Heterogeneity and homogeneity in modular neural architecture

The author propose and enhance some models of module-based dynamical neural networks. This work demonstrates the effectiveness of a modular architecture focusing on associative memory tasks. In such a task, divided sub-patterns should be stored to the corresponding modules, so each module has different intra-module connections; the module structure for an associative memory task inevitably becomes heterogeneous. This paper focuses on the “heterogeneity” and “homogeneity” in modular neural architectures. We discuss their relationship and show that they produce qualitatively different kinds of dynamics, suitable for an associative memory task and an optimization task, respectively