Dr. Chao Tian

Research Directions

We conduct research in the general area of information processing, storage, machine learning, and communication. Some of the most significant topics of current and recent research are explained below, with some of our example publications on these topics.

Machine Learning

Machine learning has made tremendous advances in recent years. However, theoretical understanding of learning algorithms, particularly deep neural networks, has not kept up at the same level. We studied the generalization error using information theoretical approach. Another area of interest is reinforcement learning, where new algorithms have been proposed dealing with diverse settings.

Joint Source-Channel Coding and Semantic Communication

The current communication network is built with the well known layered structure, however, underlying this structure is the assumption of such separations do not cause much loss of performance. As machine learning becomes prevalent, the question of whether this separation continues to be a good choice becomes less clear.

Modern information system must provide security and privacy guarantees, in addition to the regular data retrieval functionality. Furthermore, caching can be viewed as part of system to be optimized. How can coding be used to improve the efficiency of such systems? We rely on algebraic and combinatoric structures to construct codes, where information theoretic outer bounds can not only provide theoretic performance limit (some of which are obtained computationally), but also clues in code constructions.

Proving outer bounds or converses for complex information systems require specialized skills and extensive training in information theory, which means significant human diligence and intelligence. As modern computer software and hardware become more and more powerful, bordering on the edge of true artificial intelligence, we must ask whether computers can be used to at least aid this human-heavy process? We show that this is indeed possible, and in fact can be tremendously effective, by providing concrete solutions to several important problems of current research interest.


In this information age, users and digital devices are constantly producing data, and the need for reliable data storage for the data in the cloud environment has grown exponentially. Next generation of data storage systems have to be reliable, distributed and agile, with a high-availability guarantee. We are working on various aspects of this problem, which includes, e.g., new erasure codes with reduced repair bandwidth requirement, system designs for distributed data storage, secure distributed storage codes, distributed data storage for lossy compression.

Images as a special type of signals have its own characteristics, and require many special techniques in coding and processing. For example, in multiple description image coding, information theoretic optimal solutions may not suit the best for images. In 3D image processing and object reconstruction, tools such as bilateral filtering turn out to be quite effective, and optimization techniques can be more effectively used by incorporating visual clues. Even for the classical image compression problem, simpler codec proves to be useful: the TCE embedded image codec (ICASSP 04) (available in QccPack software library) has been used by many researchers.

  • C. Tian and S. Krishnan, “Accelerated bilateral filtering with block skipping,” IEEE Signal Processing Letters, Vol. 20, No. 5, pp. 419-422, May 2013.
  • U. Samarawickrama, J. Liang, and C. Tian, “M-channel multiple description coding with two-rate predictive coding and staggered quantization,” IEEE Trans. Circuits and Systems for Video Technology, Vol. 20, No. 7, pp. 933-944, Jul. 2010.
  • G. Sun, U. Samarawickrame, J. Liang, C. Tian, C. Tu, and T. D. Tran, “Multiple description coding with prediction compensation,” IEEE Trans. Image Processing, Vol. 18, No. 5, pp. 1037-1047, May 2009.
  • C. Tian, M. Masry, and H. Lipson, “Physical sketching: reconstruction and analysis of 3D objects from freehand sketches,,” Journal of Computer Aided Design, Special Issue on Computer Support for Conceptual Design, Vol. 41, No. 3, pp. 147-158, Mar. 2009.
  • C. Tian and S. S. Hemami, “A new class of multiple description scalar quantizers and its application to image coding,” IEEE Signal Processing Letters, Vol 12, No. 4, pp. 329-332, Apr. 2005.

Network information theory traditionally asks for complete characterizations of network communication problems, partly motivated by the elegant solution on point-to-point channels given by Shannon. However on most cases, such characterizations turn out to be extremely difficult to find. We develop a methodology to instead find approximate solutions, which in most cases are much tractable and in fact leads to nontrivial insight into the problem.

When multiple users are present in a communication problem, source coding techniques to combat the uncertainty in the communication systems become necessary.  Moreover, various additional information may be available to the encoder or the decoders, and “side information” is a general term to model such information. The dependence structure between the source and the side information is not fixed, and more complex coding strategy has to be used. 

Channel capacity characterization and code designs are two central themes in information and communication theory. For broadcast channels, the capacity region is unknown, and the two papers below include some progress toward solving this difficult problem. Constant weight code is a classical problem for which we developed a novel algorithm; polar code, on the other hand, is a new development in the field, based on which we developed optimal codes for communication on the parallel channel.