CSE DSI Machine Learning Seminar with Kevin Miller (BYU)
Active Learning for Data-Efficient Machine Learning: Graph-based Classification and Dynamical Systems
Machine learning applications often face constraints due to labeling costs, storage limits, or expensive experiments. Active learning (sampling) mitigates these challenges by strategically selecting which data points to label. This talk will provide an overview of the active learning paradigm, as well as explore two contexts: (1) graph- based semi-supervised classification and (2) learning dynamical systems.
Graph-based semi-supervised classification methods use similarity graphs to propagate labels from a small set of labeled nodes to the rest of the (unlabeled) dataset. Active learning in this setting identifies the most informative unlabeled nodes to refine classification. We present our recent work in designing computationally efficient active learning methods that explore the clustering structure of the graph and in turn efficiently refine class decision boundaries. We also highlight successful applications of these methods to hyperspectral and multispectral pixel classification, as well as synthetic aperture radar (SAR) classification.
Shifting to dynamical systems, we introduce an active learning algorithm that incorporates prior domain knowledge to guide sampling. By focusing exploration on regions with high empirical discrepancy between observed data and an imperfect prior model, our approach efficiently refines the learned dynamics. Time permitting, we will also present a rigorous analysis proving that our algorithm provides a consistent estimate of the underlying dynamics, along with an explicit rate of convergence for maximum predictive variance.
Kevin Miller is an Assistant Professor of Mathematics at Brigham Young University, in Provo, Utah. He received his PhD in Mathematics from the University of California, Los Angeles (UCLA) where he was supported by the National Defense Science & Engineering Graduate (NDSEG) Research Fellowship. He was a Peter J. O’Donnell Jr. postdoctoral fellow in the Oden Institute for Computational Engineering & Sciences at the University of Texas, Austin, involved with the NSF-funded Institute for the Foundations of Machine Learning (IFML). His research focuses on designing and analyzing data-efficient methods for machine learning, lying in the intersection of statistical learning theory, active learning, numerical methods, and probability.
Keller 3-180 or via Zoom.