The pendulum in Artificial Intelligence (AI) research has periodically swung from so called "neat" or mathematically rigorous approaches, and "scruffy" or more adhoc approaches. In recent years, real-world data across varied fields of science and engineering are increasingly complex, and involve a large number of variables, which has resulted in a surge of scruffier methods. This proposal develops a general "neat" framework for such modern settings by leveraging state of the art developments in two of the most popular subfields of machine learning methods: graphical models and high-dimensional statistical methods. These developments have in common that a complex model parameter is expressed as a superposition of simple components, which is then leveraged for tractable inference and learning.
Our unified framework results not only in a unified picture of these developments but also provides newer methods to work with such high-dimensional data. The research thus impacts problems across science and engineering wherever statistical machine learning approaches are being used (such as genomics, natural language processing and image analysis, to name a few). The work on a unified framework for statistical machine learning problems is highly coupled with a push for imparting training to students on what we call "comptastical" thinking. This combines both computational and statistical thinking required for addressing the problems of limited computation and limited data inherent in modern statistical AI application domains. The proposal also develops an infrastructure for component-based courses with relationally organized lecture module components.
