IPP Symposium
Scaling Nonparametric Learning with Online Variational Inference
Mike Hughes, PhD Student (Machine Learning) & Dae Il Kim, PhD Student (Graphical Models)
A primary goal of unsupervised machine learning is uncovering the hidden structure of documents, images, and social networks. Within this field, the nonparametric Bayesian modeling approach offers a particular advantage: the model complexity can grow with the amount of data seen. This lets the data "speak for itself," freeing researchers from making restrictive assumptions about the specific number of thematic topics to find in Wikipedia articles or the number of communities that exist in a social network.
Until recently, most nonparametric learning algorithms required all data to fit in memory. Groundbreaking recent work by a number of researchers has produced a promising new framework for online learning that allows training from Internet-scale streaming data. In this joint talk, we will describe active projects in our research group that apply these methods to analyze articles from the New York Times, uncover structure in social networks, and remove noise from photographs.
Dae Il Kim is a fourth year PhD candidate at Brown University advised by Erik Sudderth. He works primarily on developing Bayesian nonparametric models for relational datasets such as social networks along with an interest in developing scalable online learning algorithms.
Michael C. Hughes is a third year PhD candidate at Brown advised by Erik Sudderth. His research lies at the intersection of machine learning and computer vision, where he alternates between building expressive nonparametric Bayesian models for videos and images, and developing effective learning algorithms to train these models from data.
