基于梯度动力学的协同神经网络学习算法的改进

本文在研究协同神经网络梯度动力学过程的基础上，针对学习过程收敛速度缓慢的缺点，提出了一种改进的、基于梯度动力学的协同神经网络学习算法。该算法分析了非平衡注意参数对学习过程的影响，简化了初始伴随向量的选取；并引入最优化理论，将该问题归结为求解非线性最优化问题，用共轭梯度法代替梯度下降法，加快了学习过程的收敛。通过对汉字图像库和人脸图像库的图像识别实验表明，该算法比其他学习算法的识别率高，并能较快地收敛到极小值。

The Improved Learning Algorithm of Synergetic Neural Networks Based on Gradient Dynamics

An improved learning algorithm of synergetic neural networks is proposed after studying the synergetic gradient dynamics process. The improved algorithm analyzes the effect of the unbalanced attention parameter on the study process, which reduces the complexity of initializing the adjoint vector, and then introduces an optimization approach by using the conjugate gradient algorithm instead of the steepest decent algorithm. The experiments of the recognition of Chinese characters and face images show that the new algorithm is more efficient in convergence property and recognition rate.