Home
Speakers
Schedule
Location and Directions
Hotel Information
Prakash Ishwar
Boston University
Title: Distributed Coding for Interactive Computing
Abstract: In wireless sensor network systems where distributed inferencing and control needs to be performed, the raw-data generated at different nodes needs to be transformed and combined in a number of ways to extract actionable information. This requires performing distributed computations on the raw-data. A pure data-transfer solution approach would advocate first reliably reproducing the raw-data at decision-making nodes and then performing suitable computations to extract actionable information. Two-way interaction and statistical correlation among source, destination, and relay nodes, would be utilized, if at all, to primarily improve the reliability of data-reproduction than overall computation-efficiency. However, to maximize the overall computation-efficiency, it is necessary for nodes to interact bidirectionally, perform computations, and exploit statistical correlation in data as opposed to only generating, receiving, and forwarding data. In this talk we will explore this common wisdom through some examples of distributed function-computation problems with the goal of minimizing the total number of bits exchanged per unit time. We will highlight the role of interaction in computation-efficiency within an information-theoretic setting involving block-coding asymptotics and vanishing probability of function-computation error. We will discuss information-theoretic characterizations of the set of feasible coding-rates for these problems and explore the fascinating interplay of function-structure, correlation-structure, and interaction.