substantial amount of past research in this area dur-ing the last decade, examining how modularity in neural networks evolves (e.g., [2{4]). Our own group initially focused on using genetic algorithms and multi-objective evolutionary optimization methods to evolve module parameters that in uence acquisition of functionality in predened modules during learning [5, 6]. This past work was directed at explaining left-right asymmetries and hemi-spheric specialization in the brain. In order to go beyond such issues and address the origin of neural modular