Difference in Learning ability of Neural Nets and Logical Expressions

本研究的目的是识别神经网络与逻辑表达式在人工智能学习能力上的区别。首先分析了神经网络检测小于脉冲持续时间的时间差能力的来源。提出了实现上述检测的并行处理机制，并通过计算机仿真进行了验证。神经网络必须具有生成各种不同序列的自学习能力和从这些序列中识别出合适序列的能力。分析了这些能力产生的机理，并提出了一个神经网络模型。在实际环境中对该模型进行了计算机仿真，结果表明该模型对于序列的生成和特定序列的选择是足够的。为了在不执行网络的情况下识别网络的可达性集，提出了一类典型的逻辑类型的Petri网。换句话说，从网络的结构和初始状态来确定网络的可达性集。结合这两种理解，我们清楚地认识到时间可微性是神经网络优于逻辑表达式的最关键因素。

The purpose of this study is to identify the difference between the neural nets and the logical expressions in the learning ability for the artificial intelligence.The origin of capability of detecting the time difference smaller than the pulse duration used in the neural nets is first analyzed. The parallel processing mechanism which enables the above detection is proposed and verified by the computer simulation. It is necessary for the neural nets to have selflearning capability to generate wide variety of different sequences and capability of identifying the adequate sequence from the sequences. The mechanism for these capability is analyzed and proposed as a neural net model. The computer simulation of this model in the realistic environment shows that the proposed model is sufficient for generating the sequences and selecting a particular sequence.A class of Petri nets, which is a typical expression of the logical type, is proposed for the reachability set of the net is to be identified without carrying out the execution of the nets. In other words, the reachability set is identified for the class of the nets from the structure of the nets and the initial states.Combination of these two understandings makes us clear that the time differentiability is the most crucial to give advantage of the neural nets over the logical expressions.