AI-COM · Approximate inference in graphical models for digital communications
FP6 — Marie Curie Actions (Human Resources and Mobility)
- Duration
- 2006-10-18 → 2009-10-17
- EU contribution
- €256,032
- Participants
- 2
- Scheme
- OIF
Lines connect the coordinator with its partners. CORDIS does not always give exact coordinates for projects before 2014. These points are placed at city or country level.
Results in brief
Final Activity Report Summary - AI-COM (Approximate Inference in Graphical Models for Digital Communications)
The goal of this project was to use new developments in approximate inference (artificial intelligence) for designing digital communication systems (AI-COM). The main three results have been: 1 Precoder optimisation: In ADSL and wireless systems we need to use a linear precoder to maximize the transmission rate. These precoders are designed to make the different channels independent. We have shown that correlating the inputs the transmission grows considerably, which can produce a considerable gain in the amount of information transmitted through any multiple-input multiple-output communication system. 2 Gaussian processes for designing digital communication receivers: Gaussian processes (GPs) are novel tools in machine learning for designing general detectors and estimators. We have shown that GP need shorter training sequence to achieve the same performance than other nonlinear detectors (e.g. neural networks or SVMs) and they provide accurate probability predictions, which are fundamental for the correct performance of communication receivers. 3 Joint source and channel coding: Low-Density Parity-Check (LDPC) codes have been shown to achieve channel capacity in the most interesting cases (Binary Erasure, Binary Symmetric and Gaussian channels), we have built on these codes to propose a code that compresses the information first with an LDPC code and then uses another to protect the source. We have shown that in the finite block length this joint scheme increases the transmission rate, compare to separate approach without an increase in the complexity. Ancillary results: We have shown a new method for estimating information theoretic quantities, as differential entropy, mutual information and the divergence, from samples. We have proven the convergence of our method using waiting times distributions.
Data: CORDIS, © European Union
Project objective
Information theory guides research in communication engineering and settle its theoretical limits. Advances in machine learning generative modelling have proven useful in providing state-of-the-art results in channel coding, one of the disciplines in information theory. The objective of this Marie Curie Fellowship proposal is to explore to what extent new algorithms in approximate inference for generative models can be beneficial not only to channel coding, but also to other information theory issues. In the last couple of years, there have been significant advances in solving the approximate inference problem in graphical models with loops.These algorithms improve the approximation function and divergence measure of belief propagation, the algorithm use d to address the channel-coding problem. Moreover, most communication problems (i.e. source coding, channel coding, equalisation, multiple access) can be expressed as inference problems in graphs with loops. Hence, the application of belief propagation extensions to information theory is a natural research direction that can benefit the description of communications systems.In this Fellowship we will focus on applying these extensions to open problems in information theory. Furthermore, we expect to draw conclusions about the performance of these approximate inference algorithms and improve them to meet our goals. The multidisciplinary nature of this fellowship proposes advances in information theory and machine learning.
Original text from CORDIS.
Participants
- UNIVERSITY CARLOS III DE MADRID · GETAFE (MADRID)CoordinatorCity levelSpain
- PRINCETON UNIVERSITY · PRINCETONUnited States
Links
Data: CORDIS, © European Union