Scalable Algorithms For Distributed Statistical Inference
Animashree Anandkumar · eCommons (Cornell University) · 2009
The classical framework on distributed inference considers a set of nodes taking measurements and a fusion center making the final decision on the underlying phenomenon, without dealing with the issue of transporting the measurements to the fusion center. Such an approach introduces significant overhead in com-munication. Communicating all the raw data for inference is not scalable: in this case, the per-node average energy consumption and the total bandwidth requirement become unbounded as the network grows. We design scalable algorithms for two scenarios with guarantees for infer-ence whose communication requirements and complexity are bounded even as the network grows. This is achieved through distributed computation of a suffi-cient statistic, which results in reduction of data dimensionality while ensuring no loss in inference accuracy at the fusion center. The first scenario deals with multihop routing and fusion of spatially correlatedmeasurements, incorporated through a Markov random field model. The second scenario deals with design of medium-access control (MAC) with the aim of computing a sufficient statistic for inference over a multiple access channel. BIOGRAPHICAL SKETCH Animashree Anandkumar hails from a family of engineers and teachers, and spent her childhood in the city of Mysore in southern India. She received her B.Tech in Electrical Engineering from the Indian Institute of TechnologyMadras in 2004 with a minor in Theoretical Computer Science. She is a PhD student in Electrical Engineering at Cornell University with a minor in Applied Math-ematics. Since Fall 2004, she has been working with the Adaptive Communi-cations and Signal Processing (ACSP) group under the direction of Prof. Lang Tong. She spent two summers at IBM Watson Research, Hawthorne, NY work-ing on transaction monitoring in financial systems. She is currently visiting the Stochastic Systems Group (SSG) at MIT hosted by Prof. Alan Willsky.