柔性神经网络训练算法的研究

On Training Algorithms of Flexible Neural Networks

针对经典BP网络训练速度慢、易陷入局部最小值而无法收敛的弱点,提出用具有高度柔性的柔性神经网络代替经典网络,并以矩阵作为基本运算单位导出了柔性神经网络训练的最速下降法和LM(Levenberg Marquard)算法。矩阵作为基本运算单位的优点是可以用高效矩阵库LAPACK来编写程序,提高了数值计算的精度和速度。仿真结果表明了算法的有效性。

To the problem that flexible neural networks is a kind of network structure with high flexibility, but the training algorithms is not so rich compared with classic neural networks, using matrix as the basic arithmetic unit, the steepest descent algorithm and LM optimization algorithm are deduced. With matrix being used as the basic arithmetic unit, highly efficient LAPACK can be applied to deal with programming, which shall increase the accuracy and speed of numerical computation. Finally, a simulation example shows the validity of the algorithm, and indicates that flexible neural network, to a certain degree, overcomes the disadvantages of classic BP network training.