Abstract: Feature learning is a key issue in the field of pattern recognition. By unsupervised pre-trainning and supervised fine-tuning, the deep neural network based on autoencoders can effectively extract critical information of data to form features. A marginal Fisher analysis algorithm based on stacked denoising autoencoders has been proposed which can further improve the ability of representation learning by applying marginal Fisher analysis to the supervised fine-tuning phase. Experimental results show that the algorithm achieves better recognition results compared to the standard stacked denoising autoencoders and the deep belief networks based on restricted Boltzmann machine.

Key words: feature learning, deep learning, artificial neural network, stacked denoising autoencoders, marginal Fisher analysis

摘要： 特征学习是模式识别领域的关键问题。基于自动编码器的深度神经网络通过无监督预训练与有监督微调能够有效地提取数据中关键信息，形成特征。提出一种基于栈式去噪自编码器的边际Fisher分析算法，该算法将边际Fisher分析运用于有监督微调阶段，进一步提升算法的特征学习能力。实验结果表明，该算法与标准的栈式去噪自编码器和基于受限玻尔兹曼机的深度信念网相比，具有更好的识别效果。

关键词: 特征学习, 深度学习, 人工神经网络, 栈式去噪自动编码器, 边际Fisher分析