Data Representation: Learning Kernels from Noisy Data and Uncertain Information

Abstract : Identifying appropriate data representation is critical to many decision making problems. In this project, we focus on learning kernel-based data representation from noisy data and uncertain information. Unlike conventional approaches that represent objects by vectors, kernel representation defines a pairwise similarity between two objects, and is convenient for representing complex objects like graphs. Although many studies are devoted to learning kernel representation, none of them addresses the challenge of learning kernel representation from noisy data and uncertain information. The proposed research aims to address this challenging problem by developing (i) a kernel learning framework that are robust to data noise and information uncertainty, and (ii) efficient algorithms to solve the related optimization problems. The proposed algorithms will be evaluated in the object recognition domain. The impact of the proposed research to the US Army is significant. To counter against future threats to the safety and security of our society, we need to enhance our capabilities to detect, locate, and track such threats by extracting and representing data from noisy observation and uncertain information. The proposed research seeks to significantly advance, both theoretically and computationally, the representation and modeling of information from noisy and uncertain sources.