Fast machine learning algorithms for large data
Arindam Choudhury · ePrints Soton (University of Southampton) · 2002
Traditional machine learning has been largely concerned with developing techniques for small or modestly sized datasets. These techniques fail to scale up well for large data problems, a situation becoming increasingly common in today’s world. This thesis is concerned with the problem of learning with large data. In particular, it considers solving the three basic tasks in machine learning, viz., classification, regression and density approximation. We develop fast memory- efficient algorithmics for kernel machine training and deployment. These include considering efficient preprocessing steps for speeding up existing training algorithms as well as developing a general purpose framework for machine learning using kernel methods. Emphasis is placed on the development of computationally efficient greedy schemes which leverage state-of-the-art techniques from the field of numerical linear algebra. The algorithms presented here underline a basic premise that it is possible to efficiently train a kernel machine on large data, which generalizes well and yet has a sparse expansion leading to improved runtime performance. Empirical evidence is provided in support of this premise throughout the thesis.