论文部分内容阅读
本文提出一种用于多层前向网络的广义能量约束算法。新的能量函数不仅包括误差平方和项,而且还包括某些能使网络获得所期望的性能的约束项.讨论了某些典型的约束形式,施加在权重矢量或隐藏层输出或网络结构上.将本方法用于信号处理中广泛采用的奇异值分解(SVD)问题,取得较好效果.此方法还可以引申到主分量分析(PCA)。方法具有高度灵活性,并使矩阵的并行运算成为可能.能量约束中还可以引入有关信号特征的先验知识,从而使作信源分解时对信号和传输阵的正交要求有所降低.初步仿真结果表明本方法有良好应用前景.
This paper presents a generalized energy constraint algorithm for multi-layer forward networks. The new energy function includes not only the sum of the square terms of the error but also some of the constraints that give the network the desired performance, and discusses some of the typical forms of constraints that are imposed on the weight vector or hidden layer output or network structure. The proposed method is applied to singular value decomposition (SVD) which is widely used in signal processing and has good effect. This method can also be extended to PCA. The method is highly flexible and makes parallel computing of the matrix possible.It is also possible to introduce a priori knowledge of the signal characteristics into the energy constraint so as to reduce the orthogonality of the signal and the transmission matrix when the source is decomposed. Simulation results show that this method has good application prospects.