DEIPRO: Distributed lEarning, communication and Information PROcessing (TEC2009-14504-C02)


With the current technology trends, communication networks are evolving towards ever-complex systems consisting of a large number of heterogeneous nodes that enjoy enhanced capabilities for sensing, storing, processing and transmitting data in many sophisticated forms. This project deals with the design, analysis and assessment of distributed methods for efficiently extracting, processing and transmitting information in this new framework.

Our main goal is to advance towards the solution of a number of fundamental problems that arise in this scenario, both in terms of new formal methodologies and numerical techniques and in the demonstration of their validity by means of an adequate hardware platform. We will make intensive use of statistical signal processing, machine learning and information theory techniques. The basic problems that will be tackled and the specic goals to achieve for each one of them are briefly introduced below.

  • Compressed sensing (CS): development of new methods for the selection of (non random) projection matrices and the design of sparse signal reconstruction methods for distributed CS.

  • Local processing: development of new multitask learning classifiers by discriminative methods and new multitask learning regressors by generative (gaussian processes) methods.

  • Combination of classifiers for correlated observations: development of new local multi-input and global graphbased classifiers.

  • Distributed inference: development of distributed methods for detection, localization and tracking, suitable to work under realistic communication constraints and with special attention to the cases of unreliable or correlated measurements, by applying tools from classical detection and estimation, machine learning, information theory, matrix theory and sequential Monte Carlo approximation. Design and analysis of a methodology for inference in high-dimensional dynamic systems by combining the Bayesian framework with sequential Monte Carlo methods to deal with analytically intractable computations.

  • Global and dynamic optimization: elaboration of a general methodology to recast a broad family of highdimensional global and dynamic optimization problems into fully equivalent estimation and tracking problems for state-space systems, wherein sequential inference techniques can be exploited to achieve accurate solutions with tractable computational complexity.

  • Communications: development of new precoding, equalization and decoding methods exploiting the relationship between estimation and information theories, and inference in graphical models.

  • Coding: obtention of novel codes for joint source and channel coding, and of optimal and suboptimal methods for cooperative detection in mesh networks.

  • Performance analysis of the previous methods: asymptotic analysis of information extraction and communication methods using large deviation theory and other techniques, and modeling the uncertainty in likelihood ratio tests using both Bayesian and non-parametric (confidence interval based) frameworks.

The applicability of the obtained results will be evaluated on a multipurpouse hardware platform, configured for four different scenarios: indoor navigation, domestic monitoring, security in large-scale indoor facilities and communications.