Reduced-Complexity Greedy Distributed Information Bottleneck Algorithm

Publikationstyp
Conference Paper
Publikationsdatum
2021-07-11
Sprache
English
Author
Steiner, Steffen 
Kuehn, Volker 
Stark, Maximilian orcid-logo
Bauch, Gerhard 
Institut
Nachrichtentechnik E-8 
TORE-URI
http://hdl.handle.net/11420/10277
Start Page
361
End Page
365
Article Number
9513805
Citation
IEEE Statistical Signal Processing Workshop (SSP 2021)
Contribution to Conference
21st IEEE Statistical Signal Processing Workshop, SSP 2021 
Publisher DOI
10.1109/SSP49050.2021.9513805
Scopus ID
2-s2.0-85113555012
This paper addresses a distributed sensing scenario, widely known as the Chief Executive Officer (CEO) problem. Considering the logarithmic loss distortion measure, the distributed scalar compression can be optimized using an information bottleneck (IB) approach. The recently proposed Greedy Distributed IB (GDIB) algorithm optimizes all quantizers successively exploiting the statistics of previously designed quantizers as side-information. It was shown, that jointly optimizing the scalar quantizers results in a significant performance improvement compared to individual scalar optimization without side-information. However, processing the side-information becomes a major bottleneck as the memory complexity grows exponentially with the network size. This paper proposes a sequential compression scheme in order to compress this side-information to ensure feasibility even for larger networks. The compression is performed again by means of the information bottleneck method. Presented simulation results show that despite the compression of side-information the overall loss in relevant information can be made sufficiently small.