SCS Undergraduate Thesis Topics

Andrew Maas Drew Bagnell/Charles Kemp Learning about Related Tasks with Hierarchical Models

One of the hallmarks of human learning is the ability to apply knowledge learned in previous situations to novel scenarios. In contrast, many machine learning algorithms are explicitly designed to train and test on a single, unchanging distribution of examples. My work is aimed at better understanding how we can extend machine learning algorithms as to enhance their ability to use knowledge from previous experience in novel situations. My project explores two domains where the ability to leverage knowledge learned in previous situations enables performance improvements and new capabilities within a task domain. First, I explore the domain of learning real-world concepts with Bayesian networks. Second, I focus on the domain of signal classification, specifically for detecting seizure events in EEG data recorded from patients with epilepsy. In both of these domains, my approach relies on storing knowledge in a hierarchy with multiple levels of abstraction.

Close this window