Nebojsa Jojic
Microsoft Corp., One Microsoft Way, Redmond, WA 98052                                              
425-705-5865                
     
e-mail: jojic@microsoft.com
URL:    http://www.research.microsoft.com/users/jojic

QUICK REFERENCE

Current highlights: 

Protein epitome: We are using similar representations to build HIV vaccines. I have given presentations on this at CROI 05, HVTN 04, and several talks at Hutch, and UW. Joint work with Vladimir Jojic, David Heckerman on the ML side, and Jim Mullins, Simon Mallal on the biology side.

Epitomic Analysis of Appearance and Shape, ICCV 2003

                                                        

Deformable spectrograms

CVPR Tutorial on graphical models (also to appear in PAMI): paper, slides

 

 An audio-visual graphical model, ICASSP 2002

                                 

 

 A generative model for layered vision, CVPR 2001, detailed math     


RESEARCH INTERESTS

I am interested in machine learning approaches to automatic extraction of medium level representations of natural signals. By structuring statistical generative models to mimic the structure of the real world, the models should be able to automatically adapt to audio, visual or multimodal signals during the unsupervised model fitting (learning) stage, thus providing a medium-level representation suitable for compression, transmission, search, editing, enhanced viewing experience, etc. These models are object-based, where an object can produce sounds, have a changing appearance, move and be exposed to attenuation in audio domain, illumination in video domain, and, when other objects are present, to occlusion or additive mixing in both domains. The main requirement is complete adaptivity, and so I am trying to avoid application-specific initialization parameters. For example, the same model should be applicable to tracking a person in front of a cluttered background, and to tracking a flock of birds. The tracking task, as well as many other tasks performed jointly,  such as de-noising, dynamic mosaic building or object removal as well as separating audio sources and associating them to object appearances, are all achievable as probabilistic queries, i.e., inference of the hidden variables associated to the world structure. All this should be doable using the data itself, without special application-specific initialization procedure or the separate supervised training stage.

As of two years ago, I have also been working on biological data, such as sequences, measurements of binding energies, etc. Biological sequences exhibit similar structure to other natural signals and many of the algorithms I have previously developed for other  domains apply to a large extent to biological sequences as well. A prime example is the epitome, which is now being tested at the University of Washington and Harvard University (more precisely, Mass General Hospital), as an immunogen for HIV vaccines. (Here's a short interview I gave on MSNBC)


EDUCATION

University of Illinois, Urbana-Champaign, IL
Ph.D. , Electrical and Computer Engineering, 2001.
thesis title: "Generative models for computer vision"
M.S., Electrical and Computer Engineering, October 1997.
thesis title: "Computer Modeling, Analysis and Synthesis of Dressed Humans"
Advisor: Prof. Thomas S. Huang
 
University of Belgrade, Belgrade, Yugoslavia
B.S., Electrical Engineering, June, 1995
B.S. thesis title: "Image compression using wavelets and scalar coding of sub-bands", advisor: Prof. Miodrag Popovic
 


PUBLICATIONS (probably not a complete list)

Book Chapters:

Journals: M.S. Thesis: "Computer Modeling, Analysis and Synthesis of Dressed Humans"

Ph.D. Thesis: "Generative Models for Computer Vision"

Conferences:


HONORS
Last update:  January 2002.