Large data and zero noise limits of graph-based semi-supervised learning algorithmsMS17

Graph-based approaches to semi-supervised classification typically employ a graph Laplacian L constructed from the unlabelled data to provide clustering information. We consider Bayesian approaches wherein the prior is defined in terms of L; choice of data model then leads to probit and Bayesian level set algorithms. We investigate large data limits of these algorithms from both optimization and probabilistic points of view, and observe a common necessary condition for the limiting algorithms to be meaningful.

This presentation is part of Minisymposium “MS17 - Discrete-to-continuum graphical methods for large-data clustering, classification and segmentation
organized by: Matthew Thorpe (University of Cambridge) , Luca Calatroni (CMAP, École Polytechnique CNRS) , Daniel Tenbrinck (University of Münster) .

Matt Dunlop (Caltech)