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

Hi Elman,

This newsletter covers recent breakthroughs in signal processing and machine learning across diverse applications, from communication systems optimization to innovative sensing technologies and healthcare. We'll explore key themes and highlight some of the most impactful papers pushing the boundaries of these fields.

General Overview

This collection of papers explores advancements in signal processing and machine learning for diverse applications, ranging from communication systems and network optimization to healthcare and sensing technologies. Several papers focus on improving the efficiency and robustness of wireless communication systems. Sukhsagar et al. (2024) Sukhsagar et al. (2024) present a closed-form solution for symbol error probability in hexagonal QAM, promising higher data rates and energy efficiency for future communication systems. Focusing on integrated imaging and communication (IIAC), Tornielli Bellini et al. (2024) Tornielli Bellini et al. (2024) propose a multi-view approach for NLOS exploration, achieving high-resolution imaging through far-field acquisitions. Kong et al. (2024) Kong et al. (2024) address propagation distance estimation in radio over fiber systems with cascaded structures, considering non-linear distortions and demonstrating accurate estimation despite these challenges. Díaz-Ruiz et al. (2024) Díaz-Ruiz et al. (2024) analyze precoding matrix techniques for MIMO optimization in 5G, comparing Type I and II codebooks to enhance network capacity and reduce latency. Finally, Mukhopadhyay et al. (2024) Mukhopadhyay et al. (2024) investigate optimal ASK modulation for RIS-assisted noncoherent communication, demonstrating superior error performance with their proposed schemes.

Another prominent theme is the application of machine learning, particularly graph neural networks (GNNs), for network optimization and resource allocation. Zhang et al. (2024) Zhang et al. (2024) employ a model-based GNN for energy-efficient beamforming in ultra-dense wireless networks, demonstrating millisecond-level response and adaptability to varying channel conditions. Similarly, Wang et al. (2024) Wang et al. (2024) utilize GNNs to optimize placement and transmission design for UAV communications, showcasing scalability and effectiveness in maximizing energy efficiency. Orimogunje et al. (2024) Orimogunje et al. (2024) develop an autonomous self-trained channel state prediction method for mmWave vehicular communications, leveraging RNNs and C-V2X messages for proactive beam switching. Li et al. (2024) Li et al. (2024) contribute to the understanding of in-vehicle wireless networks by characterizing channels across multiple frequency bands, providing valuable insights for future in-vehicle communication systems.

Several contributions focus on novel sensing applications and data-driven approaches. Tatli et al. (2024) Tatli et al. (2024) explore prediabetes detection using wearable sensors, combining CGM and smartwatch data with machine learning for early identification. Tashakori et al. (2024) Tashakori et al. (2024) introduce machine learning-powered smart textile gloves for capturing complex hand movements, achieving high accuracy in tracking hand and finger articulations. Diallo and Garbinato (2024) Diallo & Garbinato (2024) propose a decentralized collaborative inertial tracking algorithm for indoor environments, leveraging data exchange between mobile devices for improved location estimation. Fan and Kopsaftopoulos (2024) Fan & Kopsaftopoulos (2024) present a data-driven framework for guided wave-based structural health monitoring, addressing the challenges of varying environmental conditions. Yazdkhasti et al. (2024) Yazdkhasti et al. (2024) develop a novel ultrasonic device for monitoring implant condition, offering a potential low-cost solution for early detection of implant loosening.

Further contributions explore innovative approaches in distributed learning, semantic communication, and signal processing. Pignata et al. (2024) Pignata et al. (2024) introduce lightweight diffusion models for resource-constrained semantic communication, enabling efficient image regeneration from compressed semantic information. Al-Zawqari et al. (2024) Al-Zawqari et al. (2024) propose Uniform Cross-Entropy optimization for automating the design of multi-band microstrip antennas, demonstrating faster convergence compared to traditional methods. Liu et al. (2024) Liu et al. (2024) present a TRIS-assisted cognitive wireless human activity recognition system, enhancing signal clarity and recognition accuracy in through-the-wall scenarios. Ito et al. (2024) Ito et al. (2024) develop a polynomial-time algorithm for multivariable quantum signal processing, providing a constructive method for implementing M-QSP. Finally, Anavangot (2024) Anavangot (2024) proposes algorithms for overpredictive signal analytics in federated learning, addressing communication constraints and privacy concerns in distributed signal processing.

Paper Highlights

Prediabetes Detection Using Wearable Sensors

Prediabetes detection in unconstrained conditions using wearable sensors by Dimitra Tatli, Vasileios Papapanagiotou, Aris Liakos, Apostolos Tsapas, Anastasios Delopoulos https://arxiv.org/abs/2410.02692

Caption: This ROC curve compares the performance of various prediabetes detection methods. SVM models trained on wearable sensor data (Θᵥᵥ, Θ_h, and Θ_h U Θ_c) are compared against a lab test-based SVM (Θ_l) and common medical practice using two blood tests (HbA1c + FPG and FPG + OGTT). The combined wearable feature set (Θ_h U Θ_c) demonstrates promising sensitivity for prediabetes screening.

Prediabetes, a precursor to type 2 diabetes, often goes undetected. This research presents a novel, non-invasive method for early prediabetes detection using wearable technology—specifically, continuous glucose monitors (CGMs) and smartwatches. This approach addresses limitations of traditional diagnostic methods, which require blood draws and fasting. The study involved 22 participants wearing both devices for 14 days. Data analysis focused on two distinct feature sets. The first, Θ_h, is derived from a dynamic model of glucose homeostasis, incorporating glucose levels, accelerometer data (for physical activity estimation as A₄ ·R·m(t)), and individual biometrics. The model is represented by the equations:

ė(t) = −A₃ – u(t) · (e(t) + ē) + F(t) – A₄ ·R·m(t)

u(t) = A₁e(t) + A₂∫^t_-∞e(-λ)(τ)dτ

The second feature set, Θ_c, mimics standard blood tests by extracting parameters from the CGM glucose curve, including mean glucose (μ_G), standard deviation of glucose (σ_G), an approximation of fasting plasma glucose (μ_FG), and a normalized postprandial glucose gradient (μ_NPGG) accounting for physical activity. Support vector machines (SVMs) were trained on various feature combinations (Θ_VV - original Van Veen features, Θ_h, and Θ_h U Θ_c) and compared against lab blood test results (Θ_l) and standard medical practice using two blood tests (HbA1c + FPG or FPG + OGTT). The combined wearable feature set (Θ_h U Θ_c) achieved high accuracy (0.86) and sensitivity (0.92), suggesting its potential as a screening tool. While lab tests (Θ_l) achieved higher overall accuracy (0.9), the wearable approach offers a promising non-invasive alternative. This study highlights the potential of wearable sensors for prediabetes detection in real-world settings, demonstrating the value of combining physiological modeling with data-driven approaches. Future work should focus on larger, more diverse populations and further integration of CGM and smartwatch functionalities.

Context-Aware Predictive Coding for WiFi Sensing

Context-Aware Predictive Coding: A Representation Learning Framework for WiFi Sensing by B. Barahimi, H. Tabassum, M. Omer, O. Waqar https://arxiv.org/abs/2410.01825

Caption: The CAPC framework processes unlabelled CSI windows through a twin-branch architecture, each using a base encoder (E) and autoregressive model (G) to generate latent and context representations. These representations are then used to calculate the CPC loss (LCPC) for temporal consistency and the Barlow Twins loss (LBT) for robustness against augmentations, ultimately learning robust representations for downstream tasks like HAR.

WiFi sensing, utilizing wireless signals for applications like human activity recognition (HAR), faces challenges due to the limited availability of labelled data and the need for models to generalize across diverse environments. This paper introduces Context-Aware Predictive Coding (CAPC), a self-supervised learning (SSL) framework designed to overcome these challenges. CAPC uniquely blends Contrastive Predictive Coding (CPC) and Barlow Twins, two powerful SSL methods. This hybrid approach learns robust data representations from unlabelled Channel State Information (CSI) data, emphasizing both temporal and contextual consistency. This is crucial for capturing the dynamic nature of CSI, especially in tasks like HAR, and ensures robustness against data distortions.

CAPC features a twin-branch architecture where each branch processes multiple augmented views of CSI samples, including a novel "Dual View" augmentation. This augmentation leverages both uplink and downlink CSI to isolate free space propagation effects and minimize the impact of electronic distortions in transceivers, enhancing generalization. The core of CAPC is a hybrid contrastive loss function combining the CPC loss (LCPC) and the Barlow Twins loss (LBT):

L = LBT + β(LĈPC + LÊPC)

where β balances the two components. The CPC loss predicts future latent representations from context embeddings, capturing temporal dynamics. The LBT loss ensures consistency between the context embeddings of the two branches, promoting robustness against augmentations and reducing redundancy.

Evaluations on SignFi and UT HAR datasets demonstrate CAPC's superior performance. In low-labelled data scenarios, it significantly outperforms supervised learning and other SSL baselines. Transfer learning studies on the UT HAR dataset highlight CAPC's exceptional cross-domain adaptability. The paper also analyzes various data augmentations for time-series and WiFi sensing, demonstrating the effectiveness of "Dual View." A sensitivity analysis shows CAPC's robustness across different hyperparameter settings. These findings position CAPC as a significant advancement in SSL for WiFi sensing, offering a promising solution for real-world applications where labelled data is scarce and environmental variations are common.

Lightweight Diffusion Models for Semantic Communication

Lightweight Diffusion Models for Resource-Constrained Semantic Communication by Giovanni Pignata, Eleonora Grassucci, Giordano Cicchetti, Danilo Comminiello https://arxiv.org/abs/2410.02491

Caption: This diagram illustrates the Q-GESCO framework for quantized generative semantic communication. It depicts the four key steps: noise-aware training of the semantic diffusion model, quantization of the model weights, noise and timestep-aware calibration of the quantized model, and finally, the inference process for semantic image transmission and reconstruction. The framework significantly reduces the model's computational and memory footprint while maintaining high-quality image generation, making it suitable for resource-constrained devices.

Generative semantic communication (GSC) offers a revolutionary approach to communication, but the computational demands of generative models, particularly diffusion models, hinder their deployment on resource-constrained devices. This paper introduces Q-GESCO (Quantized GEnerative Semantic COmmunication), a framework designed to bring the power of GSC to such devices. Q-GESCO utilizes a quantized semantic diffusion model at the receiver, which regenerates content from received semantic maps. The key innovation is the application of post-training quantization (PTQ), compressing the denoising neural network within the diffusion model by reducing parameter precision. This significantly reduces memory and computational load without compromising generative capabilities.

Q-GESCO employs noise-aware training, exposing the model to noisy semantic maps during training to simulate real-world channel conditions. The PTQ process utilizes a carefully designed calibration dataset, generated using the full-precision model and incorporating both timestep variations and simulated channel noise. This ensures robustness to varying activation distributions and accounts for the model's noise resilience. A split quantization technique targeting weights before concatenation operations further mitigates quantization error accumulation.

Experimental results on the Cityscapes dataset demonstrate Q-GESCO's effectiveness. The quantized model achieves a 75% reduction in memory and a 79% reduction in FLOPS compared to its full-precision counterpart, while maintaining comparable or even superior image quality (measured by LPIPS and FID), even under noisy channel conditions. Ablation studies validate the effectiveness of the noise-aware calibration process. Q-GESCO's lightweight design and robust performance make it a promising solution for bringing GSC to a wider range of applications and devices.

Conclusion

This newsletter highlighted key advancements in signal processing and machine learning. A recurring theme is the focus on efficiency and robustness, as seen in the development of lightweight diffusion models for semantic communication (Q-GESCO) and the innovative context-aware predictive coding (CAPC) for WiFi sensing. These approaches address the practical challenges of deploying complex models in resource-constrained environments. Furthermore, the application of these techniques extends beyond communication systems, as exemplified by the promising results in prediabetes detection using wearable sensors. These innovations collectively demonstrate the potential of data-driven approaches to transform various domains, from healthcare to wireless communication, by enabling more efficient, robust, and intelligent systems.