What's new?

[Very outdated! I'm no longer doing search or face recognition, but all sorts of large scale statistical inference.]

Collaborative filtering

Netflix. The constant challenge. I'm still at it, and Ole Winther and I are still Gibbs sampling and constructing mean field approximations. Blaise Thomson is in the loop too. See our talk at the NIPS 2007 workshop (link at end of "Research" page). Now to write it all up...

Free energy and perturbations

Recently, Sudderth, Wainwright and Willsky (Loop Series and Bethe Variational Bounds in Attractive Graphical Models) showed that belief propagation (Bethe-Peierls free energy) can in certain cases give a lower bound to a partition function, or, in Bayesian language, a log marginal likelihood. They used Chertkov and Chernyak's loop expansion to show their result; the bounds were also much better than standard variational or mean field bounds! Expectation propagation is belief propagation's bigger brother. When will EP give a bound? In the "Research" page there is a link to a paper by Manfred Opper, Ole Winther and myself. We already have a Feynman diagram expansion... This is too interesting. I wish I had more time!

Image similarity search

Very much bobbing around the top of the pile at the moment. Imense now has a nice image similarity search system, and I can spell "Java" backwards now! Lots of wavelets, and heuristics, and random projections... There will be a web front-end soon, but in the meantime, we're happy to demo what we have! It is already being used commercially. I'm also interested in inverted indexing and search, which brings me to "Hashing".


How do we make retrieval fast and very scalable? How do we design machine learning methods that run in O(log N) or constant time?

Face recognition and all that

For face recognition in full form, please see the advanced options in the Imense search engine. I'm really happy with the results, and as usual Chris Town has done a sterling job of putting all sorts of probabilities from lots of classifiers into our index. I was also toying with automatic face clustering and labelling; here is a screenshot: