Competitive decomposition of input space in a competitive modular multinet system

A competitive modular multinet structure is introduced as a local learning framework. Here, there is not a control switching mechanism between modules. Instead, the modules are encouraged to specialize in sub-regions of feature space competitively. In my decomposition scheme, the subregions are created, developed, shrunk or vanished during learning process, based on an interaction in a pool of networks. Just after specialization of networks in certain sub-regions, a selector is trained to learn the mapping between sub-regions and experts, which helps the multinet system to be used over test set. In my model, there is a balance between quantity as well as learning capacity of networks and complexity of feature space. Furthermore, task simplification and decision-making efficiency are both achieved. I also show that my model benefits from a smooth transition in the boundary of neighbor experts that improves the performance at boundary patterns. The proposed method is used for a regression toy problem as well as a Persian handwritten digit recognition task. The results of a comparative study reveal the superior performance-to-cost ratio of the proposed method.