Advanced Networks Colloquium: Tara Javidi, Information Acquisition & Sequential Refinement of Belief
Friday, October 28, 2016
1146 AV Williams Building
Information Acquisition and Sequential Refinement of Belief
Electrical and Computer Engineering
University of California, San Diego
Information acquisition problems form a class of stochastic decision problems in which a decision maker is faced with utilizing a stochastically varying (and uncontrollable) environment. However, the state of the environment, due to the limited nature of the measurements in terms of dimension/ complexity/cost/accuracy, is only partially known to the decision maker. In this setting, to best utilize the system, our decision maker has to carefully control the acquisition process, which results in uncertain noisy measurements and outcomes. In other words, the utilization of the state depends on how well the acquisition process dynamically refines the belief about the underlying stochastically varying environment. This problem arises in a broad spectrum of applications such as medical diagnosis, cognitive radio, sensor management, active learning, generalized noisy search, and noisy group testing. A generalization of hidden Markov models and partially observable Markov decision problems, information acquisition is both an information problem as well as a control one.
In this talk, we start with active hypothesis testing as a special case of information acquisition. This problem has been studied in various areas of applied mathematics, statistics, and engineering. The first part of the talk discusses the historical developments due to Wald, Blackwell, DeGroot’s, and Chernoff and the missing link of information acquisition rate. In this context, we go back and connect DeGroots information utility framework with the Shannon theoretic concept of uncertainty reduction to introduce a symmetrized divergence measure: Extrinsic Jensen-Shannon (EJS) divergence. We use this approach to revisit Chernoff’s treatment of the problem and provide a rate-reliability generalization. Going back to our original dynamic setting, we discuss a formal equivalence between the information acquisition problem and generalized tracking of a time varying state (hypothesis) with partial observation as well as the connections to the filter stability problem.
We focus the second part of the talk, interactively with the help of the audience, to detail recent advances in one of the following applications of interest: Noisy Search, Active Learning from Noisy Labels, Distributed Social Learning, and Adaptive Object Detection in Computer Vision.
This work was done in collaborations with my PhD students as well as K. Chaudhury, A. Goldsmith, Y. Kaspi, A. Sarwate, O. Shayevitz, and M. Wigger.
Tara Javidi studied electrical engineering at Sharif University of Technology, Tehran, Iran from 1992 to 1996. She received her MS degrees in electrical engineering (systems) and in applied mathematics (stochastic analysis) from the University of Michigan, Ann Arbor, in 1998 and 1999, respectively. She received a Ph.D. in electrical engineering and computer science from the University of Michigan, Ann Arbor, in 2002. Her research interests are in stochastic control theory, information theory with feedback, and wireless and computer networks.