A nonexclusive task decomposition method for modular neural networks

Modular neural networks (MNNs) architectures have been developed aiming to outperform single neural nets. One of the main drawbacks in the construction of the MNNs is the task decomposition which consists in divide the problem into simpler sub-problems. This paper proposes a novel task decomposition method in which the classes of the problem can be divided redundantly. Thus, two different expert modules can have the same class. This is specially interesting for problems that have multimodal classes. The proposed MNN, called Redundant Pattern Distributor, is compared with other ones over many databases and the results show its effectiveness.