Neocognitron学习算法分析

现有的对Ｎｅｏｃｏｇｎｉｔｒｏｎ的分析都采用代数法，因而无法研究它的动态特性．本文把Ｎｅｏｃｏｇｎｉｔｒｏｎ及其学习算法推广到连续时域，借助微分方程来研究Ｎｅｏｃｏｇｎｉｔｒｏｎ．文中给出了无教师学习算法一般情况下，学习过程中Ｕｓ层神经元输出变化规律的微分方程，指出其增加的条件，并推出权ａ、ｂ初始值选择的一个必要条件；进一步得出无教师算法代表稳定后和有教师学习情况下Ｕｓ变化的一种等效显式函数，指出此时学习过程是Ｕｓ层神经元输出向一个系数的逼近过程，且有教师学习过程的最后状态与可变权初值和学习率无关．并讨论了影响Ｕｓ终值和逼近速度的因素．

ON THE LEARNING PROCESS OF THE NEOCOGNITRON

xisting analyses of the Neocognitron failed to discuss the dynamic characteristic during learning because they were all confined to algebraic method. This paper intruduces differentiation to analyse the Neocognitron. For the unsupervised learning, this paper derives a defferential equation to describe us, a condition of us increasing, and an inequality that the initial value of variable weight of a, b must satisfy. For fixed representation or the supervised learning, we obtain an u. function of time. We show that, in this case, learning is a process in which us approximates a coefficience independence of q and the initial value of weight a,b.