任意前馈神经网络的稀疏化算法

Algorithm of Melting Sparsly of Wanton Feedforward Neural Network

主要讨论前馈神经网络的稀疏化,即如何确定和删除网络中冗余的神经元和连接。首先给出前馈神经网络的数学定义,并将偏序和拓扑排序引入到前馈神经网络的学习算法和稀疏化算法中。在此基础上提出了冗余神经元和连接的判断依据,并按照自构形和自调整的策略,提出了适用于前馈神经网络的自构形学习算法和自调整删减算法。实验结果表明,上述的稀疏化算法不仅能够有效地删除网络中冗余的神经元和连接,而且能够改善网络的性能。

This text discusses melt sparsely of the feed forward neural network,that is how to determine and delete the network＇s redundant neuron and joining, gives the mathematics define of feed forward neural network, and introduces the Lean towards preface and Arrange in an order topologically to the Study algorithm and Sparse to take the algorithm of feed forward neural network. Based on this the judgment basis of the redundant neuron and joining is put forward. And according to the tactics of since topography and adjusted by oneself, proposes the study algorithms in since topography and adjust and delete the algorithm of reducing by oneself which is suitable for feed forward neural network. The experimental result indicates ,the above-mentioned algorithm of melting spars lies can not merely delete the redundant neuron in the network and join effectively ,but also improve the performance of the network.