Subject: Cutting-Edge Advancements in Signal Processing, Communications, and Machine Learning
Hi Elman,
This newsletter covers recent preprints exploring advancements in signal processing, communication systems, and machine learning, with applications in radar, wireless networks, and robotics.
This collection of preprints showcases exciting progress in signal processing, communication systems, and machine learning, particularly in radar, wireless networks, and robotics. Many papers concentrate on enhancing existing systems. For example, Kitchen et al. (2025) Kitchen et al. (2025) propose a distributed algorithm for joint localization and tracking using MIMO radars. Their mean field approach achieves impressive accuracy even at low SNRs. Similarly, Masiero et al. (2025) Masiero et al. (2025) introduce a computationally efficient sigma point-based algorithm for MP-SLAM in MIMO systems, offering a compelling alternative to particle-based methods. In the realm of communication networks, Martins de Jesus et al. (2025) Martins de Jesus et al. (2025) analyze age-of-information (AoI)-aware multiple access mechanisms in multi-relay networks, presenting novel algorithms that minimize signaling overhead while approaching theoretical AoI lower bounds. Additionally, Tapie and del Hougne (2025) Tapie & del Hougne (2025) present a scalable solution for multiport antenna array characterization using a PCB-realized tunable load network, effectively extending VNA capabilities.
Machine learning also takes center stage. Dufrène et al. (2025) Dufrène et al. (2025) explore learning linear block codes optimized for belief propagation decoders, introducing novel optimizers for discrete-valued weights and a comprehensive performance assessment framework. Huang et al. (2025) Huang et al. (2025) leverage MoE-based large language models for zero-shot multi-task semantic communication, demonstrating improved generalization and accuracy on unseen tasks.
Specific challenges in wireless communication are also addressed. Several works investigate physical layer security enhancements, novel metasurface applications, and efficient resource allocation in various network configurations. These include studies on cognitive radio-enabled NTNs, multiplexed backscatter communication, and dense LEO network coexistence. Furthermore, advanced signal processing techniques are explored for radar applications, including direct multiobject tracking, non-cooperative radar signal parsing, and chirp parameter retrieval. Finally, innovative waveform and filter design strategies for ISAC and RIS-aided communications are presented.
Beyond wireless communication, the preprints delve into diverse applications. Lenz et al. (2025) Lenz et al. (2025) connect speech acoustics with radar-measured vibrations, opening new possibilities for radar-based speech processing. Other works explore distributed reinforcement learning, deaf-mute assistive technology, and the secrecy performance of α-F channels. Finally, several papers focus on robotics, including decoding monkey neural data into robotic arm movement and developing inductive position sensors.
Communication-Efficient Distributed On-Device LLM Inference Over Wireless Networks by Kai Zhang, Hengtao He, Shenghui Song, Jun Zhang, Khaled B. Letaief https://arxiv.org/abs/2503.14882
This paper proposes a groundbreaking framework for deploying large language models (LLMs) on resource-constrained edge devices. The sheer size and computational demands of LLMs typically necessitate cloud-based inference, introducing latency and privacy concerns. This work tackles these challenges by employing tensor parallelism and over-the-air computation (AirComp). Tensor parallelism distributes the LLM's weight matrices across multiple devices for parallel processing. However, this introduces frequent all-reduce operations, which can create a communication bottleneck. AirComp cleverly addresses this by exploiting the superposition property of wireless channels, enabling simultaneous transmission and aggregation of data, thus significantly reducing communication overhead.
A key contribution is the formulation of a joint model assignment and transceiver optimization problem to minimize the average transmission mean squared error (MSE): MSE(â,z) = E[||ẑ - z||²], where ẑ is the received signal after Aircomp and z is the desired target summation. This problem is inherently complex, being a mixed-timescale stochastic non-convex optimization problem. The authors propose an efficient two-stage algorithm: the first stage handles long-term model assignment based on statistical channel information using semidefinite relaxation (SDR) and stochastic successive convex approximation (SCA); the second stage optimizes short-term transceiver variables based on instantaneous channel state information. The algorithm is theoretically sound, with proven almost sure convergence to a stationary point. Furthermore, the framework extends to multi-antenna edge devices, leveraging spatial multiplexing for enhanced communication efficiency.
Simulations using state-of-the-art LLMs (LLaMA2 and LLaMA3) and a real-world text dataset (WikiText-2) validate the effectiveness of this approach. The results demonstrate remarkable performance gains compared to benchmark schemes, including digital and uncoded FDMA all-reduce methods. The AirComp-based approach achieves up to 5x inference speed acceleration and improves inference accuracy (measured by perplexity). Importantly, the framework maintains low and stable perplexity even with increasing numbers of devices, showcasing its scalability. Comparisons with centralized inference further highlight the latency advantages, especially for large-scale models. This research opens exciting possibilities for deploying powerful LLMs in resource-constrained environments.
MetaFAP: Meta-Learning for Frequency Agnostic Prediction of Metasurface Properties by Rafid Umayer Murshed, Md Shoaib Akhter Rafi, Sakib Reza, Mohammad Saquib, Ifana Mahbub https://arxiv.org/abs/2503.14866
Reconfigurable intelligent surfaces (RIS) are transforming wireless communications by enabling dynamic control over electromagnetic waves. However, the frequency-dependent nature of meta-atom resonances presents a major challenge, particularly for broadband and multi-band operation in dynamic spectrum access and carrier aggregation scenarios. Traditional full-wave EM simulations are computationally intensive, and standard deep learning models often struggle with overfitting and poor generalization across frequency bands. This paper introduces MetaFAP, a novel meta-learning framework for predicting metasurface properties, addressing these limitations.
MetaFAP utilizes model-agnostic meta-learning (MAML) to predict properties such as reflectance, transmittance, and absorbance. By training on diverse frequency tasks, MetaFAP learns generalizable patterns and adapts swiftly to new spectral conditions with minimal data. The MAML framework employs a two-loop learning process: an inner loop for task-specific adaptation (θ ← θ - α∇<sub>θ</sub>L<sub>T<sub>i</sub></sub>(θ)) and an outer loop for meta-objective optimization (min<sub>θ</sub> Σ<sup>K</sup><sub>i=1</sub> L<sub>T<sub>i</sub></sub>(θ - α∇<sub>θ</sub>L<sub>T<sub>i</sub></sub>(θ))). This allows the model to learn an initialization that facilitates rapid adaptation to new frequency bands.
Experimental results demonstrate MetaFAP’s superiority over traditional machine learning and deep learning methods, achieving an order-of-magnitude reduction in MSE and MAE (MSE as low as 0.0079) and maintaining high Pearson correlations (around 80%). Critically, MetaFAP achieves inference in under 0.15 ms, significantly faster than traditional simulations, which can take minutes per unit cell and scale poorly with array size. An ablation study further validates the importance of both frequency-dependent and independent features in the model architecture. MetaFAP's speed and accuracy pave the way for practical, scalable, and adaptive metasurface implementations in dynamic wireless environments, particularly for applications like adaptive beamforming and real-time reconfiguration.
Robust Transmission of Punctured Text with Large Language Model-based Recovery by Sojeong Park, Hyeonho Noh, Hyun Jong Yang https://arxiv.org/abs/2503.14831
Semantic communication aims to improve efficiency by transmitting only task-relevant features. However, existing methods often suffer from data dependency, performing poorly on data outside their training set. This paper introduces a novel, data-independent text transmission model that leverages the power of LLMs for robust recovery. Instead of transmitting semantic features, the model selectively transmits characters and reconstructs the full text at the receiver using a pre-trained LLM. This approach bypasses the need for data-specific training, enhancing robustness and generalizability.
Central to this model is the Important Character Extractor (ICE). ICE strategically selects which characters to transmit to maximize the LLM's reconstruction accuracy. This selection process aims to minimize the number of possible words (K) the LLM might consider when filling in missing characters. The filter selection within ICE maximizes the inner product with the importance score vector, s: f = arg max_{f∈{f1, f2,..., fм}} f^Ts, where f is the selected filter and M is the number of filters. This targeted selection significantly improves the LLM's recovery performance compared to random character omission.
Evaluations on text reconstruction and Q&A tasks using the Europarl and SQuAD 1.1 datasets demonstrate the model’s effectiveness. It outperforms traditional bit-based communication, particularly in low SNR conditions, and surpasses existing semantic communication models like DeepSC, demonstrating robustness across datasets and tasks. For instance, at 3dB SNR, the proposed model performs better on the Q&A task than DeepSC while maintaining comparable sentence similarity. This data-independent approach, coupled with ICE's intelligent character selection, leverages the LLM's knowledge and context understanding for effective text reconstruction, enhancing robustness and transmission efficiency.
This newsletter highlights a convergence of advancements across signal processing, communications, and machine learning. The trend towards leveraging the power of LLMs, as seen in the distributed inference and robust text transmission papers, is particularly noteworthy. Furthermore, the development of innovative techniques like AirComp and MetaFAP demonstrates a focus on addressing practical challenges in deploying complex systems in resource-constrained environments. These advancements collectively promise significant improvements in efficiency, robustness, and scalability across various applications, from wireless networks and radar systems to robotics and human-computer interaction.