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

Hi Elman,

This newsletter dives into a collection of recent preprints showcasing exciting advancements in signal processing and machine learning across a variety of applications.

General Overview

This collection of preprints explores advancements in signal processing and machine learning across diverse applications, from wireless communications and sensing to biomedical engineering and material science. Several papers focus on enhancing the efficiency and performance of wireless systems. Wang and Liu (Wang et al., 2024) introduce an Autoencoder Gated Recurrent Unit (AEGRU) model for decoding motor movements from neural recordings, achieving state-of-the-art performance on Neurobench. Tong et al. (Tong et al., 2024) propose a novel symbol-level precoding scheme for multi-user MIMO systems with maximum likelihood receivers, demonstrating significant improvements over traditional block diagonalization methods. Two papers, Rajabalifardi et al. (Rajabalifardi et al., 2024) and Bhattacharya et al. (Bhattacharya et al., 2024), address the challenges of low-rank channels in fixed wireless access by introducing optimized power allocation and time-sharing algorithms for multi-carrier NOMA, achieving substantial gains in data rate and energy efficiency. Jin et al. (Jin et al., 2024) propose LinFormer, a lightweight transformer architecture for time-aware MIMO channel prediction, replacing the computationally intensive attention mechanism with a time-aware multi-layer perceptron (TMLP).

Another prominent theme is the application of advanced signal processing techniques for sensing and estimation. Jouni et al. (Jouni et al., 2024) present a comprehensive Bayesian framework for multiple-beam interference spectroscopy, offering a unified approach to spectrum reconstruction from interferometric measurements. Lin et al. (Lin et al., 2024) develop a joint parameter and channel tracking algorithm for LEO satellite communications, exploiting on-orbit characteristics to reduce pilot overhead. Wu et al. (Wu et al., 2024) propose location and map-assisted wideband phase and time calibration methods for distributed antenna systems, addressing the crucial challenge of synchronization in large-scale MIMO networks. Yu and Wu (Yu & Wu, 2024) introduce a novel target detection method by unfolding the CFAR detector with a state space model, combining the strengths of signal processing and deep learning. Sheng et al. (Sheng et al., 2024) propose a subset random sampling scheme for finite time-vertex graph signals, addressing the challenge of sampling and reconstruction with unknown spectral support.

Several contributions focus on specific applications and leverage diverse methodologies. Al-Nabulsi et al. (Al-Nabulsi et al., 2024) utilize bioimpedance and deep learning for the diagnosis of knee osteoarthritis. Omid et al. (Omid et al., 2024) employ deep reinforcement learning for downlink transmit precoding in satellite systems, mitigating the impact of delayed CSI. Chen et al. (Chen et al., 2024) introduce the D-Subspace algorithm for online learning over distributed networks. Berghi and Jackson (Berghi & Jackson, 2024) explore the use of reverberation and visual depth cues for sound event localization and detection. Mortaheb et al. (Mortaheb et al., 2024) propose rAge-k, a communication-efficient federated learning algorithm using age of information. Kang et al. (Kang et al., 2024) introduce PK-YOLO, a pretrained knowledge-guided YOLO model for brain tumor detection in MRI slices. Zhu and Zhou (Zhu & Zhou, 2024) investigate exponentially consistent statistical classification of continuous sequences with distribution uncertainty.

Further contributions explore novel applications of established techniques. Zheng et al. (Zheng et al., 2024) investigate the use of RIS in dual-polarized MIMO systems, analyzing the required surface size for performance gains. Li et al. (Li et al., 2024) propose a joint communications and jamming scheme for unauthorized UAV countermeasures in MIMO cellular systems. Shim et al. (Shim et al., 2024) apply labeled random finite sets for property estimation in geotechnical databases. Nguyen et al. (Nguyen et al., 2024) propose channel-coded precoding for multi-user MISO systems. Topal et al. (Topal et al., 2024) investigate precoding schemes for multi-target integrated sensing and communication in massive MIMO systems. Saeizadeh et al. (Saeizadeh et al., 2024) develop an AI-assisted agile propagation model for real-time digital twin wireless networks. Yang (Yang, 2024) proposes a novel algorithm to address the mismatch problem in compressed sensing.

Finally, several papers focus on methodological advancements in data processing and analysis. Yuan et al. (Yuan et al., 2024) introduce Adaptive NAD, a dynamic threshold-based two-layer online unsupervised anomaly detector. Yang et al. (Yang et al., 2024) present an ISAC prototype system for multi-domain cooperative communication networks. Nandi et al. (Nandi et al., 2024) introduce KALAM, a toolkit for automating the synthesis of analog computing systems. Fang and Yu (Fang & Yu, 2024) demonstrate centimeter-level geometry reconstruction and material identification in 300 GHz monostatic sensing. Jawaid et al. (Jawaid et al., 2024) propose using stochastic processes for surface data imputation. Wiame et al. (Wiame et al., 2024) introduce a novel approach to error correction decoding in interference-limited wireless systems using alpha-stable noise models. McKinney et al. (McKinney et al., 2024) present an unsupervised multimodal fusion method for in-process sensor data in advanced manufacturing. Chowdhury et al. (Chowdhury et al., 2024) introduce fast algorithms for hyperspectral neutron tomography, significantly reducing computation time and improving reconstruction quality. Chen et al. (Chen et al., 2024) derive Chernoff fusion equations for Bernoulli Gaussian max filters, addressing the challenge of track fusion with unknown statistical dependence. These diverse contributions highlight the ongoing advancements in signal processing and machine learning, pushing the boundaries of what's possible in various domains.

Paper Highlights

AI-Assisted Agile Propagation Modeling for Real-time Digital Twin Wireless Networks

AI-assisted Agile Propagation Modeling for Real-time Digital Twin Wireless Networks by Ali Saeizadeh, Miead Tehrani-Moayyed, Davide Villa, J. Gordon Beattie Jr., Ian C. Wong, Pedram Johari, Eric W. Anderson, Stefano Basagni, Tommaso Melodia https://arxiv.org/abs/2410.22437

Caption: This diagram illustrates the AI-driven real-time propagation modeling framework. It shows the training process of the U-Net model using elevation maps, rough propagation models from ray tracing, and the subsequent calibration using measurement data to achieve high-fidelity propagation predictions. The framework leverages 3D geo-information and ray tracing simulations to train a deep learning model capable of fast and accurate propagation estimation.

Accurate and real-time channel modeling is paramount for optimizing modern wireless communication systems, especially within the context of digital twins (DTs). Traditional methods, such as ray tracing and field measurements, struggle to keep pace with the demands of dynamic environments due to their computational intensity and cost. This paper introduces a groundbreaking AI-driven approach that utilizes deep learning to achieve real-time, high-fidelity propagation modeling, paving the way for more agile and adaptable wireless network design.

The researchers have developed a modified U-Net model trained on a comprehensive dataset generated using ray tracing simulations and empirical measurements. A key innovation of this model is its input strategy. Instead of relying on binary building maps and one-hot encoded transmitter locations like previous approaches, this model uses two primary inputs: an elevation map providing rich 3D geographical information, and a rough propagation estimate derived from fast, real-time ray tracing. This combination allows the model to effectively capture spatial features while maintaining generalizability across various scenarios. Furthermore, by strategically placing the transmitter at the center of the input, the model's complexity is reduced, leading to enhanced performance. The model's target output is the path gain (PG), defined as PG(t) = PRX(t) – PTX(t), where PRX(t) represents the received power and PTX(t) represents the transmitted power.

The model's training and testing were conducted using a dataset representing the Northeastern University campus in Boston. To further assess its generalizability, additional testing was performed on a completely different environment: Fenway Park. A rigorous 10-fold cross-validation process was employed to determine the optimal training split ratio. The results demonstrate a substantial improvement over traditional propagation estimation methods. The AI-driven model achieved a normalized Root Mean Squared Error (RMSE) of less than 0.035 dB over a sprawling 37,210 square meter area. Even more impressive is the model's speed: it achieved these results with processing times of just 46 ms on a GPU and 183 ms on a CPU, a stark contrast to the 387.6 seconds required by traditional high-fidelity ray tracing.

The researchers also demonstrated the model's adaptability to real-world data by refining it with a small set of actual measurements. This calibration process further reduced the median error to an impressive 0.0113 dB, highlighting the model's practicality for real-world deployments. This ability to incorporate and adapt to real-world measurements sets this approach apart from traditional ray tracing software, which typically lacks such flexibility. This AI-assisted approach marks a significant leap forward in real-time propagation modeling. Its speed, accuracy, and adaptability make it a powerful tool for creating and maintaining real-time digital twins of wireless networks. This capability will enable more efficient network design, optimization, and performance management in dynamic environments, ultimately enhancing user experiences in future wireless systems.

SleepNetZero: Zero-Burden Zero-Shot Reliable Sleep Staging With Neural Networks Based on Ballistocardiograms

SleepNetZero: Zero-Burden Zero-Shot Reliable Sleep Staging With Neural Networks Based on Ballistocardiograms by Shuzhen Li, Yuxin Chen, Xuesong Chen, Ruiyang Gao, Yupeng Zhang, Chao Yu, Yunfei Li, Ziyi Ye, Weijun Huang, Hongliang Yi, Yue Leng, Yi Wu https://arxiv.org/abs/2410.22646

Caption: This diagram illustrates the SleepNetZero architecture for zero-burden sleep staging using BCG signals. It shows the process of extracting heartbeat, breath, and body movement components from PSG and BCG data, applying data augmentation, and using a deep learning model (ResNet, Transformer, MLP) to predict sleep stages. The model is trained on PSG data and then applied to BCG data, bridging the sensor gap through novel augmentation techniques.

Sleep monitoring is essential for maintaining good health, with sleep staging being a fundamental metric in this process. While traditional methods utilizing medical sensors like EEG and ECG can be effective, they often present challenges such as an unnatural user experience, complex deployment procedures, and high costs. Ballistocardiography (BCG), which uses piezoelectric sensor signals, offers a non-invasive, user-friendly, and easily deployable alternative suitable for long-term home monitoring. However, reliable BCG-based sleep staging has been difficult due to the limited availability of sleep monitoring data specifically for BCG. A restricted training dataset hinders the model's ability to generalize across diverse populations. Furthermore, transferring models trained on other data sources to BCG presents challenges in ensuring robustness.

This paper introduces SleepNetZero, a zero-shot learning approach for sleep staging that directly addresses these challenges. To overcome the generalization problem, the authors propose a series of BCG feature extraction methods that align BCG components with corresponding respiratory, cardiac, and movement channels in polysomnography (PSG) data. This innovative approach allows models to be trained on extensive and diverse PSG datasets, benefiting from the wealth of available PSG data. To address the migration challenge, the authors employ data augmentation techniques, significantly enhancing the model's generalizability. The core of SleepNetZero is a sophisticated deep learning model comprising ResNet-based feature extractors for each component (heartbeat, breathing, and body movement), a Transformer encoder for capturing contextual information, and a linear classifier for final sleep stage prediction. The local representation $z_t$ is formed by:

$z_t = \text{concatenate}(f_1(\text{stack}(x_1^{(t)}, x_3^{(t)})), f_2(\text{stack}(x_2^{(t)}, x_3^{(t)})), f_3(\text{stack}(x_1^{(t)}, x_2^{(t)}, x_3^{(t)})))$

where $x_1$ and $x_2$ are continuous signals representing heartbeat and breath, respectively, and $x_3$ is a discrete signal representing body movement, and $f_1, f_2, f_3$ are the feature extractors.

The model was rigorously trained and tested on a large combined dataset (NSRR) comprising 12,393 records from 9,637 different subjects, drawn from multiple publicly available PSG datasets. The results on the NSRR test set were impressive, achieving an accuracy of 0.803 and a Cohen's Kappa of 0.718. Importantly, to evaluate real-world performance, SleepNetZero was deployed in a physical prototype (monitoring pads) and tested in an actual hospital setting with 265 users. In this real-world scenario, the model achieved an accuracy of 0.697 and a Cohen's Kappa of 0.589, demonstrating its practical viability. Further analysis showed that the model maintained its robustness across diverse demographics and varying levels of sleep apnea severity. While the current model faces challenges in accurately identifying N1 sleep stages due to inherent low annotation consistency and data imbalance in this stage, the overall results are highly promising. Future research could explore self-supervised learning on larger, unannotated BCG datasets to further improve performance and expand the application to sleep disorder diagnosis.

Channel-Coded Precoding for Multi-User MISO Systems

Channel-Coded Precoding for Multi-User MISO Systems by Ly V. Nguyen, Junil Choi, Bjorn Ottersten, A. Lee Swindlehurst https://arxiv.org/abs/2410.22640

Caption: This figure showcases the constellation points of received symbols for a multi-user MISO system using 4-QAM. The proposed Channel-Coded Precoding (CCP, solid symbols) demonstrates improved clustering around the ideal constellation points compared to conventional Symbol-Level Precoding (SLP, transparent symbols), indicating a lower bit error rate and highlighting CCP's ability to leverage channel coding for enhanced performance. The arrows indicate how erroneous symbols with CCP are closer to the correct symbol than with SLP, making them more likely to be corrected by the channel code.

Traditional precoding techniques primarily focus on minimizing the symbol error rate (SER), often neglecting the powerful error-correcting capabilities of channel codes. This paper introduces a paradigm shift in precoding with Channel-Coded Precoding (CCP), a novel framework designed to directly minimize the information bit error rate (BER) in multi-user multiple-input single-output (MISO) systems. CCP cleverly leverages the inherent error correction provided by channel codes, allowing for some data symbol errors at the receiver, as long as the overall information BER is improved. This innovative approach expands the degrees of freedom in transmit signal design, leading to enhanced system performance.

The CCP framework is developed for both one-bit (εc = 1) and multi-bit (εc > 1) error-correcting capacities. For the one-bit case (εc = 1), closed-form expressions for the probability of correctly recovering information bits are derived for various QAM constellations. A projected gradient (PG)-based algorithm is employed to optimize the CCP design criterion: max{X} min_k P[||c_k - ĉ_k||_0 ≤ ε_c | X] subject to ||X||_F^2 ≤ P, where P represents the probability, c_k and ĉ_k denote the transmitted and estimated codewords respectively, and P represents the total transmit power constraint. For the multi-bit case (εc > 1), the transmission block is strategically divided into sub-blocks, and the probability of having no more than one bit error per sub-block is maximized. This simplification makes the optimization problem more tractable while still benefiting from the code's error-correcting capabilities. Furthermore, a robust CCP framework is developed to address scenarios with imperfect channel state information (CSI), explicitly taking into account the effects of both noise and channel estimation errors.

Simulation results clearly demonstrate the superiority of CCP over conventional precoding methods such as Maximum Ratio Transmission (MRT), Zero Forcing (ZF), Minimum Mean Square Error (MMSE), and Symbol-Level Precoding (SLP). For example, with a repetition code of length 3, K=N=4, and 8-QAM modulation, CCP achieves a significantly lower BER compared to SLP. In a scenario with εc = 1, CCP provides a remarkable gain of over 3dB compared to SLP. Further analysis reveals that CCP's performance gains are most pronounced with lower code rates and in heavily loaded systems where the number of users approaches the number of transmit antennas. For εc > 1, the results show that there exists an optimal sub-block length that balances the trade-off between exploiting the error-correcting capacity and limiting the probability of exceeding it. The robust CCP design also exhibits significant improvements over the non-robust version in the presence of CSI errors, achieving a BER reduction of approximately 2dB. This work highlights the significant potential of incorporating channel coding knowledge into precoding design. By strategically allowing for correctable symbol errors, CCP unlocks new degrees of freedom, enabling more efficient transmit signal designs. This information-directed precoding framework represents a promising avenue for future research, paving the way for more sophisticated joint optimization of channel coding and precoding strategies.

Conclusion

This newsletter highlights the innovative applications of signal processing and machine learning within diverse fields. From enhancing wireless communication through AI-driven propagation modeling and channel-coded precoding, to revolutionizing healthcare with zero-burden sleep staging using BCG signals, these advancements showcase the transformative potential of these technologies. The development of real-time digital twin networks, facilitated by AI-assisted propagation modeling, promises to revolutionize network optimization and management. Simultaneously, the introduction of Channel-Coded Precoding signifies a paradigm shift in precoding design, leveraging the power of channel codes to enhance BER performance in multi-user MISO systems. Furthermore, the advent of SleepNetZero demonstrates the potential of BCG-based sleep staging to provide a convenient and accurate alternative to traditional methods, enabling widespread access to sleep monitoring. These cutting-edge developments underscore the ongoing convergence of signal processing, machine learning, and various application domains, paving the way for a future where intelligent systems can address complex challenges and improve quality of life across multiple sectors.