Dynamic integration of modular neural network's sub-networks

This paper proposes a general definition for modular neural networks, which can encompass a large variety of modular neural networks in the literature. Issues rela ted to integration of modular neural networks are addressed, and rive integration methods are presented based on the principle of 'divide and conquer' and adaptive ensemble. The five methods are categorized according to (1) distance measurement (including absolute distance measurement and relative distance measurement). (2) number of modules to be integrated (i.e. some methods integrate all modules, while others integrate partial modules). (3) integration strategy (e.g. data-driven integration or data and knowledge-d riven integration). Simulation results demonstrate effectiveness of the methods.