Neural-Network-Based Feature Learning: Auto-Encoder

Abstract

In recent years, deep learning has garnered tremendous success in a variety of application domains. As a representative of unsupervised learning in deep learning, auto-encoder (AE) is favored by many researchers because of its good feature learning ability and the ability to process a large amount of unlabeled data and save manpower and material resources. In this chapter, the AE and its variant versions are elaborated. In particular, the basic theory and characteristics of denoising auto-encoder (DAE) and stacked denoising auto-encoders (SDAE) are introduced. Then, combined with various application fields, the application of various AEs is classified and summarized. In the experimental part, we introduced the application of AE in fault diagnosis in detail and also proved that the performance of traditional AE is better than PCA.