Subject: Cutting-Edge Advancements in Signal Processing, Communication, and Machine Learning

Hi Elman,

This newsletter covers recent exciting developments in signal processing, communication systems, and machine learning, emphasizing data-driven approaches for enhanced performance and robustness.

General Overview

This collection of preprints explores advancements in signal processing, communication systems, and machine learning applications, with a notable focus on leveraging data-driven approaches for enhanced performance and robustness. Several studies delve into the intricacies of channel estimation and beamforming, particularly in the context of massive MIMO and reconfigurable intelligent surfaces (RIS). Zhang et al. (2024) investigate 6DMA-aided hybrid beamforming, demonstrating significant sum-rate improvements through joint optimization of antenna position, orientation, and beamformers. Gomes et al. (2024) tackle the challenge of non-reciprocal RIS-assisted MIMO channel estimation, proposing a novel three-phase protocol and Tucker decomposition-based algorithm. Complementing these efforts, Qin & Zhu (2024) provide an optimal error analysis for IRS-assisted MIMO channel estimation, establishing theoretical guarantees for stable recovery of channel matrices. These works collectively advance the understanding and practical implementation of RIS technology for future wireless communication systems.

Another prominent theme is the application of deep learning to various signal processing tasks. d'Ascoli et al. (2024) achieve remarkable results in decoding individual words from non-invasive brain recordings, leveraging a novel deep learning pipeline and a large dataset. In the realm of medical imaging, Friot--Giroux et al. (2024) assess deep learning methods for enhancing low-dose dental CBCT volumes, demonstrating the potential of 3D U-Nets for artifact reduction and detail preservation. Several studies also explore the use of deep learning for fault diagnosis, including Ali et al. (2024)'s weighted probability ensemble deep learning approach for induction motors and Sadek et al. (2024)'s multi-classification of high-frequency oscillations using iEEG signals. These contributions highlight the versatility and effectiveness of deep learning across diverse applications.

Beyond deep learning, several innovative signal processing techniques are introduced. Chan et al. (2024) propose a generalization of the Hilbert transform on graphs, enabling the exploitation of phase information in graph signals. Weiss et al. (2024) introduce extremum encoding for joint baseband signal compression and time-delay estimation in distributed systems. Weiss (2024) further investigates robustness in blind modulo analog-to-digital conversion. These works contribute novel methodologies to address specific challenges in signal processing.

The intersection of communication and sensing is also explored, particularly in the context of integrated sensing and communication (ISAC). Delamou et al. (2024) propose a joint adaptive OFDM and reinforcement learning design for autonomous vehicles, leveraging age of updates to optimize both communication and sensing performance. Elrashidy et al. (2024) investigate unsupervised learning for beamforming in cell-free ISAC. Khosroshahi et al. (2024) focus on improving localization accuracy in multistatic ISAC using 5G NR signals. These contributions highlight the growing interest in developing efficient and robust ISAC systems.

Finally, several papers address practical applications and system-level considerations. Prod'homme & del Hougne (2024) present an updatable closed-form evaluation method for complex multi-port network connections. Pavurala et al. (2024) develop an adaptive signal analysis method for automated subsurface defect detection. Luo et al. (2024) explore the practical implementation of underwater acoustic reconfigurable intelligent surfaces. These diverse applications demonstrate the practical impact of signal processing and communication research across various domains. Overall, these preprints showcase significant advancements in signal processing, communication, and machine learning, pushing the boundaries of theoretical understanding and practical implementation.

Paper Highlights

Decoding Words Directly from Brain Waves

Decoding individual words from non-invasive brain recordings across 723 participants by Stéphane d'Ascoli, Corentin Bel, Jérémy Rapin, Hubert Banville, Yohann Benchetrit, Christophe Pallier, Jean-Rémi King https://arxiv.org/abs/2412.17829

Caption: This image visually represents the methodology and data used in a study on decoding words from brain waves. It details the datasets used, organized by modality (listening/reading), device (EEG/MEG), and language, along with the model architecture incorporating CNNs, subject-specific layers, a transformer, and contrastive learning against a vocabulary.

This research introduces a groundbreaking deep learning model that decodes individual words from non-invasive brain recordings (EEG and MEG) with remarkable accuracy. This achievement addresses the long-standing challenge of extracting precise word-level information from noisy brain signals, surpassing previous limitations of invasive methods or sentence-level decoding. The model's training and evaluation involved an unprecedentedly large dataset of 723 participants exposed to five million words across English, French, and Dutch, encompassing both listening and reading tasks.

The model's success stems from its innovative use of contrastive learning to establish mappings between brain activity and semantic word representations, leveraging a pre-trained multilingual language model. A subject-specific layer accounts for inter-subject variability, while a transformer layer incorporates crucial contextual information from surrounding words.

The results demonstrate a significant advancement in non-invasive word decoding, consistently outperforming existing methods like linear models and standard convolutional networks (e.g., EEGNet). Remarkably, the model exhibits "zero-shot" decoding, predicting words not present in the training data, indicating its ability to learn semantic features rather than simply memorizing patterns. The study also highlights the influence of experimental design, with MEG outperforming EEG and reading tasks proving easier to decode than listening. Crucially, collecting more data per participant yielded greater benefits than increasing the number of participants with limited data. Performance also scaled with both the training data size and the number of test words used for averaging predictions, suggesting further improvements with larger datasets and enhanced signal processing.

Analysis of the model's predictions reveals its reliance on word semantics, with incorrect predictions often semantically similar to the target word. However, the model also captures syntactic and surface properties (part-of-speech, word length, and even individual letters), particularly during reading tasks. This multi-faceted decoding strategy likely contributes to the model's robust performance across languages and tasks. This is further supported by the statistically significant observation that incorrect predictions share the same parts-of-speech and word length as the true word more often than chance (p < 0.005). This work represents a significant stride towards practical, non-invasive BCIs for natural language processing.

Underwater Acoustic Reconfigurable Intelligent Surfaces

Underwater Acoustic Reconfigurable Intelligent Surfaces: from Principle to Practice by Yu Luo, Lina Pu, Junming Diao, Chun-Hung Liu, Aijun Song https://arxiv.org/abs/2412.17865

Caption: This diagram illustrates the architecture of a 1-bit phase coding Underwater Acoustic Reconfigurable Intelligent Surface (UA-RIS). Each unit (U_k) incorporates PZT pads sandwiched between front and tail masses, connected to a load network (N_k) controlled by a microcontroller (MCU) via an I/O extender using I²C communication. This setup allows for dynamic phase switching (0° or 180°) of reflected acoustic waves, offering a more energy-efficient alternative to conventional PZT cylinder transducers for enhanced underwater communication.

This research explores the innovative application of Reconfigurable Intelligent Surfaces (RIS) to underwater acoustic (UA) communication, aiming to overcome the limitations of traditional methods in achieving high data rates over long distances while minimizing environmental impact. Unlike their radio frequency counterparts, UA-RIS must address the unique challenges posed by the physics of acoustic waves, including lower frequencies, limited power availability, and potential harm to marine life. This necessitates a distinct approach to architecture and design principles.

The authors propose a novel 1-bit phase coding UA-RIS architecture, utilizing a load network to dynamically control the phase of reflected acoustic waves (0° or 180°) by switching NMOS transistors connected to each reflection unit. This offers improved flexibility and energy efficiency compared to varactor-based systems commonly used in RF-RIS, which are unsuitable for the lower frequencies and higher impedance of underwater acoustics. The prototype UA-RIS consists of 24 acoustic elements in a 6x4 grid, each capable of independent phase switching.

Experimental validation in both tank and lake environments demonstrated the UA-RIS's ability to manipulate acoustic waves. Tank tests showed that aligning 12 active reflection units in phase increased the received signal's peak-to-peak voltage to 1.7V, while inverting the phase reduced it to 0.2V, demonstrating effective signal control. Lake tests further confirmed this, with received signal strength varying significantly (280mV, 100mV, and 33mV) based on the UA-RIS coding and receiver angular deviation.

While these results showcase the feasibility of manipulating acoustic waves using UA-RIS, the authors acknowledge limitations and outline future research directions. More controlled environments are needed for quantitative performance analysis, and hardware assembly inaccuracies must be addressed. Future work will focus on large-scale deployments and higher operational frequencies, requiring redesigned reflection units and optimized system components. This research marks a crucial step towards integrating UA-RIS into practical underwater communication systems, enabling high-rate, long-range transmission with minimal environmental impact.

Joint Downlink-Uplink Channel Estimation for Non-Reciprocal RIS-Assisted Communications

Joint Downlink-Uplink Channel Estimation for Non-Reciprocal RIS-Assisted Communications by Paulo R. B. Gomes, Amarilton L. Magalhães, André L. F. de Almeida https://arxiv.org/abs/2412.16301

This paper addresses the critical challenge of channel estimation (CE) in non-reciprocal RIS-assisted MIMO systems, where channel characteristics differ between downlink (DL) and uplink (UL) transmissions. Existing approaches often rely on the simplifying assumption of channel reciprocity, which doesn't hold in practice due to hardware limitations. This work tackles the joint DL and UL CE problem, acknowledging the distinct nature of channels in both directions.

The authors introduce a novel three-phase closed-loop CE protocol that shifts the computational burden to the base station (BS), freeing the resource-constrained user terminal (UT). In Phase 1, the BS transmits pilot sequences while the DL RIS employs varying responses. Phase 2 involves the UT applying simple linear coding to the received signal before transmission back to the BS. Finally, in Phase 3, the BS receives the UL signal reflected by the RIS with varying UL scattering patterns. This protocol enables the BS to gather sufficient information for joint DL and UL CE without demanding complex processing at the UT.

A two-stage CE algorithm based on fourth-order Tucker decomposition is proposed. Stage 1 uses trilinear alternating least squares (TALS) to iteratively estimate the non-reciprocal BS-RIS channels (H_d and H_u). Stage 2 employs Khatri-Rao factorization (KRF) to extract the RIS-UT channels (G_d and G_u) from the estimated structured matrix. This two-stage approach effectively decouples the estimation problem, facilitating accurate recovery of all four non-reciprocal channels.

Simulations demonstrate the superior performance of the proposed method compared to traditional FDD LS-based and tensor-based techniques, achieving significantly lower normalized mean squared error (NMSE). For example, in a specific system configuration, the proposed method showed an NMSE improvement of approximately two orders of magnitude compared to LS-KRF with imperfect feedback. The ability to accurately estimate individual non-reciprocal channels, even with error propagation between stages, further highlights the method's robustness. This accurate and efficient CE approach paves the way for optimizing beamforming and other crucial aspects of non-reciprocal RIS-assisted communication.

Conclusion

This newsletter highlights significant progress in signal processing and communication. From decoding words from brain waves using deep learning to enhancing underwater acoustic communication with reconfigurable intelligent surfaces and tackling the intricacies of non-reciprocal channel estimation in RIS-assisted MIMO systems, these works push the boundaries of both theoretical understanding and practical implementation. The common thread is the innovative use of data-driven approaches to overcome limitations and enhance performance in various applications. These advancements hold immense potential for shaping the future of communication and signal processing technologies.

Best,

[Your Name/Newsletter Name]