反馈神经网络的一种反向传播算法

曾有人提出利用线性最小二乘法解决神经网络求解优化问题，如用于 Ｂ Ｐ、 Ｈｏｐｆｉｅｌｄ 网络求解。但在许多实际应用中，典型的反向传播法速度较慢。顾提到的反馈神经网络新型算法基于线性代数方法，使用线性最小二乘技巧获得每个神经元的激活函数的权重和，并根据理想权重和与实际权重和的“子误差”来逐层调整权重。在描述时间序列等应用方面，其速度比经典的反向传播要快几个数量级。

An Algorithm of Dynamic Linear Least Squares Backpropagation for Feedback Neural Networks

Make use of linear least squares to solve the optimal problems of neural network has been put forward. For example, it is used to ask for the answer of BP network and Hopfield network. But in many applications, the speed of typical backpropagation training method is slower. Advanced new training method of recurrent neural network is based on linear algebyaic in this paper. Using linear least squares techniques to get wighted sum of activation function of each neuron, and according to the sub\|error between the desire sum and actual sum to adjust the weights one by one. As respect of describing the time order, the speed is faster than that of the essence backpropagation.