Anomaly detection: sparse representation for high dimensional data

Abstract

In this thesis, I investigated in three different anomaly aware sparse representation approaches. The first approach focuses on algorithmic development for the low-rank matrix completion problem. It has been shown that in the l0-search for low- rank matrix completion, the singular points in the objective function are the major reasons for failures. While different methods have been proposed to handle singular points, rigorous analysis has shown that there is a need for further improvement. To address the singularity issue, we propose a new objective function that is continuous everywhere. The new objective function is a good approximation of the original objective function in the sense that in the limit, the lower level sets of the new objective function are the closure of those of the original objective function. We formulate the matrix completion problem as the minimization of the new objective function and design a quasi-Newton method to solve it. Simulations demonstrate that the new method achieves excellent numerical performance. The second part discusses dictionary learning algorithms to solve the blind source separation (BSS) problem. For the proof of concepts, the focus is on the scenario where the number of mixtures is not less than that of sources. Based on the assumption that the sources are sparsely represented by some dictionaries, we present a joint source separation and dictionary learning algorithm (SparseBSS) to separate the noise corrupted mixed sources with very little extra information. We also discuss the singularity issue in the dictionary learning process which is one major reason for algorithm failure. Finally, two approaches are presented to address the singularity issue. The last approach focuses on algorithmic approaches to solve the robust face recognition problem where the test face image can be corrupted by arbitrary sparse noise. The standard approach is to formulate the problem as a sparse recovery problem and solve it using l1-minimization. As an alternative, the approximate message passing (AMP) algorithm had been tested but resulted in pessimistic results. The contribution of this part is to successfully solve the robust face recognition problem using the AMP framework. The recently developed adaptive damping technique has been adopted to address the issue that AMP normally only works well with Gaussian matrices. Statistical models are designed to capture the nature of the signal more authentically. Expectation maximization (EM) method has been used to learn the unknown hyper-parameters of the statistical model in an online fashion. Simulations demonstrate that our method achieves better recognition performance than the already impressive benchmark l1-minimization, is robust to the initial values of hyper-parameters, and exhibits low computational cost.