【摘要】 福岛邦彦的神经认知机 (Neocognitron)能用于有变形或位移的模式识别 .然而 ,在原始神经认知机中许多参数及训练模式都是凭经验设定的 .该文旨在系统地剖析神经认知机的工作机制并有效地将进化计算结合进来以提高其性能 .首先 ,通过分析神经认知机的学习机制指出原始神经认知机忽略了训练模式间的相关性分析 ;然后 ,结合协作进化为神经认知机搜索合理的参数和训练模式 .实验结果表明了该文方法的有效性 .因为参数和训练模式是由进化计算获得而非人为设定的 ,所以这种改进型神经认知机有更广阔的应用领域更多还原

【Abstract】 Fukushima’s Neocognitron has the ability to recognize visual patterns even if they are distorted or shifted. However, many parameters as well as the training patterns are designed empirically in the original Neocognitron. Evolutionary computation will be incorporated into Neocognitron in this paper. First, by analyzing the learning mechanism of Neocognitron, we point out the correlation amongst the training patterns was ignored in the original Neocognitron. And then, the cooperative coevolution is incorporated to search reasonable parameters and training patterns for Neocognitron. Experimental results show this proposed methodology effective and efficient. Because the parameters and the training patterns are acquired by evolutionary computation rather than domain expert, such an advanced Neocognitron can be applied to a number of application problems.更多还原