Distributed Compression using the Information Bottleneck Principle
Steffen Steiner, Volker Kuehn · 2021
This paper considers a scenario being widely known as the Chief Executive Officer (CEO) problem under log-loss distortion. It consists of a set of distributed sensors observing statistically dependent data and forwarding their measurements to a common receiver. As these links are capacity limited, the sensors have to compress or quantize the data. This paper contributes an algorithm for the quantizer design based on the information bottleneck (IB) method. This algorithm successively designs scalar quantizers using the statistics of other quantizers as side-information. Furthermore, it allows individual compression rate adjustments in order to fulfill given rate-constraints of the capacity limited links. Numerical results demonstrate the benefit compared to individual scalar optimized quantization.