Sparse algorithms of Random Weight Networks and applications

This paper studies sparse algorithms for training Random Weight Networks (RWN) and their applications. The proposed algorithms contain three principal steps: initialization of networks structure, simplification of RWN structure based on sparse coding, and relearning process with renewed nodes. A key of the algorithms is sparse coding of hidden layer neurons by adding an initialization process to simplify the networks structure. Specially, the new algorithms, to some extent, can avoid the over-fitting phenomenon efficiently. As applications, the algorithms are used to diagnose the fault of switch reluctance motor (SRM) and to recognize the human face. Compared with the traditional back-propagation (BP) and RWN algorithms, the experimental results show that the proposed algorithms have effective performances on the accuracy or time. These methodologies can also be conceived as support tools for the practical fault diagnosis of SRM and the human face pattern recognition.