Research

Interests

I'm interested in statistical machine learning and probabilistic models of human cognition. My long-term goal is to build intelligent computer programs that are inspired by human cognition.

C.V.

I try to keep my curriculum vitae up-to-date. You can view a recent version of it in either PDF or HTML.

Peer-reviewed publications

Kevin R. Canini, Lei Shi, and Thomas L. Griffiths, Online inference of topics with latent Dirichlet allocation, Proceedings of the 12th International Conference on Artificial Intelligence and Statistics (AISTATS), April 2009.

Thomas L. Griffiths, Adam N. Sanborn, Kevin R. Canini, and Daniel J. Navarro, Categorization as nonparametric Bayesian density estimation, M. Oaksford and N. Chater (Eds.), The Probabilistic Mind: Prospects for Rational Models of Cognition, Oxford: Oxford University Press. March 2007.

Thomas L. Griffiths, Kevin R. Canini, Adam N. Sanborn, and Daniel J. Navarro, Unifying rational models of categorization via the hierarchical Dirichlet process, Proceedings of the 29th Annual Conference of the Cognitive Science Society, February 2007.

Conference posters

Kevin R. Canini, Lei Shi, and Thomas L. Griffiths, Online inference of topics with latent Dirichlet allocation, 12th International Conference on Artificial Intelligence and Statistics (AISTATS), April 2009.

Kevin R. Canini and Thomas L. Griffiths, The hierarchical Dirichlet process as a model of human category learning, NIPS 2008 workshop: Machine Learning Meets Human Learning, December 12, 2008.

Other papers

Master's thesis: Kevin R. Canini, Modeling categorization as a Dirichlet process mixture, Technical Report UCB/EECS-2007-69, EECS Department, University of California, Berkeley. May 2007.

Jacob Abernethy, Kevin R. Canini, John C. Langford, and Aleksandr Simma, "Online Collaborative Filtering", February 2007.

Peter Bodík, Michael P. Armbrust, Kevin R. Canini, Armando Fox, Michael I. Jordan, and David A. Patterson, A case for adaptive datacenters to conserve energy and improve reliability, Technical Report UCB/EECS-2008-127, EECS Department, University of California, Berkeley. September 2008.

Projects

Human categorization modeling

We are exploring hierarchical Dirichlet process (HDP)1 mixture models as a way to explain human performance on categorization tasks. HDPs provide multiple benefits over traditional prototype and exemplar models of categorization, including automatically tuning representational complexities to the available data and sharing structures among multiple categories. We are currently running a human subject experiment to collect data on the transfer effect of learning overlapping systems of categories.

Colleagues

Online topic modeling

We have introduced two online algorithms – an incremental Gibbs sampler and a particle filter – for automatically inferring the topics in a collection of documents using the latent Dirichlet allocation (LDA)2. They can be used for things like clustering news articles in real-time or alerting users about ongoing multi-party conversations with relevant topics. They are easily parallelizable and can be adjusted to trade off runtime vs. performance.

For fun, I've created a video illustrating an expectation-maximization (EM) algorithm solving a topic modeling problem using the pLSA model (which is LDA without priors). Check it out here. The black triangle is the space of all possible distributions over 3 words. Each vertex corresponds to putting 100% probability on a single word. The red triangle's vertices are topics, and the blue circles are approximations to the true documents, which are the green asterisks.

Colleagues

Undergraduate research

References

[1] Yee Whye Teh, Michael I. Jordan, Matthew J. Beal, and David M. Blei, Hierarchical Dirichlet processes, Journal of the American Statistical Association, Vol. 101 (476), 2004.

[2] David M. Blei, Andrew Y. Ng, and Michael I. Jordan, Latent Dirichlet allocation, Journal of Machine Learning Research, Vol. 3, 2003.