论文部分内容阅读
二进神经网络中提取知识主要体现为对输入输出逻辑关系的提取,而逻辑关系的表达方式分为蕴含性规则和等价性规则,文中对比了蕴含性规则和等价性规则的差异;以KT方法为例,讨论了蕴含性规则在表达二进神经网络内在知识时,对某些具有明确逻辑意义的二进神经网络,并不是最清晰的表达方式。对这些逻辑关系,采用等价性规则可以简洁清晰地解决问题,所以对于二进神经网络神经元表达的逻辑关系建立可能的等价性规则提取方法是有意义的。CH判据是一种提取等价性规则的方法,但CH判据是充分性判据,对二进神经元的权系数有约束条件,因此不适用于任何学习算法的学习结果。为解决这些问题,文中研究了二进神经网络表达几类等价逻辑关系的充要性判据,并根据这些判据提出了提取等价性规则的WTA方法。在使用WTA方法时,必须预先对二进神经元进行必要的剪枝。文中证明了剪枝定理,并通过二个例子说明了用WTA方法进行规则提取的过程。
Binary neural network to extract knowledge is mainly reflected in the input and output of the logical relationship between the extraction, and the logic of the expression is divided into three kinds of rules and equivalence rules of implication, the article compares the difference between the implied rules and equivalence rules; KT method is taken as an example to discuss the implicit rules in the expression of the intrinsic knowledge of the binary neural network, some of the explicit logical meaning of the binary neural network is not the clearest way of expression. For these logical relations, we can solve the problem succinctly and clearly by using the equivalence rules. Therefore, it is significant to establish a possible equivalence rule extraction method for the logical relation of neuron expression in binary neural networks. CH criterion is a method of extracting equivalence rules. However, the CH criterion is a sufficient criterion, which has constraints on the weight coefficients of binary neurons and therefore is not suitable for the learning results of any learning algorithm. In order to solve these problems, this paper studies the binary neural network to express several kinds of equivalence logical relations of the necessary and sufficient criteria, and based on these criteria proposed WTA method to extract the equivalence rules. When using the WTA method, the necessary pruning of the binary neurons must be done in advance. The pruning theorem is proved in this paper, and the process of rule extraction by WTA method is illustrated by two examples.