OPTIMIZING L1 LOSS REGULARIZER AND ITS APPLICATION TO EEG INVERSE PROBLEM
View/ Open
Date
2020-07-21Author
Mainali, Kiran Kumar
0000-0002-8510-8234
Metadata
Show full item recordAbstract
Sparse reconstruction occurs frequently in science and engineering and real-world applications, including statistics, finance, imaging, biological system, compressed sensing, and, today more than ever, machine learning and data science in general. Mathematically, they are often modeled as l1-minimization
problems. There are a number of existing numerical methods that can efficiently solve such l1-minimization problems, such as Alternating Direction Methods of Multipliers (ADMM), Fast Iterative Shrinkage Thresholding Algorithm (FISTA), and Homotopy algorithm.
In this dissertation, we will introduce a special type of l1-minimization problem called the Sylvester Least Absolute
Shrinkage and Selection Operator (SLASSO) problem. In theory, an SLASSO problem can be converted to a standard LASSO problem and then solved by any
existing numerical method, but the converted LASSO problem is too
large scale to be practical even if the SLASSO problem is modest. The first contribution of this dissertation is a novel method to solve an SLASSO problem without conversion, making it practical to solve a fairly large sized SLASSO problem.
Our second contribution is a new structured Electroencephalogram
(EEG)/Magnetoencephalogram (MEG) Source Imaging (ESI) model that groups the time-varying signals of a similar structure and uses the
mixed norm estimation for accurate results. The model is then
solved alternatingly. Numerical simulations compare favorably with
the state-of-the-art ESI methods, demonstrating the effectiveness
of the model and efficient numerical treatment.