Research Project Title:
Universal Features for Graphs and Markov Chains via Modal Decomposition
abstract:Many modern problems in inference and systems design require access to complex distributions over a large domain. In general, memory constraints and intractable computations make it infeasible to model joint distributions within the domain. Current models make a strong and often unfounded assumption of mutual independence at the cost of inaccurate approximations. The goal of this project is to develop a graphical model that accounts for dependence over the domain and does so in a way that minimizes memory usage and query latency. To do so, we propose a tree-structured model of the data and introduce a modal decomposition drawn from principles in information geometry to factor the distribution in a way that is accurate and does not exhaust memory.
"This SuperUROP project is exciting because it allows me to explore concepts from and contribute to active research fields in information theory and machine learning. I hope to learn more about graphical models and information geometry and gain experience to improve my research skills."