基于多层自动编码机的fisher判别分析
Fisher Discriminant Analysis Based on Stacked AutoEncoders

在常见的特征提取方法中，Fisher判别分析（Fisher Discriminant Analysis，FDA）只能提取线性特征，基于核的方法具有提取非线性特征的能力，但对核函数类型及其参数十分敏感. 文中研究如何有效提取数据特征，提出了一种基于多层自动编码机（Stacked AutoEncoders，SAE）和Fisher标准的特征提取算法，该算法中所使用的深度学习网络模型在训练过程中结合无监督特征提取SAE以及有监督的特征提取FDA. 通过与多层自动编码机、极限学习机（Extreme Learning Machine，ELM）等模型提取的特征进行对比，在数据集Pendigits、mnist、ORL和AR上利用支持向量机对数据特征进行分类，结果表明基于SAE的Fisher变换（FDA-SAE）在分类结果准确率以及分类时间上都有较好的效果. 特别是在小数据集AR上，当样本特征较少的情况下效果非常明显.
：Among the methods of common-featured extraction, Fisher Discriminant Analysis(FDA) can only extract the linear features. While the method based on kernel has the ability to extract nonlinear features, but it is very sensitive to the type of kernel function and its parameters. The effective feature extraction of data is concerned in this paper. A feature extraction algorithm based on Stacked AutoEncoders(SAE) and FDA is proposed. The training process of the deep-learning network model in this algorithm combines unsupervised features extracted by SAE with supervised features extracted by FDA. Through the comparison of features extracted from multilayer automatic encoding machine, extreme learning machine (Extreme Learning Machine, ELM) and other models, by using support vector machine to classify the data features in the data set Pendigits, MNIST, ORL, AR, FDA based on SAE (FDA-SAE) has good performance in the classification accuracy and runtime. Especially in small dataset AR, the effect is especially obvious under the condition of less sample features