Past Meetup

BNPy: Bayesian Nonparametric clustering with Python

Location visible to members

Details

On February 8th, Mike Hughes, of Harvard University, will present BNPy: a Python package for Bayesian NonParametric Clustering. The event will be held at Acquia Offices in Downtown Boston.

Abstract

How can we go beyond the k-means algorithm to find clusters in huge datasets with hierarchical or sequential structure? In this talk, I'll present work on our Python software package, BNPy, which fits modern Bayesian nonparametric models (the Dirichlet process mixture model and its extensions) in ways that scale well to millions of examples and avoid local optima by adding and removing clusters during a single training run. I will demonstrate our new variational inference algorithms on tasks like discovering salient topics from news articles, image patch modeling, and motion capture analysis. Time allowing, I'll also present recent applications of similar models to predicting interventions in the intensive care unit (ICU).

Speaker Bio

Mike Hughes is currently a postdoctoral researcher at Harvard University, where he applies machine learning to improve clinical decision-making with Prof. Finale Doshi-Velez. He completed a Ph.D. in the Department of Computer Science at Brown University in May 2016, advised by Prof. Erik Sudderth. His research interests include building flexible nonparametric models, developing effective optimization algorithms for non-convex problems, and applications in medicine and the sciences. He completed his undergraduate studies at Olin College of Engineering (2010). You can find his papers and code on the web at www.michaelchughes.com (http://www.michaelchughes.com/).

Agenda:

• 6pm: Meet and greet. Networking.

• 6:30pm: Talk by Mike Hughes and Q&A.

• 7:30pm: Networking

• 8:30pm: End of event.

Sponsor:

This event is sponsored by Acquia (https://www.acquia.com/).

Note to all attendees:

All attendees need to present your ID at the security desk in the building lobby in order to be able to access the elevators to the 10th floor. Names of registered attendees will be provided to building security in advance in order to speed up this process so please be sure to register in good time.