【摘要】 针对神经网络进行小样本数据训练时出现文本表征精确度低及特征过拟合,易造成权值全局最优和局部最优的不平衡现象,提出一种基于粒子群的双向长短期记忆网络(Bi-PSO)算法,利用Bi-LSTM对序列数据中长短期距离依赖信息的能力,对文本特征矩阵进行最小残差化处理得到降维矩阵,并通过粒子群算法获取降维矩阵中特征向量的全局最优和局部最优权重,最终进行权重类间、类内距离的迭代计算获得最优特征子集.仿真实验表明:Bi-PSO算法在文本特征拟合精度上得到了提升,算法精确度比Bi-LSTM平均提高了2.225%,在处理样本数目为200～600小样本数据集时拟合效果良好.更多还原

【Abstract】 In order to solve the low accuracy of text representation and overfit of features in the training of one-shot learning data by neural network, a bidirectional long and short term memory algorithm based on particle swarm optimization is proposed. The characteristics of one-shot learning data set in neural network training are overfitted, which can easily lead to the global optimal and local optimal imbalance of weights. Firstly, the dimension reduction matrix is obtained by minimizing the residual processing of the text feature matrix by using the ability of Bi-LSTM to rely on the information in the long and short distance of the sequence data. Secondly, the global optimal and local optimal weights of the feature vector in the dimension reduction matrix are obtained by particle swarm optimization algorithm. Finally, the iterative calculation of the intra-class distance between the weight classes is carried out to obtain the optimal feature subset. The simulation results show that the Bi-PSO algorithm improves the accuracy of text feature fitting, and the accuracy of the algorithm is 2.225% higher than that of Bi-LSTM on average, the fitting effect is good when dealing with the small sample data set with the number of 200～600 samples.更多还原