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

Hi Elman,

This newsletter covers recent advancements in signal processing, communication systems, and machine learning, highlighting emerging technologies and impactful optimization techniques.

General Overview

This collection of papers explores advancements across various domains of signal processing, communication systems, and machine learning, with a particular emphasis on emerging technologies and optimization techniques. Several papers focus on improving the efficiency and accuracy of existing methods. Li and Principe (2025) introduce SPEED, a novel approach for kernel adaptive filtering that addresses scalability issues by providing an explicit, reduced-dimensionality Euclidean representation of the reproducing kernel Hilbert space (RKHS). Similarly, Inada et al. (2025) propose an efficient algorithm for full-waveform inversion with a total variation constraint, leveraging a primal-dual splitting method to avoid computationally expensive inner loops and approximations. For sparse channel estimation in massive MIMO systems, Arjas and Atzeni (2025) introduce enhanced sparse Bayesian learning (E-SBL and M-E-SBL) methods, demonstrating improved mean squared error performance compared to baseline SBL and variational message passing.

Another prominent theme is the integration of sensing and communication functionalities, particularly in the context of emerging wireless technologies. Bai et al. (2025) propose a multi-modal intelligent channel model for 6G multi-UAV-to-multi-vehicle communications, incorporating terrestrial and aerial traffic density parameters and leveraging LiDAR data for accurate channel characterization. Hussain et al. (2025) present a unified approach to near-field beam-training and sensing in ISAC systems, employing dual-purpose codebooks and a customized space-time adaptive processing technique. Ribeiro et al. (2025) explore mobility management in ISAC networks, proposing handover strategies for maintaining sensing continuity and mitigating interference. Lv et al. (2025) investigate multipath exploitation for target detection in OFDM-ISAC systems, developing a weighted GLRT detector and a joint optimization framework for subcarrier power allocation and detector weights.

The application of machine learning and deep learning techniques is also explored across several papers. Zhao et al. (2025) present a systematic review of machine learning methods for multimodal EEG data in clinical applications, highlighting the potential for improved diagnostic accuracy. Zhang and Liu (2025) introduce SDN-Net, a subject disentanglement neural network for reconstructing speech envelopes from EEG, addressing cross-subject variability. León-López et al. (2025) propose a Bayesian multifractal image segmentation method, demonstrating superior performance compared to traditional and deep learning-based approaches. For path loss prediction, Ethier et al. (2025) leverage machine learning with extended features derived from geographic information system data, improving prediction accuracy and model generalization. Kahya et al. (2025) develop FARE, a deep learning framework for radar-based face recognition and out-of-distribution detection.

Several papers address specific challenges in wireless communication systems. Alghananim et al. (2025) introduce the first calibration model for Bluetooth Angle of Arrival (AoA) positioning, enabling practical application of this technology for enhanced indoor localization. Wu et al. (2025) provide a comprehensive study on the deployment and coordination of intelligent reflecting surfaces (IRSs) for wireless networks, examining single- and multi-reflection architectures and discussing practical challenges. Yin et al. (2025) investigate low-complexity waveform, modulation, and coding designs for 3GPP Ambient IoT, focusing on backscatter communications. Sau and Ghosh (2025) propose a set cover-based RIS deployment strategy for mmWave D2D communication, demonstrating the benefits of double reflections. Zhang et al. (2025) propose a diffusion-enhanced decision Transformer framework for RIS-assisted systems, combining a diffusion model for CSI acquisition and a decision Transformer for phase-shift optimization.

Paper Highlights

Transforming Indoor Localization with Advanced Transformer Architecture

Transforming Indoor Localization: Advanced Transformer Architecture for NLOS Dominated Wireless Environments with Distributed Sensors by Saad Masrur, Jung-Fu (Thomas)Cheng, Atieh R. Khamesi, Ismail Guvenc https://arxiv.org/abs/2501.07774

Caption: The image illustrates the Sensor Snapshot Tokenization (SST) method, where each sensor's Power Delay Profile (PDP) over time is treated as an independent token. The red dashed lines delineate the time snapshots for each token, capturing the unique temporal characteristics of the signal received by each sensor. This tokenization approach preserves variable-specific representations, enabling the Transformer model to effectively capture multivariate correlations crucial for accurate indoor localization.

Indoor localization, especially in Non-Line-of-Sight (NLOS) environments, presents persistent challenges. While deep learning (DL) offers potential solutions, computational complexity often limits its applicability on resource-constrained devices. This paper introduces a novel Transformer architecture designed for high accuracy and computational efficiency in NLOS-dominated wireless environments, utilizing distributed sensors.

The key innovation is Sensor Snapshot Tokenization (SST). Instead of aggregating data, SST treats each sensor's Power Delay Profile (PDP) as an individual token. This preserves variable-specific information, allowing the Transformer's attention mechanism to capture crucial multivariate correlations for accurate localization. The model learns the mapping ŷ = f(P, Ψ; Θ), where ŷ represents the device's 2D coordinates, P is the sensor position vector, Ψ is the received PDP matrix, and Θ are the model parameters. Preprocessing involves power compression of the PDP, p[d] = S(2||{p[d]}||)^-1 p[d], addressing the wide dynamic range of received signal powers.

The architecture also features a Lightweight Swish-Gated Linear Unit-based Transformer (L-SwiGLU Transformer). This model utilizes a Swish-Gated Linear Unit (SwiGLU) based MLP, SwiGLU(Z^MHA) = (Swish(Z^MHA W₁) ⊙ (Z^MHA W₂))W₃, enhancing feature selection and noise reduction. Root Mean Square Normalization, RMSNorm(Z^(j-1)) = Z^(j-1) / RMS(Z^(j-1)), replaces Layer Normalization for improved stability. A Global Average Pooling (GAP) layer replaces the class token, reducing overhead while maintaining global representation. Positional embeddings are omitted due to the inherent positional information within the SST tokens.

Evaluations in a simulated 3GPP indoor factory scenario (InF-DH) highlight the approach's effectiveness. SST significantly outperforms conventional methods, achieving a 90th percentile positioning error (A90) of 0.388m with a large Vanilla Transformer. The L-SwiGLU ViT further reduces this to 0.355m (8.51% improvement). Impressively, the small L-SwiGLU model surpasses a 14.1 times larger model with conventional tokenization by 46.13%, showcasing its remarkable computational efficiency.

AI-RAN: Converging AI and RAN for Next-Generation Networks

AI-RAN: Transforming RAN with AI-driven Computing Infrastructure by Lopamudra Kundu, Xingqin Lin, Rajesh Gadiyar, Jean-Francois Lacasse, Shuvo Chowdhury https://arxiv.org/abs/2501.09007

The Radio Access Network (RAN) is evolving from a communication-centric model to a converged compute-communication platform. This paper introduces AI-RAN, integrating RAN and AI workloads on shared infrastructure for enhanced performance and resource utilization. This convergence allows dynamic resource allocation between RAN and AI, optimizing efficiency and adaptability.

AI-RAN manifests in three forms: AI-for-RAN, AI-on-RAN, and AI-and-RAN. AI-for-RAN uses AI to optimize RAN performance within the protocol stack (e.g., channel estimation, beamforming). AI-on-RAN leverages RAN infrastructure for AI applications, creating new revenue streams through AIaaS. AI-and-RAN focuses on efficient resource sharing between AI and RAN workloads, minimizing operational costs.

Key requirements for AI-RAN include accelerated computing (especially GPUs), a software-defined, cloud-native design for flexibility and scalability, joint orchestration of communication and computing resources, and native AI support, including network digital twins (NDTs).

The proposed AI-RAN architecture utilizes standard datacenter racks with AI-RAN servers (CPUs, GPUs, DPUs, SSDs) interconnected via ethernet. The software stack comprises a cloud OS, computing and networking platforms with APIs, and an end-to-end orchestrator managing both RAN and AI workloads.

A proof-of-concept using NVIDIA Grace-Hopper GH200 servers demonstrates concurrent RAN and AI processing. A 5G network interacting with an LLM-powered chatbot and digital human showcased simultaneous operation without performance degradation. Multi-Instance GPU (MIG) divided resources between 5G RAN (40%) and AI (60%), demonstrating efficient utilization and potential for expanded service offerings. The results showed a mean GPU utilization of 29.8% for AI, 11.3% for 5G RAN, and 41.1% combined, with maximums of 56.9%, 19.3%, and 76.0% respectively, highlighting AI-RAN's potential for maximizing resource utilization and cost reduction.

Bluetooth AoA Positioning: A Real-World Calibration Model for Enhanced Accuracy

The first calibration model for bluetooth angle of arrival: Enhancing positioning accuracy in indoor environments by Ma'mon Saeed Alghananim, Yuxiang Feng, Washington Yotto Ochieng https://arxiv.org/abs/2501.08805

Caption: This diagram illustrates the architecture of a novel Bluetooth Angle of Arrival (AoA) calibration model, a key component for achieving accurate indoor positioning. The model estimates anchor orientation parameters using a least squares approach, enabling the transformation of measured angles from the anchor's frame of reference to the user's coordinate system. This calibrated information is then fed into the AoA positioning algorithm, improving the accuracy of tag localization.

Indoor positioning is increasingly vital for IoT applications, and Bluetooth Angle of Arrival (AoA) offers a cost-effective path to sub-meter accuracy. However, real-world deployment has been limited by the lack of a method to calibrate AoA anchor orientation. This paper presents the first calibration model for Bluetooth AoA, bridging a crucial gap and enabling practical application. The model estimates the 3D rotation matrices between anchor and user coordinate systems, allowing for accurate angle transformation.

The calibration model leverages the AoA positioning model, treating anchor orientation parameters (rotation around x, y, and z axes – 𝜑, 𝜃, ψ) as unknowns. Using tags with known locations, unit vectors between tags and anchors are calculated in both coordinate systems. The calibration functional model is:

F(ψ, θ, φ) = Rz(ψ)Ry(θ)Rx(φ)u – v = 0

where Rz(ψ), Ry(θ), Rx(φ) are rotation matrices, u is the anchor's unit vector, and v is the user's unit vector. A nonlinear least squares approach iteratively refines the rotation angle estimates.

The paper also details a complete end-to-end AoA positioning architecture, including setup (anchor configuration, installation, calibration) and the positioning model. The calibrated anchor orientations and a least squares method are used to estimate tag positions based on measured angles.

Experiments in a 55m² classroom with Ublox XPLR-AOA-2 anchors and tags validated the model. Calibration yielded standard deviations of 0.94° to 2.49° for the rotation angles. Static positioning achieved horizontal errors of 0.33m to 1.25m, with better accuracy in areas with strong geometry and shorter anchor ranges. Kinematic tests showed average across-path horizontal and vertical errors of 1.06m and 1.18m respectively. Systematic error (bias) was identified as a significant error contributor, suggesting future research on bias mitigation.

Conclusion

This newsletter highlights a convergence of advancements across signal processing, communication, and machine learning. From optimized kernel adaptive filtering and full-waveform inversion to the innovative integration of AI and RAN infrastructure, these papers demonstrate a push towards greater efficiency and performance. The development of a Bluetooth AoA calibration model marks a significant step towards practical, high-accuracy indoor localization. The common thread is a focus on addressing real-world challenges, paving the way for more robust and impactful applications in various fields. The innovative use of Transformer architectures for indoor localization and the convergence of AI and RAN infrastructure represent significant steps towards more intelligent and efficient systems. These advancements promise to shape the future of wireless communication and signal processing.