一种新的基于梯度动力学的协同神经网络学习算法

A Novel Learning Algorithm of Synergetic Neural Networks Based on Gradient Dynamics

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,介绍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,提出了适时地用共轭梯度法代替梯度下降法的算法,加快了学习过程的收敛。通过对标准人脸图像库的图像识别实验表明该算法较之其他学习算法有较高的识别率,并能较快地收敛到极小值。

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