Subject: Cutting-Edge Advancements in Communication and Signal Processing
Hi Elman,
General Overview
This newsletter explores recent advancements across various domains within communication and signal processing, with a particular focus on the integration of artificial intelligence (AI) and novel signal processing techniques. Several works concentrate on enhancing the robustness and efficiency of wireless communication systems. For instance, Wang et al. (2024) Wang et al. (2024) introduce an XAI-based defense mechanism against adversarial attacks in automatic modulation classification (AMC), leveraging SHAP values to identify and neutralize negative information within adversarial samples, subsequently fine-tuning the model for improved robustness. Von Bank et al. (2024) von Bank et al. (2024) address decoding efficiency by proposing a spiking neural network (SNN) based belief propagation decoder for LDPC codes, achieving performance comparable to normalized sum-product decoding while exhibiting robustness to SNR mismatch. In the realm of spectrum sensing, Song & Gao (2024) Song & Gao (2024) present GBSense, a GHz-bandwidth compressed spectrum sensing system employing periodic nonuniform sampling for efficient wideband signal capture.
Another key theme is the fusion of diverse data sources and advanced signal processing for enhanced system performance. Guo et al. (2024) Guo et al. (2024) develop an RSSI-assisted CSI-based passenger counting system using multiple Wi-Fi receivers, achieving improved accuracy through a novel CSI feature fusion module. Anand et al. (2024) Anand et al. (2024) investigate EEG-based turn intention detection, demonstrating high accuracy in classifying movement intentions using statistical measures and Hjorth parameters extracted from preprocessed EEG signals. Khalil et al. (2024) Khalil et al. (2024) explore adaptive power allocation in spaceborne-assisted NOMA systems, proposing a transition model combining Gaussian Mixture Models and Log-Normal distributions to optimize power distribution between spaceborne and terrestrial signals.
The challenges and opportunities presented by emerging 6G technologies are also addressed. Saleh et al. (2024) Saleh et al. (2024) investigate RIS-aided single-LEO localization, developing a signal model incorporating Doppler effects and a low-complexity estimator. Gao et al. (2024) Gao et al. (2024) propose a vision-aided RAT selection framework for ISAC systems, utilizing computer vision for user localization and activity recognition to optimize resource allocation. Yu et al. (2024) Yu et al. (2024) introduce BUPTCMCC-6G-DataAI+, a generative channel dataset for 6G AI air interface research, providing deterministic data for various 6G scenarios. Chen et al. (2024) Chen et al. (2024) analyze near-field communications for XL-MIMO from a beamspace perspective, utilizing fractional Fourier transform for antenna space mapping.
Further contributions explore novel applications and theoretical advancements, including works on biomedical signal processing, sustainable ICT, and advanced waveform design. These cover topics such as Grover adaptive search with spin variables Fujiwara & Ishikawa (2024), neighbor discovery and beam alignment techniques in mmWave-enabled UAV swarm networks Alam & Moh (2024), channel charting-based channel prediction for distributed massive MIMO Stephan et al. (2024), multi-block UAMP detection algorithm for AFDM under fractional delay-Doppler channels Xu et al. (2024), a DRL-based approach for communication-control codesign in large-scale WNCSs Pang et al. (2024), an RL-based energy-efficient transmission parameter selection method for distributed LoRa networks Airiyoshi et al. (2024), high-precision decoding of spinal motor neurons Grison et al. (2024), user behavior dynamics for environmentally sustainable ICT Roth & Sezgin (2024), and DRIP, a family of space-time ISAC waveforms with controlled PAPR Wang et al. (2024). Several other papers contribute to diverse areas such as fiber-longitudinal optical power monitoring, spectrum sharing in vehicular networks, line spectral analysis, active transmissive RIS design, magnetic distortion resistant orientation estimation, dual-sniffer passive localization, constrained time series generation, physiological signal classification using speech foundation models, dynamic backoff optimization for MIMO-enabled grant-free NOMA, AI-aided Kalman filters, enhanced IoT communication via smart antennas, emotion-augmented audio captioning, signal source evaluation for brain-computer interfaces, and CSI-based self-quarantine monitoring. These works collectively demonstrate a significant push towards intelligent and efficient signal processing and communication systems.
Paper Highlights
AI-Aided Kalman Filters
AI-Aided Kalman Filters by Nir Shlezinger, Guy Revach, Anubhab Ghosh, Saikat Chatterjee, Shuo Tang, Tales Imbiriba, Jindrich Dunik, Ondrej Straka, Pau Closas, Yonina C. Eldar https://arxiv.org/abs/2410.12289
Caption: This figure illustrates three filtering paradigms: a) Model-Based Filter using system identification to derive a state-space model, b) AI Filter relying solely on data-driven weights within a neural network, and c) AI-Augmented Filter combining data, neural networks, and approximate state-space models in both task-oriented (DNN within the filter) and SS-oriented (DNN for system identification) approaches. These approaches represent the evolution of state estimation techniques, from traditional model-based methods to data-driven AI filters and finally to hybrid approaches leveraging the strengths of both.
Kalman filters (KFs) have long been essential tools for state estimation, particularly in linear Gaussian systems. They provide optimal performance and interpretable results with reliable uncertainty measures. However, their dependence on accurate state-space (SS) models limits their effectiveness in real-world scenarios where system dynamics are often complex, non-linear, and only partially understood. The emergence of AI, and specifically deep learning (DL), offers a promising avenue to overcome these limitations. AI-aided KFs combine the strengths of model-based KFs with the adaptability and data-driven capabilities of deep neural networks (DNNs).
This paper categorizes AI-aided KF design approaches into two primary families: task-oriented and SS model-oriented. Task-oriented methods transform Kalman-type state estimation into a trainable machine learning architecture by incorporating DNNs. This might involve using a DNN to pre-process observations, mapping them into a latent space suitable for KF processing (z<sub>t</sub> = F<sub>θ</sub>(y<sub>t</sub>) ≈ Hx<sub>t</sub> + w<sub>t</sub>), or learning correction terms for internal KF calculations. Alternatively, integrated DNN architectures can replace specific KF components, such as the Kalman gain (KG) with learned modules (x<sub>t</sub> = x̂<sub>t|t-1</sub> + K<sub>t</sub>(θ)(y<sub>t</sub> - ŷ<sub>t|t-1</sub>)), effectively turning the entire filter into a trainable ML model.
SS model-oriented approaches, on the other hand, concentrate on using DNNs to learn, refine, or augment the underlying statistical model. This includes data-driven models (DDMs) learned entirely from data, physics-informed neural networks (PINNs) that incorporate physical constraints, and augmented physics-based models (APBMs) that combine physics-based models with data-driven components (x<sub>t+1</sub> = g(f<sub>t</sub>(x<sub>t</sub>), x<sub>t</sub>, d<sub>t</sub>; θ<sub>APBM</sub>) + v<sub>t</sub><sup>PBM</sup>). These approaches offer varying degrees of model interpretability and adaptability.
The paper compares these different design approaches, highlighting the trade-offs. While end-to-end DNNs offer flexibility, they lack interpretability and adaptability. AI-augmented KFs, especially those with learned KGs, provide a balance between performance and interpretability, effectively handling non-linearities and uncertainties. SS-oriented methods, particularly APBMs, retain physical meaning and adaptability while benefiting from data-driven refinement. The level of domain knowledge required also varies significantly across these approaches.
A quantitative study, focusing on tracking a chaotic Lorenz attractor with sampling mismatch, demonstrates the effectiveness of AI-aided KFs. Model-based filters struggle with this mismatch, while AI-aided methods, including KalmanNet, DANSE, and APBM, outperform them significantly. The paper concludes by outlining future research directions, including handling time-varying SS models, incorporating non-Markovian dynamics, addressing non-Gaussian noises, developing distributed AI-aided KFs, and ensuring robust training against outliers.
Intramuscular High-Density Micro-Electrode Arrays Enable High-Precision Decoding and Mapping of Spinal Motor Neurons to Reveal Hand Control
Intramuscular High-Density Micro-Electrode Arrays Enable High-Precision Decoding and Mapping of Spinal Motor Neurons to Reveal Hand Control by Agnese Grison, Jaime Ibanez Pereda, Silvia Muceli, Aritra Kundu, Farah Baracat, Giacomo Indiveri, Elisa Donati, Dario Farina https://arxiv.org/abs/2410.11016
Caption: Figure a displays the normalized confusion matrices for two participants (S1 and S2) using root mean square (r.m.s.) features from high-density intramuscular electromyography (HD-iEMG) for gesture classification, achieving accuracies ranging from 71% to 97%. Figure b shows the improved classification results using motor unit discharge timings decoded from HD-iEMG, achieving near-perfect accuracy (>99%) for both participants across a range of gesture classes relevant for prosthetic control. This demonstrates the superior performance of the HD-iEMG-based decoding approach for precise control in neural interfaces.
This research introduces a groundbreaking neural decoding system using high-density intramuscular electromyography (HD-iEMG) to achieve unprecedented precision in decoding and mapping spinal motor neuron activity related to hand control. Researchers implanted three micro-electrode arrays (40 electrodes each) into the forearm muscles of two healthy participants. This is the largest reported number of implanted electrodes in forearm muscles, allowing for extensive sampling of muscle activity. The system combines this extensive sampling with advanced techniques for neural decomposition, analysis, and classification, enabling accurate detection and interpretation of spinal motor neuron spiking activity.
The study revealed that multi-digit tasks elicit unique motor neuron recruitment patterns, distinct from simple combinations of single-digit patterns. This task-specific recruitment was quantified, showing that the majority of active motor units were unique to the specific multi-digit task. Motor units from single-digit tasks contributed minimally (2.3±3.7%) to the motor units recruited in multi-digit tasks. This key observation led to the hypothesis that hand tasks could be classified with high precision based on decoded neural activity. The researchers tested this hypothesis by classifying various gestures, including individual finger movements and two-finger combinations.
The results were remarkable. Using motor unit discharge timings decoded from HD-iEMG, the system achieved perfect classification accuracy (100%) for 12 targeted muscle classes in both participants. Even for a larger set of 16 classes, the system maintained high accuracy (>96%). These results significantly outperformed classification based on global features from HD-iEMG and HD-sEMG, which achieved accuracies ranging from 71% to 97%. This superior performance highlights the potential of this approach for precise and robust control in neural interfaces, particularly for prosthetic control.
While the study was conducted offline and with a small sample size, the near-perfect classification accuracy achieved using decoded motor unit activity paves the way for future research on real-time implementation and validation in larger populations. Addressing challenges like the computational intensity of real-time HD-iEMG decomposition, potentially through integration with neuromorphic hardware, will be crucial for translating this promising technology into practical applications.
GBSense: A GHz-Bandwidth Compressed Spectrum Sensing System
GBSense: A GHz-Bandwidth Compressed Spectrum Sensing System by Zihang Song, Yue Gao https://arxiv.org/abs/2410.11495
Caption: This diagram illustrates the time-interleaved analog-to-digital conversion (TI-ADC) at the heart of GBSense. It shows how the input signal x(t) is sampled at periodically non-uniform intervals ( t = nT + c<sub>p</sub>(T/L) ) using P parallel ADCs, each with a unique clock offset c<sub>p</sub>. These samples (Z<sub>cp</sub>) are then aligned for spectrum reconstruction.
GBSense is a novel compressed spectrum sensing (CSS) system designed for high-bandwidth signals. Traditional spectrum sensing, reliant on Nyquist sampling, requires high sampling rates for wideband signals, leading to high power consumption and expensive hardware. GBSense addresses this challenge by using sub-Nyquist sampling, exploiting signal sparsity in the frequency domain to reduce the required sampling rate without compromising reconstruction accuracy. Instead of complex analog delay circuits commonly used in multicoset sampling (MCS), GBSense utilizes time-interleaved analog-to-digital conversion (TI-ADC). This allows for precise, real-time adjustable delay control by simply modifying the sampling clocks, avoiding complex and often inflexible analog components.
The key to GBSense lies in its efficient implementation of periodic nonuniform sampling (PNS). By adjusting the P sampling clocks to unique offsets (c<sub>p</sub>T/L) relative to a common reference, the system achieves the desired non-uniform sampling pattern. The minimum achievable delay resolution is Δt<sub>min</sub> = 1/(2f<sub>VCO</sub>), where f<sub>VCO</sub> is the voltage-controlled oscillator frequency. The system uses the JESD204B high-speed serial interface for precise data realignment from the multiple ADCs at the receiver, overcoming challenges posed by non-aligned digitization instances and varying lane latencies. Data decimation is also employed to reduce the computational load for spectrum reconstruction without increasing the risk of aliasing.
Laboratory tests confirm GBSense's effectiveness. Using a 2 GHz radio frequency bandwidth signal, the system achieved accurate spectrum reconstruction with an average sampling rate of just 400 MHz. The results demonstrated 100% accurate reconstruction for spectrum occupancy below 100 MHz and over 80% accuracy for occupancy up to 200 MHz. An integrated CSS system built around GBSense and a low-power Raspberry Pi processor exhibited a low processing latency of approximately 30 ms per frame, showcasing its real-time performance capabilities. This makes GBSense a promising solution for dynamic spectrum access in next-generation wireless communication systems.
Conclusion
This newsletter showcases a convergence of AI and advanced signal processing techniques driving innovation across various communication domains. From enhancing the robustness of AMC through XAI to revolutionizing spectrum sensing with compressed sampling techniques like GBSense, these advancements are pushing the boundaries of efficiency and performance. The highlighted papers exemplify this trend, with AI-aided Kalman filters demonstrating improved state estimation in complex dynamic systems, high-density intramuscular electrodes enabling precise decoding of motor neuron activity for advanced neural interfaces, and GBSense offering a practical solution for efficient wideband spectrum sensing. These breakthroughs collectively pave the way for more intelligent, robust, and efficient communication and signal processing systems in the future.