| One way to improve the capability of the neural network is to incorporate human knowledge. Knowledge-based neural networks can utilize both the knowledge from human experts and the knowledge inherent in the data. Neural networks mapped from a set of production rules are referred to as rule-based neural networks. Fuzziness and context are two distinct aspects of the knowledge of human experts. However, previous research on rule-based neural networks did not consider these two important aspects. Thus, a fuzzification layer is proposed to be added to the rule-based neural network for encoding the fuzziness of continuous input variables. The fuzzification layer can be superimposed onto any back-propagation networks. Moreover, a neural model with an adaptive memory structure follows the rule-based neural network in order to perform context processing. The adaptive memory is an extension of the gamma model (a neural model for temporal processing) for learning both the right and left contexts. A hybrid information processing system is constructed based on the above-mentioned neural models, and is evaluated on the electroencephalogram (EEG) sleep staging problem. Different experiments are designed to test the system. Compared with the previous results on the same sleep records, the present results demonstrate the following advantages. First, knowledge-based neural networks out-perform both symbolic knowledge-based systems and traditional neural networks. Secondly, fuzzy encoding of continuous data gives better performance than crispy data representation. Thirdly, context processing effectively enhances the results of information processing. The developed system is promising for sleep staging, but the design principles are generally applicable. |
