Discovery Through Constraints: Imposing Constraints on Autoencoders for Data Representation and Dictionary Learning

摘要

Without the proper choice of constraints, autoencoders (AEs) are capable of learning identity mapping or overcomplete representations. The features learned by this architecture may be local, isolated, or primitive. The extraction of features, however, can be controlled by judiciously enforcing some desired attributes in the form of constraints on its parameters. This article gives an overview of AEs and such constraints for data representation. It also puts AE learning in the broader context of dictionary learning.