Nov 2010: This page is a bit out of date; I am hoping to update is soon.

My research interests are briefly summarized below, followed by more detailed descriptions of specific research projects I have been involved in.

• Machine Learning

  I am interested in both core machine learning (theoretical analysis of computation involved in statistical learning and design of new algorithms and model classes) and applied machine learning, where the primary goal is to provide a practitioner in a particular problem domain with suitable tools for the problem at hand. I believe that theory and applications of learning greatly benefit from each other. In particular, a challenging practical problem can often provide inspiration for a new way of thinking which in turn leads to conceptually new models or algorithms.

  • Learning similarity, as a task-specific concept (this was the subject of my PhD thesis).
  • Non-parametric methods for classification, density estimation and regression.
  • Modeling structured time series in unsupervised and semi-supervised scenarios.

• Brain-machine interfaces

  My current work at Brown with Michael Black focuses on developing mathematical methods for decoding neural (cortical) code for movement and using it for direct cortical control of motor activity in artificial systems. The primary application for this is in neuro-motor prosthetics for patients whose motor cortex is intact but who have lost control control of motor function due to injury or decease. We would like to bypass the damaged neural pathways by means of computation: decode the "commands" issued by the brain and translate them into commands to, say, a robotic manipulator. Of course, such understanding of the neural code would also have great scientific implications. We are working in collaboration with Donoghue Lab and Cyberkinetics, Inc.

• Computational Vision

  Much of my past work has been on problems in computer vision. I remain very interested in this area, in particular the following topics.

  • Articulated body modeling and pose estimation. I have worked on new, example-based approaches to these problems; details below.
  
  
  • Visual object categorization and recognition.

• Neural decoding of dexterous hand manipulation. This is the main project of my postdoctoral work, done in collaboration with Michael Black, Carlos Vargas-Irwin and John Donoghue. We are working on learning structure (motor primitives) in natural hand movements and, at the same time, try to decode the movements from simultaneously recorded neural signal. An inportant context of this work is brain-machine interfaces however it is of great scientific interest in its own right.



• Risk of approximate nearest-neighbor classification. In many practical applications, empirical performance is excellent, given enough data. That, however, leads to often infeasible computational expense involved in searching a database for neighbors. One remedy is to use fast approximate search algorithms. With John Fisher, we have started investigating the tradeoff between the speedup gained by approximation and the potential loss in accuracy. Preliminary results were presented at the 2006 Learning workshop at Snowbird; see our poster.



• Example-based articulated pose estimation and tracking. This is an ongoing project, in collaboration with L. Taycher, D. Demirdjian and T. Darrell. We have developed a framework that allows to accurately and very efficiently search a large database of images of people for examples with pose similar to the pose of the person in an single input image. The approach, Parameter-Sensitive Hashing, is based on learning hash functions that are sensitive to the similarity in pose space. The key contributions in this work is the connection between hashing and classification of pairs, and the notion of learning embeddings that reflect distance in the parameter space for regression problem involved in pose estimation. We have explored a number of methods for integrating such example-based mechanisms into motion-based articulated tracking systems.
  
  
• Learning task-specific similarity - subject of my PhD thesis. I am interested in the retrieval setup, whereby a "user" provides the algorithm with a set of pairs of examples known to be similar under some (hidden) notion of similarity, and possibly another set of pairs of dissimilar examples. Then, given a set of reference examples, the goal is for previously unseen query to retrieve reference examples similar to the query - that is, the ones that the user would deem similar - with high precision/recall. The core of this problem is how to learn task-specific similarity, and how to do the search with respect to the learned similarity in an efficient way that would allow dealing with very large data sets. Example applications of this approach are example-based pose estimation, Motion-based animation and visual fragment similarity.
  
  
• Face recognition from sets of images. This is an ongoing project; I have worked on this with John Fisher and Trevor Darrell, and recently we have been collaborating with Ognjen Arandjelovic and Roberto Cipolla from the University of Cambridge. We are insterested in a scenario in which a classification algorithm gets to see a set of observations that are known to all belong to the same (unknown) class - for instance, if a surveillance camera collects a set of face snapshots from a person. Our goal is to develop an approach to classification that would use the information contained in a set.

---

Past projects




• Motion-based animation interface, collaboration with Liu Ren, Paul Viola, Jessica Hodgins and Haspeter Pfister. In this work we extended the approach taken in PSH and introduced a number of important improvements, namely the greedy learning of similarity classifiers, which improves the performance significantly by properly dealing with dependencies), and the integration with temporal context by using the similarity classifiers in conjunction with the motion graph.
  
  
• Inferring 3D Structure with a Statistical Image-Based Shape Model. Joint work with Kristen Grauman and Trevor Darrell.



• Hypercuts - boosted dyadic discriminants. A new ensemble classifier, which can be seen as an SVM contructed one pair of opposing support vectors at a time. Joint work with Baback Moghaddam at MERL.



• Integrated face and gait recognition with multiple views. Joint work with Lily Lee and Trevor Darrell.