论文部分内容阅读
受神经生理学和发育生物学中有关事实的启发,在BP网络学习过程中,将动物视觉系统中感受野的空间结构形式用于初始权值的设置,训练其完成图形识别问题。T-C问题的计算机实验结果表明,这种感受野型的初始化,不论是全局联系还是局域联系以及隐单元个数的多少,输入图形的尺度变换情况下,都比随机化权值的学习速度有显著的提高。
Inspired by the relevant facts in neurophysiology and developmental biology, in the process of BP network learning, the spatial structure of the receptive field in the animal visual system is used to set the initial weight, and training is performed to complete the problem of pattern recognition. The computer experiment results of TC show that the initialization of this receptive field type, regardless of global or local association and the number of hidden units, is much faster than the learning speed of randomized weights Improved significantly.