一种具有对称结构的神经网络的构造方法研究

Research on Construction Algorithm of Neural Network with Symmetry Structure

文章提出了一种神经网络构造方法—改进的OBD(OptimalBrainDamage)算法,力求使网络的结构(网络权值矩阵)具有三角对称性,定义了衡量网络对称性的三角对称度。一方面,该算法可以提高网络的收敛速度,另一方面,由于网络的对称性,可以有效地提高网络的硬件和软件可实现性。将该算法应用于系统辨识,结果表明在本文提出的算法的训练下,网络结构近似为三角对称,同时不影响网络的学习能力。

An improved OBD algorithm is presented to construct neural network in this paper, which makes the structure of the NN tridiagonal symmetry and the convergence speed faster. Tridiagonal degree is mathematically defined. On the other hand, due to the symmetry structure, the hardware of the NN is easy to realized. Applying this algorithm in an example of system identification, after trained by the improved OBD algorithm, network is of approximately tridiagonal symmetry, while the learning efficiency is still guaranteed.