Subject: Cutting-Edge Research in Wireless Communications, Sensing, and Machine Learning

Hi Elman,

This newsletter covers recent preprints exploring signal processing, communication systems, and machine learning applications in wireless networks. We'll delve into novel approaches in channel estimation, beamforming, localization, security, and the exciting potential of Reconfigurable Intelligent Surfaces (RIS) and Terahertz (THz) communication.

General Overview

This collection of preprints explores various aspects of signal processing, communication systems, and machine learning applications in wireless networks, including novel approaches to channel estimation, beamforming, localization, and security. Several papers focus on the challenges and opportunities presented by emerging technologies like Reconfigurable Intelligent Surfaces (RIS) and Terahertz (THz) communication. For instance, Inwood et al. (2024) (Inwood et al., 2024) investigate phase selection for multi-frequency multi-user RIS systems, proposing iterative methods that outperform existing approaches in specific scenarios, particularly those with strong Line-of-Sight channels and clustered users. Meanwhile, Xu et al. (2024) (Xu et al., 2024) delve into the design and application of Non-Reciprocal RIS (NR-RIS), demonstrating their potential for non-reciprocal beamsteering and channel reciprocity attacks. In the realm of THz communication, Bhattacharya & Gupta (2024) (Bhattacharya & Gupta, 2024) introduce a deep learning-based approach for THz channel estimation and beamforming prediction using sub-6GHz channel information, achieving near-optimal spectral efficiency. Zhao et al. (2024) (Zhao et al., 2024) further address the issue of hardware imperfections in THz hybrid beamforming, proposing a two-stage DNN-based compensation algorithm that significantly mitigates performance degradation.

Another prominent theme is the application of advanced signal processing techniques for localization and sensing. Spanos et al. (2024a) (Spanos et al., 2024a) and Spanos et al. (2024b) (Spanos et al., 2024b) present a USRP-based testbed for 5G positioning, utilizing Angle of Arrival (AoA) estimation with uplink signals and ultra-wideband ranging. Their work addresses practical challenges such as phase misalignment and high sampling rates, demonstrating accurate pedestrian localization. Aghashahi et al. (2024) (Aghashahi et al., 2024) propose algorithms for single antenna tracking and localization of RIS-enabled vehicular users, achieving significant improvements in localization accuracy compared to traditional Time of Arrival (ToA) methods. Furthermore, Wu et al. (2024) (Wu et al., 2024) introduce an FFT-enhanced low-complexity super-resolution sensing algorithm for near-field source localization, combining coarse and fine granularity spectrum peak search for efficient and high-resolution target localization.

Several contributions explore the intersection of machine learning and signal processing for diverse applications. Boumeftah et al. (2024) (Boumeftah et al., 2024) propose a machine learning-based approach for detecting on-orbit jamming in GEO satellite links, employing random forest with Principal Component Analysis (PCA) and adaptive thresholding. Pandukabhaya et al. (2024) (Pandukabhaya et al., 2024) present a framework for optimizing psychomotor learning using wearable IMU sensors, linking motion trajectories to performance metrics for skill optimization. In the biomedical domain, Turpin et al. (2024) (Turpin et al., 2024) investigate the effects of direct electrical stimulation on the human brain by analyzing direct cortical responses and axono-cortical evoked potentials. Similarly, Lopez Alcaraz & Strodthoff (2024) (Lopez Alcaraz & Strodthoff, 2024) develop multimodal deep learning models to estimate and monitor laboratory values from ECG signals, demonstrating the potential for non-invasive health monitoring.

Beyond these core themes, the papers also address specific challenges in communication systems. Chang et al. (2024) (Chang et al., 2024) investigate end-to-end learning for MU-MIMO systems, jointly optimizing learned constellations and detectors to achieve near-ML performance. Qian & Zhao (2024) (Qian & Zhao, 2024) propose a data processing efficiency-aware algorithm for user association and resource allocation in blockchain-enabled metaverse over wireless communications. Shrestha et al. (2024) (Shrestha et al., 2024) tackle the problem of downlink MIMO channel estimation from limited feedback, establishing recoverability conditions and proposing an ADMM-based algorithm. Finally, Zhou et al. (2024) (Zhou et al., 2024) propose a general sensing-assisted channel estimation framework for distributed MIMO networks, considering both LoS and NLoS scenarios.

Paper Highlights

Goal-oriented Semantic Communications for Metaverse Construction via Generative AI and Optimal Transport

Goal-oriented Semantic Communications for Metaverse Construction via Generative AI and Optimal Transport by Zhe Wang, Nan Li, Yansha Deng, A. Hamid Aghvami https://arxiv.org/abs/2411.16187

Caption: This diagram illustrates the Goal-Oriented Semantic Communication (GSC) framework for efficient metaverse construction. It showcases the key components, including semantic information extraction, OT-enabled semantic denoising, and neural network models utilizing Stable Diffusion and NeRF for scenery reconstruction. The framework minimizes transmission latency and improves accuracy by transmitting only essential semantic information, as demonstrated by the processing pipeline from physical scenery capture to reconstructed metaverse scenery.

The metaverse presents exciting possibilities for immersive experiences, but its reliance on real-time updates and personalized content creates a massive demand for data transmission. Traditional bit-oriented communication networks struggle to handle this volume, hindering the interactivity and realism of metaverse applications. This paper proposes a goal-oriented semantic communication (GSC) framework that leverages generative AI and optimal transport to address this critical challenge. Rather than transmitting raw data, the GSC framework extracts and transmits only the essential semantic information needed to reconstruct the metaverse experience at the receiver, leading to significant bandwidth savings.

The proposed GSC framework consists of several key components working in concert. A semantic encoder, based on an hourglass network (HgNet), analyzes the metaverse scenery and extracts key points representing the positions and movements of objects. This extracted semantic information is then transmitted wirelessly to the receiver. At the receiver, a semantic decoder, powered by Stable Diffusion and Neural Radiance Fields (NeRF), reconstructs the metaverse scenery using the received semantic information and a shared knowledge base. To combat the inevitable noise introduced by wireless channels, an optimal transport (OT)-enabled semantic denoiser is employed. This denoiser refines the received semantic information by minimizing the difference between its distribution and the distribution of the transmitted information, ensuring a more accurate reconstruction of the metaverse scenery. The framework's effectiveness is demonstrated in a simulated industrial factory environment, featuring both moving objects and a stationary background.

The performance of the GSC framework is evaluated using metrics such as key point error (KPE), point-to-point (P2Point) error, and transmission latency. The core objective is to minimize the geometric disparity between the point cloud representation at the transmitter (Pt) and the receiver (Pr), expressed as min_Pr C(Pt, Pr), where C(·) denotes a modified chamfer distance measure. The results demonstrate the significant advantages of the GSC framework over conventional methods. Notably, the GSC framework with OT denoising achieves a 45.6% reduction in KPE, a 44.7% improvement in P2Point error, and a remarkable 92.6% reduction in transmission latency compared to the conventional approach. These improvements underscore the framework's ability to achieve efficient and reliable metaverse construction. By transmitting only the essential semantic information, the GSC framework significantly reduces the burden on wireless networks, paving the way for real-time updates and interactions within the metaverse. The OT-enabled denoiser further enhances robustness by mitigating the impact of channel noise. The proposed GSC framework represents a significant advancement toward realizing the full potential of the metaverse, offering a path to seamless and immersive real-time experiences.

ChatBCI: A P300 Speller BCI Leveraging Large Language Models for Improved Sentence Composition in Realistic Scenarios

ChatBCI: A P300 Speller BCI Leveraging Large Language Models for Improved Sentence Composition in Realistic Scenarios by Jiazhen Hong, Weinan Wang, Laleh Najafizadeh https://arxiv.org/abs/2411.15395

Caption: This diagram illustrates the architecture of ChatBCI, a novel P300 speller enhanced by a large language model (LLM). The system integrates a traditional P300 visual interface with LLM-generated word predictions, allowing users to select entire words or characters. This interaction loop between the Stimulation Computer, Recording Computer, and ChatGPT facilitates efficient and contextually relevant sentence composition.

P300 speller BCIs offer a vital communication pathway for individuals with motor disabilities, enabling them to type by selecting characters on a screen through brain activity detected via EEG. However, traditional P300 spellers are limited by their letter-by-letter input method, leading to user fatigue and slow communication speeds. ChatBCI introduces a groundbreaking approach by incorporating large language models (LLMs), specifically GPT-3.5, to significantly improve sentence composition. This innovative system features a redesigned GUI that displays both individual characters and LLM-suggested words, allowing users to select entire words or even predict subsequent words based on context, thus minimizing keystrokes and accelerating communication.

The core of ChatBCI lies in its seamless integration with GPT-3.5 through remote queries. A carefully designed prompt template enables ChatBCI to dynamically switch between word completion, suggesting possible endings for partially typed words, and word prediction, suggesting the next word in a sentence. As the user selects suggested words, the GUI updates, and a new query is sent to GPT-3.5 to refresh the word suggestions, maintaining contextual relevance throughout the sentence composition process. This continuous feedback loop ensures efficient and context-aware predictive typing. Stepwise Linear Discriminant Analysis (SWLDA) is employed for P300 classification, providing robust key selection from the GUI.

Seven subjects participated in online spelling tasks comparing ChatBCI with a traditional letter-by-letter speller. In a copy-spelling task, ChatBCI demonstrated remarkable improvements, reducing the time to complete the task by 62.14% and increasing the typing speed (measured by the lower bound information transfer rate, ITR*-1) by an impressive 198.96%. In a more realistic improvisation task, where subjects composed sentences starting with "H", ChatBCI achieved an average typing speed of 8.53 characters/min and an average keystroke savings of 80.68%. The keystroke savings deficit ratio (KS-DR), a novel metric introduced in this paper to quantify the gap between achieved and theoretical maximum keystroke savings (KS-WPmax, calculated as (Number of Characters in Sentence - Number of Words in Sentence) / (Number of Characters in Sentence) * 100), was less than 2.50% in the improvisation task, indicating that ChatBCI operates near its optimal predictive capacity. The information transfer rate (ITR*) was calculated using the formula: ITR* = (B / (T/60)) * α, where α = (# of characters in the sentence) / (# of selections).

The results of this study highlight the transformative potential of ChatBCI in improving real-time communication for individuals with disabilities. By harnessing the power of LLMs, ChatBCI offers a more efficient, user-friendly, and adaptable spelling experience. The introduction of keystroke analysis and the KS-DR metric provides valuable tools for evaluating and optimizing predictive typing systems. Future research directions include incorporating dynamic stopping techniques to further enhance ChatBCI's performance.

LightLLM: A Versatile Large Language Model for Predictive Light Sensing

LightLLM: A Versatile Large Language Model for Predictive Light Sensing by Jiawei Hu, Hong Jia, Mahbub Hassan, Lina Yao, Brano Kusy, Wen Hu https://arxiv.org/abs/2411.15211

Caption: The figure illustrates three LightLLM architecture variations: (a) LightLLM with a Task-Specific Encoder (TSE), Latent Fusion Layer (LFL), and Large Language Model (LLM); (b) Light_Transformer, replacing the LFL and LLM with a Transformer; and (c) Light_w/oLLM, utilizing only the TSE. These variations highlight the modularity and adaptability of the LightLLM framework for different predictive light sensing tasks.

Predictive light sensing (PLS) is becoming increasingly crucial in intelligent and sustainable systems, enabling applications such as indoor localization and solar energy forecasting. However, traditional deep learning models designed for PLS are often task-specific, require extensive training data, and struggle to generalize to new, unseen environments. This paper introduces LightLLM, a novel framework that leverages the power of pre-trained large language models (LLMs) to overcome these limitations. LightLLM incorporates a task-specific sensor data encoder, a contextual prompt, and a fusion layer to combine sensor data with environmental information into a unified representation. This combined representation is then processed by a frozen LLM, fine-tuned with Low-Rank Adaptation (LoRA) to minimize computational overhead.

LightLLM's architecture consists of several key components. Task-specific encoders, such as Graph Neural Networks (GNNs) for localization, Temporal Convolutional Networks (TCNs) for forecasting, and Convolutional Neural Networks (CNNs) for energy estimation, extract relevant features from various input data types. Task-specific knowledge prompts provide contextual information, like spatial relationships or environmental conditions, to guide the LLM's understanding of the task. A Latent Fusion Layer (LFL) integrates the encoded features and prompt embeddings using a multi-head attention mechanism, aligning the data with the LLM's latent space. Finally, LoRA enables efficient fine-tuning of the LLM for specific tasks without altering the pre-trained LLM's parameters, minimizing computational cost and retraining effort.

The authors evaluate LightLLM's performance on three real-world PLS tasks: light-based indoor localization, outdoor solar forecasting, and indoor solar estimation. For indoor localization, LightLLM demonstrates a remarkable 4.4x improvement in accuracy compared to the state-of-the-art Iris system when tested in unseen environments. In outdoor solar forecasting using the SKIPP'D dataset, LightLLM surpasses SkyGPT and TimeLLM, achieving a 33.7% improvement in Forecast Skill (FS) in seen environments and 31.4% in unseen environments, measured by Continuous Ranked Probability Score (CRPS). For indoor solar estimation, LightLLM significantly reduces the Mean Absolute Percentage Error (MAPE) to 26.33% in unseen environments, a substantial improvement over traditional machine learning models, which exhibited MAPE values exceeding 140%. A baseline for indoor solar estimation is calculated using the formula for the total current generated by a solar cell: Current = A ∫_λmin^λmax α(λ) ⋅ I(λ) ⋅ λ ⋅ dλ, where A is the surface area, α(λ) is the spectral absorption rate, I(λ) is the incident spectrum, and λ is the wavelength.

An ablation study further validates the importance of each component within the LightLLM framework. Directly prompting LLMs like GPT-4, even with chain-of-thought prompting, proved less effective than LightLLM's specialized architecture. Removing the knowledge graph, LFL, task-specific encoder, or LoRA resulted in performance degradation. The study also revealed that larger LLMs, such as LLaMA-7B, generally outperformed smaller models like GPT-2 within the LightLLM framework. LightLLM offers a promising approach for leveraging the power of LLMs in PLS. Its specialized architecture allows for efficient adaptation to various sensing tasks with minimal retraining, highlighting the potential of LLM-based frameworks to address the challenges of generalization and adaptability in real-world PLS applications. Future research could explore the use of even larger LLMs and further optimize the framework for specific sensor modalities and tasks.

Conclusion

This newsletter highlighted the growing trend of integrating data-driven approaches, especially deep learning, into various aspects of wireless communication and sensing. The impactful papers discussed showcased novel applications of LLMs, demonstrating their versatility in addressing complex challenges ranging from efficient metaverse construction and enhanced BCI communication to predictive light sensing. The GSC framework offers a paradigm shift in metaverse data transmission by prioritizing semantic information over raw data, leading to significant improvements in latency and accuracy. ChatBCI demonstrates the potential of LLMs to revolutionize assistive communication technologies, providing a more efficient and user-friendly experience for individuals with motor disabilities. LightLLM showcases the adaptability of LLMs to diverse sensing tasks, pushing the boundaries of performance in localization, forecasting, and energy estimation. These contributions underscore the transformative power of LLMs in shaping the future of wireless communication and sensing technologies.