Emin Martinian
Mitsubishi Electric Research Laboratories, Cambridge, Mass.
Thursday, February 23, 2006
4:15pm
Packard 101
Abstract:
I will describe some principles of distributed information and statistical inference which enable technologies such as multi-camera compression, wireless communication, error correcting codes, lossy data compression, computer vision, authentication, and data security.
First, I will introduce a model with a source X and some quality side information Q. When an observer describes X to a receiver, the goal is to most accurately convey the important components of X indicated by Q. For example, Q can capture perceptual effects in the human audio-visual system or relative importance based on context. If both observer and receiver know Q, the obvious solution is to spend more bits to describe important parts of X. But what if only the observer or only the receiver knows Q? Surprisingly, there is often no penalty for such distributed information and the resulting signal representations can significantly improve data communication.
Next, I will describe how to design lossy data compression systems using random graphs. Quantization using graphs corresponds to statistical inference on Bayesian networks. But encoding algorithms based on belief propagation and graph constructions based on low density parity check codes both perform poorly. By modifying belief propagation and using large deviations techniques to design good graph ensembles, however, one can obtain provably optimal systems.
Finally, I will discuss secure storage of biometrics such as iris images or fingerprints that are often used for authentication, access control, and encryption. While it is well known that passwords should never be stored in the clear, current systems often store biometrics in the clear and are easily compromised. Instead, secure systems can be constructed by combining distributed information, graphical models, and statistical inference in a novel way.
Short Biography:
Dr. Emin Martinian earned a bachelor degree in EE/CS from the University of California at Berkeley in 1997. After a year and a half at the startup OPC Technologies, he entered MIT in 1998, receiving a master degree in 2000 and a doctoral degree in 2004. His master research covered multimedia authentication, and his doctoral thesis considered dynamic information and constraints in source and channel coding.
Since completing his doctorate, he has been working on applications of distributed information processing and statistical inference to problems in biometric security, video processing, distribution, and compression at Mitsubishi Electric Research Laboratories in Cambridge, MA. His broader research interests include information theory, belief propagation, graphical models, cryptography, digital communications, and signal processing. While at MIT Dr Martinian held an NSF Graduate Fellowship and received the Capocelli Best Paper Award from the 2004 Data Compression Conference.