Avis de Décès d’André Lannes
Chers collègues, C’est avec une grande tristesse que nous faisons part du décès d’André Lannes...
Nous vous rappelons que, afin de garantir l'accès de tous les inscrits aux salles de réunion, l'inscription aux réunions est gratuite mais obligatoire.
38 personnes membres du GdR ISIS, et 21 personnes non membres du GdR, sont inscrits à cette réunion.
Capacité de la salle : 80 personnes.
Méthodes bayésiennes approximées et variationnelles pour la détection, l'estimation et le décodage
Le but de cette journée est de dresser un panorama des méthodes issues de l'inférence bayésienne approximée et variationnelle pour des applications en détection (MIMO, NOMA, multi-utilisateurs, OFDM, égalisation mono-porteuse), estimation et décodage canal.
Ces méthodes, comme par exemple expectation ou belief propagation, AMP/OAMP et leurs dérivés, offrent un riche éventail de méthodes qui ont montré leur très bon compromis performance- complexité. De plus, ces méthodes s'étendent naturellement à un contexte d'apprentissage machine dans lequel de nombreuses extensions ont été proposées (par exemple, versions dépliées profondes ou basées sur les concepts de "graph networks").
L'objectif de cette journée est, pour les différents acteurs académiques et industriels, d'échanger autour de ces méthodes et de leurs applications en télécommunications, de présenter leurs contributions récentes et éventuellement de convenir d'actions de collaboration.
La journée s'articulera autour des évènements suivants :
- un "tutorial" qui permettra d'avoir une vue synthétique du domaine
- une série de présentations orales d'acteurs académiques et industriels, juniors (doctorants/post-doctorants) ou confirmés.
(Attention changement d'amphithéatre par rapport aux précédentes annonces)
La journée se déroulera finalement dans l'amphithéâtre Fabry-Pérot, accès 4,
au CNAM Paris, 292 rue Saint Martin, 75003 PARIS.
jeudi 28/09/2023, de 9H00 à 17h00.
L'inscription à cette journée est gratuite mais obligatoire sur le site du GdR ISIS.
Participation à distance sous Teams:
https://teams.microsoft.com/l/meetup-join/19%3ameeting_MjI2Zjg1NzctMjhhMS00MzZmLWFlOTQtYzIyMDMwMTU1M2M5%40thread.v2/0?context=%7b%22Tid%22%3a%22b323bcb4-6d58-4f25-87bf-6366c3d689af%22%2c%22Oid%22%3a%2299b009eb-eba9-4c53-a381-f31bd7755556%22%7d
ID de la réunion : 382 266 711 851
Code secret : mQ7Kjf
- Pascal Chevalier, CNAM - CEDRIC,
- Romain Tajan, Bordeaux INP - IMS,
- Antonio Cipriano, THALES SIX GTS,
- Charly Poulliat, Toulouse INP - IRIT.
Organisation de la journée :
"On the advances in message-passing algorithms and practical iterative receiver design."
"Great Expectations of Expectation-Propagation (lights and shadows in the EP algorithm applied to Communications Engineering)"
"When Bayes meets Kullback-Leibler: a Tale of Message Passing and Alternating Optimization"
"Sparse-DFT and WHT Precoding with Iterative Detection for Highly Frequency-Selective Channels"
"On low complexity message passing algorithms for data detection"
"Expectation Propagation on Modifed Factor Graphs using Matrix Decomposition"
"On the advances in message-passing algorithms and practical iterative receiver design."
This is a two part tutorial, with the first part covering the notable theoretical developments on variational Bayesian inference and message passing algorithms used in signal processing. These techniques are highly versatile for designing iterative processes suited for practical low-complexity inference in a wide range of signal processing fields. The tutorial covers the evolution of fundamental techniques like belief propagation (BP) and mean field (MF) up to the recently emerging approximate message passing (AMP) algorithms and neural-network aided inference techniques.
"Great Expectations of Expectation-Propagation (lights and shadows in the EP algorithm applied to Communications Engineering)"
Approximate message passing and variational inference methods have become increasingly popular, in the last twenty years, to solve communications engineering tasks. Since the landmark paper by Kschischang, Frey and Loeliger in 2001, a number of detection, estimation and decoding problems have been tackled with the Sum-Product Algorithm, the Loopy Belief Propagation and other approximate message passing algorithms, sometimes by mixing two or more of them within the same factor graph.
Among the approximate varational inference techniques, Expectation Propagation (EP) - first proposed by Thomas Minka in 2001 - has been successfully applied to a variety of telecom problems, demonstrating its potentials and flexibility. Albeit, in some scenarios, the straightforward application of EP can hardly succeed without proper adjustment strategies.
In this talk, we shall revise the experience of our research group in appplying EP to a number of different transmission systems, affected by different impairments. We shall highlight some reasons for the strength and potentials of EP as well as for its weak sides; in the attempt to shed further light on this promising variational Bayesian method.
In the areas of communications and compressed sensing, the demand for effective approximate Bayesian estimation techniques is paramount. Sparse channel modeling extends traditional model selection, enabling optimized models based on available training data. Compressed sensing techniques extend Linear Minimum Mean Squared Error (LMMSE) estimation by a hierarchical Bayesian formulation. In multi-user detection or blind channel estimation, going beyond LMMSE and Gaussian models represents a leap.
One of the approaches in the realm of approximate Bayesian estimation is Variational Bayes (VB), a relatively straightforward method. VB can be seen as an extension of the Expectation-Maximization (EM) technique to scenarios involving random parameters, thereby yielding not only point estimates but also approximate posterior distributions. Notably, while VB yields accurate means in Gaussian problems, it tends to underestimate variances significantly.
An even more refined technique for approximate Bayesian estimation is Expectation Propagation (EP). Both VB and EP share the underlying concept of minimizing the Kullback-Leibler Divergence (KLD), albeit with different sequencing of the true and approximating probability density functions. However, EP is a heuristic approach to minimizing a more desirable KLD, which is called the Bethe Free Energy (BFE). Exact alternating constrained minimization of the BFE leads to Belief Propagation (BP), whereas EP deviates from the alternating cost function and furthermore restricts approximating pdfs to be in an exponential family.
Taking a fresh look at alternating minimization of a KLD, the Central Limit Theorem leads to Gaussianity of the extrinsics in the marginal posteriors in moderate asymptotic settings. This in turn leads to what we call Gaussian Extrinsic Propagation, which sheds new light on characterizing performance beyond the loose Bayesian Cramer-Rao bound. Focusing on the Generalized Linear Model, assuming a n.i.i.d. (sign) statistical model for the measurement matrix allows asymptotically to find the variances without matrix inversions. This leads to Approximate Message Passing (AMP) in which the Onsager correction term w.r.t. Jacobi iterations for solving the normal equations for the mean is related to the Componentwise Conditionally Unbiased MMSE estimation.
Reformulating AMP to correspond to alternating minimization of an asymptotic version of the BFE leads to a provably convergent algorithm.Alternating minimization becomes tricky in the presence of constraints and we shed some light on the desirable behavior of the Alternating Directions Method of Multipliers (ADMM) approach. Extending the class of measurement matrices to Haar distributed unitary factors in the SVD allows to model more ill-conditioned problems, typically handled via the Vector AMP algorithm which assumes uniform variance profiles. To get correct individual variances is possible with a Unitary AMP in which the AMP variance predictions need to be corrected based on Haar Large System Analysis.
We hope that this exploration of advances will inspire to extend the scope of problems tackled by these techniques, ultimately paving the way for enhanced Bayesian estimation methodologies.
"Sparse-DFT and WHT Precoding with Iterative Detection for Highly Frequency-Selective Channels"
Channel fading poses a challenge in wireless communication, demanding effective strategies to combat its effects. Precoding techniques, such as those based on the discrete Fourier transform (DFT) and Walsh-Hadamard transform (WHT), have been used for spreading data symbol energy across time and/or frequency domains. To harness diversity and address inter-symbol-interference (ISI), iterative receivers play a pivotal role. In this talk, we introduce a new maximum a-posteriori (MAP) detection method tailored for SWH precoding. This approach offers superior performance and complexity trade-offs compared to state-of-the-art expectation propagation (EP) based receivers for DFT precoding, with notable advantages for QPSK and 16-QAM, while having complexity limitations for higher QAM orders. Lastly, we discuss the sparse-DFT implementation, which offers better complexity and performance than the DFT system with an EP-based receiver.
"On low complexity message passing algorithms for data detection"
Since the complexity of the classical message passing algorithm grows exponentially with the constellation size, many researchers have studied low complexity algorithms with a good trade-off between complexity and performance. In this talk, we will first review different low complexity message passing algorithms that have been proposed in the litterature mainly focusing on Gaussian Approximation approaches where the exchanged messages between the variable and function nodes are approximated by Gaussian distribution. We will then show that those solutions can be efficiently applied for data detection in different applications including Sparse Code Multiple Access (SCMA) and Orthogonal Time Frequency Space (OTFS) modulation.
"Expectation Propagation on Modifed Factor Graphs using Matrix Decomposition"
This presentation focuses on Expectation Propagation (EP) on modified factor graphs, through observation pre-processings with matrix decompositions. TThe application of EP on graphs modified by QR decomposition (QRD), Golub-Kahan Decomposition (GKD), and Singular Value Decomposition (SVD) pre-processing for MIMO detection yields vastly different messages and detectors. A discussion on the possible schedulings of message exchange is presented as well as an overall performance-to-complexity trade off.
In the context of wireless massive machine-type communications (mMTC) and ultra-reliable low latency communications (uRLLC), the development of grant-free protocols to support random access is crucial. To this end, a key problem that must be addressed is Active User Detection and Channel Estimation (AUDaCE). The active user detection problem consists in the identification of the subset of users requesting an access to the network based on the observation of their received faded and noisy pilot sequences; the channel estimation problem aims at estimating the baseband channel coefficients between each active user and the base station's antennas. In order to jointly solve these problems, tools from the bayesian compressed sensing literature have become very popular. More precisely, a mix between belief propagation (BP) and generalized approximate message passing (GAMP), namely Hybrid GAMP (HGAMP), is leveraged to perform such a task. Furthermore, the active user detection problem is studied under the assumption of heterogeneous correlated activity where users are likely to request access to the network together. The HGAMP algorithm is then built taking into account this aspect, which is particularly relevant for mMTC networks. Extensive Monte-Carlo simulations show the gains in terms of user detection and channel estimation with this approach against other state-of-the-art bayesian algorithm that do not account for correlated user activity.
"EP-based receivers for Faster-Than-Nyquist Signaling"
Faster-Than-Nyquist (FTN) Signaling is a symbol-time compression technique that allows transmitting data at higher throughput than with traditional Nyquist communications, at the cost of inflicting increased Inter Symbol Interference (ISI), to be handled at the receiver. This presentation focuses on EP-based receivers for FTN signaling. In particular, time-domain linear, widely linear and frequency-domain block-wise equalizers are compared in terms of performance (Bit Error Rate, BER) and computational complexity. Although frequency-domain EP receivers have low complexity (quasi-linear with the number of transmitted symbols), their performance is negatively affected when the FTN compression factor becomes too small. In contrast, time-domain block and widely linear solutions converge faster, but at the cost of higher computational complexity (cubic with the size of the transmitted symbols). The final part of the presentation deals with improving the performance / complexity trade-off of time-domain block solutions, by evaluating methods for reducing the complexity.