NEURAL NETS AND TASK DECOMPOSITION

摘要: Publisher Summary When using neural networks (NN) for classification tasks, it might be useful to incorporate some knowledge about the phenomena into network architecture. The is decomposition of task simpler subtasks, each one being processed by its own NN module. There many reasons justifying such decomposition. First, impossible solve a complex unique technique, and secondly, sometimes possible decompose smaller easier solve. Modular connectionist architectures have been proposed. nets present deficiencies when different modules or units are used trained sequentially. A module failure could lead poor results whole architecture not optimum global task. It reasonable optimize training components together introduce system automatic error correction mechanism. This chapter presents modular architecture, which able identify modules' errors redirect data towards other modules: net revise final answer. had designed signal that cannot solved classifier. has tested on synthetic data, representative main problems found in original