TU Berlin

Fachgebiet Theoretische Grundlagen der KommunikationstechnikNon-Negative Structured Regression

Schriftzug CommIT


zur Navigation

General Project Information

  • DAAD program: Subject-Related Partnerships with Institutions of Higher Education in Developing Countries
  • Project Title: Non-Negative Structured Regression with Applications in Communication and Data Science
  • Cooperation partner:  

Dr. Bubacarr Bah

German Research Chair of Mathematics with specialization in Data Science

African Institute for Mathematical Sciences (AIMS) South Africa

Applied Mathematics, Stellenbosch University:

Website: https://aims.ac.za/research/data-science-and-information-systems/dr-bubacarr-bah/

  • Project managers:

Prof. Giuseppe Caire, Ph.D. and Dr. Peter Jung

Communications and Information Theory Group

Electrical Engineering and Computer Science

Technical University of Berlin

Project summary

In this project we propose to design efficient algorithms for the reconstruction of redundant-encoded signals in wireless communication and network properties in data science. The main motivation of this line of work comes from model-based compressed sensing (CS) with non-negativity priors. CS is based on the fact that the intrinsic dimension of many digital signals or large data sets is typically far less than their ambient dimensions, for example the sparse representation of images, videos, audio data, network status information like activity and novel coding techniques for wireless communication. Traditionally, redundancy and structure in the data is exploited after the acquisition (measurement), which may be very costly in terms storage and bandwidth. CS instead attempts to overcome this by performing sampling and compression simultaneously, i.e., acquisition from a sub-Nyquist perspective. CS works well with provable guarantees for dense matrices. However, in communication engineering and data science problems related to complex networks structured sparse matrices are used due to more efficient storage and processing. Therefore we focus also on binary and sparse matrices, formed for example by sampling Hadamard matrices, expander matrices, etc. For such type of matrices CS often reduces to linear sketching, which has been applied in data streaming, and graph sketching. Furthermore, sparsity and compressibility can be regarded as first order structure of signals and objects of interest. In practice, most objects of interest exhibit second other structures like block-sparsity, tree-sparsity, non-negativity, etc. Both, the communication and complex networks problems we consider have often a conic constraint like non-negativity, hence the title of the project.



Schnellnavigation zur Seite über Nummerneingabe