CAREER: A New Look at the Fundamental Limits of Lossy Network Compression

Cornell University, Ithaca NY

Investigators

Abstract

CAREER: A New Look at the Fundamental Limits of Lossy Network Compression Lossy compression plays a key role in our information economy. By far, most of the information that we generate as a society represents pictures, sounds, and videos, and for this kind of data, lossy compression yields a tremendous reduction in transmission and storage requirements. The aim of this project is to understand the fundamental limits of lossy compression,especially in the context of networks, which dominate today's communication infrastructure. Previous work on this problem has not led to a fundamental understanding of lossy network compression nor has it had a significant impact on the design of practical systems. This research overcomes the limitations of prior work using two novel approaches. First, the highly general models of the past are eschewed in favor of canonical, concrete problems that are simultaneously more tractable and more relevant to applications. Building on recent results of the PI, the work develops a comprehensive and conclusive understanding of the fundamental limits of lossy network compression for Gaussian sources under quadratic distortion constraints and discrete sources under erasure distortion constraints. Second, these fundamental limits are studied under more realistic scenarios in which the source, the network, and the allowable distortion all change in unpredictable ways. To help students learn how to create probabilistic and information-theoretic models that are tractable yet also applicable to real systems, existing courses at both the graduate and undergraduate level are being remade to include greater emphasis on modeling. The project also includes significant outreach, involving an interactive tutorial on compression delivered to women and underrepresented-minority engineering undergraduates and rural high-school students.