A Deep Manifold Representation for Information Discovery

Information discovery plays a vital role in the success of various machine learning and data-driven tasks. In essence, it is a process of exploring useful knowledge from input data sources. In this paper, a deep manifold representation method is proposed for information discovery, consisting of multistage manifold filters, a hashing transform and a histogram operation. Specifically, a manifold method, locality preserving projections (LPP) is utilized for constructing multistage filter banks in different layers, followed by a hashing transform and a histogram operation to generate the final manifold representation. In the proposed method, not only is the intrinsic local structure revealed by LPP, but also the abstract representation from different levels is explored by multistage filters, leading to better data representation for information discovery. To demonstrate the effectiveness of the proposed strategy, we conduct experiments on two visual analysis and recognition tasks. Experimental results show that the proposed method is superior to the related methods.