前馈网络的一种线性化快速学习算法

针对前馈神经网络的反向传播（ＢＰ）学习算法收敛速度慢的缺点，提出了一种新的学习算法即线性化快速学习算法。这种学习算法在神经网络学习的初期，采用标准ＢＰ学习算法。而当神经网络接近最优点时，由于此时其连接权重调节幅度很小，因此采用对各层神经元的非线性作用函数进行泰勒级数展开，并取其一阶展开式近似逼近原函数，从而使其非线性作用函数转化为线性作用函数，简化了网络学习过程的计算量，加速了网络的学习速度。文中最后给出了采用线性化算法与标准ＢＰ算法对正弦函数的学习过程。

Faster linearization learning algorithm for feedforward network

To overcome disadvantage of slow convergence rate of back propagation (BP) algorithm for feedforward network, a new learning procedure which is called faster linearization learning algorithm is presented. The algorithm adopted standard BP algorithm in the beginning of network training, but it linearized the nonlinear processing elements when the network weights that varied very small at that time were near the optimal values. So the network nonlinear processing elements were expanded by Taylor series and expressed approximately by a first order Taylor series. The algorithm simplified the network training process and accelerated network training speed. The network training process to map the relationship of sine function using both linearization learning algorithm and back propagation algorithm is provided.