Recently, there has been a surge in the development of advanced intelligent generative content (AIGC), especially large language models (LLMs). For many downstream tasks, it is necessary to fine-tune LLMs using private data. While federated learning offers a promising privacy-preserving solution to LLM fine-tuning, the substantial size of an LLM, combined with high computational and communication demands, makes it hard to apply to handle downstream tasks. More importantly, private edge servers often possess varying computing and network resources in real-world scenarios, introducing additional complexities to LLM fine-tuning. To tackle these problems, we design and implement an automated federated pipeline, named FedPipe, to fine-tune LLMs on heterogeneous edge servers with minimal training cost and no additional inference latency. FedPipe firstly identifies the weights to be fine-tuned based on their contributions to the LLM training. It then configures a low-rank adapter for each selected weight within the resource constraints of the edge server and aggregates these local adapters from all edge servers to fine-tune the whole LLM. Finally, it appropriately quantizes the parameters of LLM to reduce memory space according to the requirements of edge servers. Extensive experiments demonstrate that FedPipe expedites model training and achieves higher accuracy than the state-of-the-art benchmarks.

Low Earth Orbit satellite Internet has recently been deployed, providing worldwide service with non-terrestrial networks. With the large-scale deployment of both non-terrestrial and terrestrial networks, limited spectrum resources will not be allocated enough. Consequently , dynamic spectrum sharing is crucial for their coexistence in the same spectrum, where accurate spectrum sensing is essential. However, spectrum sensing in space is more challenging than in terrestrial networks due to variable channel conditions, making single-satellite sensing unstable. Therefore, we first attempt to design a collaborative sensing scheme utilizing diverse data from multiple satellites. However, it is non-trivial to achieve this collaboration due to heterogeneous channel quality, considerable raw sampling data, and packet loss. To address the above challenges, we first establish connections between the satellites by modeling their sensing data as a graph and devising a graph neural network-based algorithm to achieve effective spectrum sensing. Meanwhile, we establish a joint sub-Nyquist sampling and autoencoder data compression framework to reduce the amount of transmitted sensing data. Finally, we propose a contrastive learning-based mechanism compensates for missing packets. Extensive experiments demonstrate that our proposed strategy can achieve efficient spectrum sensing performance and outperform the conventional deep learning algorithm in spectrum sensing accuracy.

Due to sophisticated deployments of all kinds of wireless networks (e.g., 5G, Wi-Fi, Bluetooth, LEO satellite, etc.), multiband signals distribute in a large bandwidth (e.g., from 70 MHz to 8 GHz). Consequently, for network monitoring and spectrum sharing applications, a sniffer for extracting physical layer information, such as structure of packet, with low sampling rate (especially, sub-Nyquist sampling) can significantly improve their cost-and energy-efficiency. However, to achieve a multiband signals sniffer is really a challenge. To this end, we propose Sums, a system that can sniff and analyze multiband signals in a blind manner. Our Sums takes advantage of hardware and algorithm co-design, multi-coset sub-Nyquist sampling hardware, and a multi-task deep learning framework. The hardware component breaks the Nyquist rule to sample GHz bandwidth, but only pays for a 50 MSPS sampling rate. Our multi-task learning framework directly tackles the sampling data to perform spectrum sensing, physical layer protocol recognition, and demodulation for deep inspection from multiband signals. Extensive experiments demonstrate that Sums achieves higher accuracy than the state-of-theart baselines in spectrum sensing, modulation classification, and demodulation. As a result, our Sums can help researchers and end-users to diagnose or troubleshoot their problems of wireless infrastructures deployments in practice.

We present the implementation of four FPGA-accelerated convolutional neural network (CNN) models for onboard cloud detection in resource-constrained CubeSat missions, leveraging Xilinx's Vitis AI (VAI) framework and Deep Learning Processing Unit (DPU), a programmable engine with pre-implemented, parameterizable IP cores optimized for deep neural networks, on a Zynq UltraScale+ MPSoC. This study explores both pixel-wise (Pixel-Net and Patch-Net) and image-wise (U-Net and Scene-Net) models to benchmark trade-offs in accuracy, latency, and model complexity. Applying channel pruning, we achieved substantial reductions in model parameters (up to 98.6%) and floating-point operations (up to 90.7%) with minimal accuracy loss. Furthermore, the VAI tool was used to quantize the models to 8-bit precision, ensuring optimized hardware performance with negligible impact on accuracy. All models retained high accuracy post-FPGA integration, with a cumulative maximum accuracy drop of only 0.6% after quantization and pruning. The image-wise Scene-Net and U-Net models demonstrated strong real-time inference capabilities, achieving frame rates per second of 57.14 and 37.45, respectively, with power consumption of around 2.5 W, surpassing state-of-the-art onboard cloud detection solutions. Our approach underscores the potential of DPU-based hardware accelerators to expand the processing capabilities of small satellites, enabling efficient and flexible onboard CNN-based applications.

Given the wide adoption of multimodal sensors (e.g., camera, lidar, radar) by autonomous vehicles (AVs), deep analytics to fuse their outputs for a robust perception become imperative. However, existing fusion methods often make two assumptions rarely holding in practice: i) similar data distributions for all inputs and ii) constant availability for all sensors. Because, for example, lidars have various resolutions and failures of radars may occur, such variability often results in significant performance degradation in fusion. To this end, we present t-READi, an adaptive inference system that accommodates the variability of multimodal sensory data and thus enables robust and efficient perception. t-READi identifies variation-sensitive yet structure-specific model parameters; it then adapts only these parameters while keeping the rest intact. t-READi also leverages a cross-modality contrastive learning method to compensate for the loss from missing modalities. Both functions are implemented to maintain compatibility with existing multimodal deep fusion methods. The extensive experiments evidently demonstrate that compared with the status quo approaches, t-READi not only improves the average inference accuracy by more than 6% but also reduces the inference latency by almost 15× with the cost of only 5% extra memory overhead in the worst case under realistic data and modal variations.

Challenges with community mobility are among the most prevalent disabilities worldwide, yet the current standard of care-passive orthotics-have remained largely unchanged for hundreds of years. Powered orthoses have been developed to address these shortcomings, but challenges with reliability, safety, weight, noise, and cost have stunted commercial translation. In this work, we present the Variable Stiffness Orthosis (VSO), a quasi-passive, ankle-foot orthosis that strikes a balance between powered and passive orthoses in terms of functionality and commercial practicality. The VSO can render customized, nonlinear torque-angle relationships via passive cam-based modules, which can feature extreme or even negative stiffness. A passive cam switching mechanism also decouples energy storage and return, allowing push off to be augmented with energy recycled from early stance phase, and changing equilibrium angle to simultaneously promote swing-phase foot clearance and standing stability. The VSO also features step-to-step stiffness adjustments spanning the softest to stiffest commercial AFOs via a motorized spring support. Pilot testing was performed on two participants with and without sciatic nerve injury (SNI). Both participants had activity-dependent stiffness preferences that spanned a large stiffness range. Preliminary results showed that using the VSO led to reduced foot drop, increased self-selected speed, increased total ankle moments, reduced biological moments, reduced toe striking, and reduced steppage in the participant with SNI compared to daily-use AFO and shoes-only conditions.

With the rapid development of satellite communication systems, the development of cognitive radios for satellite systems is significant. However, traditional data sets are no longer able to meet the unique challenges of signal identification and demodulation in satellite channels. This paper proposes a general dataset for satellite cognitive radio (RML24), the first dataset designed specifically for deep learning applications in satellite signal identification and demodulation. RML24 integrates telemetry and communication signals in Telemetry, Tracking, and Command (TT & C) systems and simulates the effects of signal impairments in real satellite channels. The dataset utilizes a software-defined radio (SDR) platform and radio frequency (RF) transceivers to perform rigorous over-the-air measurements and validate the collected data. RML24 provides researchers with fundamental data and modeling benchmarks for widely used modulation classification models in the hope of facilitating the algorithmic validation and development of intelligent and adaptive satellite communication systems and advancing the development of data-driven satellite communication technologies. The experimental results demonstrate that RML24 is more consistent with the data distribution of real-world signals, and the trained model is more capable of generalization.

Contextual reasoning in language models has historically faced challenges in maintaining coherence over extended sequences and adapting to evolving contextual requirements. Context-Dependent Memory Synthesis (CDMS) offers a novel mechanism that integrates dynamic memory pathways into transformer-based architectures, enabling improved retention, retrieval, and synthesis of contextual information. The framework incorporates multi-layered memory modules and adaptive attention mechanisms, allowing language models to dynamically encode and retrieve contextually relevant information across diverse linguistic tasks. Experimental evaluations demonstrated significant reductions in perplexity and enhanced memory utilization, highlighting the efficiency of CDMS in handling longterm dependencies. Robustness against noisy inputs was achieved through dynamic memory updates, ensuring coherence even in challenging scenarios with imperfect data. Scalability assessments revealed consistent performance across varying model sizes, with notable improvements in latency and throughput metrics. The generalization capabilities of the architecture were validated through tasks spanning scientific, legal, and conversational domains, achieving high accuracy and retention rates without extensive fine-tuning. Computational benchmarks confirmed the feasibility of integrating CDMS into existing transformer frameworks, offering a balance between resource efficiency and taskspecific adaptability. Quantitative results also demonstrated low variability in latency across diverse tasks, affirming the system's stability and responsiveness. The findings demonstrate the transformative potential of CDMS in addressing key limitations in traditional language model architectures while providing a robust foundation for future innovations in neural memory systems.

Dementia has had impacts across the world for the last few decades.  The condition is characterized as progressively deteriorating in mental cognition, functional, and behavioral aspects of life.  Caring for those with this condition is a difficult, distressing, and expensive process.  This paper’s purpose is to review technological interventions for dementia and future directions with this idea.  In this paper, many interventions, most specifically online games, virtual reality, and augmented reality will be explored.

With the rapid development of artificial intelligence and the dramatic growth of communication services, the sixth-generation (6G) wireless network needs to handle communication tasks more flexibly and efficiently, significantly exacerbating the challenge of resource allocation. For the access network scenarios in 6G networks, the existing single-layer reinforcement learning resource allocation algorithms are hard to satisfy the diverse demands of users due to the complex and variable state space. Therefore, we propose a reinforcement learning-based two-timescale resource allocation scheme, aiming to jointly enhance the quality of service and system resource utilization. The proposed method comprises an upper-layer controller that allocates network resources to lower-layer controllers on a large time scale. Then, lower-layer controllers refine the resources based on user service types on a smaller time scale. To implement the proposed two-timescale allocation scheme, we propose a two-layer reinforcement learning framework consisting of a deep deterministic policy gradient (DDPG) and a dueling deep Q network (Dueling-DQN). Furthermore, recognizing that coupling multiple reinforcement learning processes may slow down algorithm convergence, we employ asynchronous training, transfer learning, and prediction-based action space simplification to expedite the model’s convergence speed. Finally, we build a phototyping network to verify the performance of the proposed small-timescale and the large-timescale allocation algorithms. Our proposed scheme demonstrates significant improvements in both resource utilization and quality of service compared to existing schemes.

Besides extremely high throughput, Wi-Fi 7 is also aimed at providing users a more deterministic behavior, characterized by shorter average latency and smaller jitters. A key mechanism to achieve this is multi-link operation, which brings simultaneous multi-band communication to client stations as well. In this paper, traffic steering policies are briefly reviewed and grouped into general classes, each one with its advantages and limitations. A basic mechanism for supporting dynamic steering is then described, which is simple enough to allow implementation in real Wi-Fi chipsets but highly flexible at the same time. Its operation can be driven by the host on a per-packet basis, and this permits to optimize spectrum usage depending on the requirements of applications and the traffic pattern they generate.

Executable QR codes, also known as eQR codes or just sQRy, are a special kind of QR codes that embed programs conceived to run on mobile devices like smartphones. Since the program is directly encoded in binary form within the QR code, it can be executed even when the reading device is not provided with Internet access. The applications of this technology are manifold, and range from smart user guides to advisory systems. The first programming language made available for eQR is QRtree, which enables the implementation of decision trees aimed, for example, at guiding the user in operating/maintaining a complex machinery or for reaching a specific location. In this work, an additional language is proposed, we term QRind, which was specifically devised for Industry. It permits to integrate distinct computational blocks into the QR code, e.g., machine learning models to enable predictive maintenance and algorithms to ease machinery usage. QRind permits the Industry 4.0/5.0 paradigms to be implemented, in part, also in those cases where Internet is unavailable.

The potential of AI in medication prediction through Electronic Health Records (EHRs) is a subject of substantial interest and ongoing research. This review critically examines the challenges and future directions needed to transform AIbased medication prediction, which is a hype in its current state rather than a practical tool for real-world application. Unlike previous literature that primarily addresses broad AI challenges, this paper focuses on the intricacies of medication prediction. We structure our analysis around three research questions. RQ1 investigates the prevailing and emerging challenges, highlighting the extensive scope for improvement in model interpretability, scalability, transferability, and domain-specific issues like the cold-start problem, drug interactions, personalization, dosage optimization, etc. RQ2 synthesizes insights from existing research, noting that performance, personalization, and domain-knowledge integration are the most frequently tackled problems, with typical metrics used for evaluation. RQ3 explores future developments required to move from theoretical potential to practical implementation, emphasizing the necessity of addressing data standardization, scalability, interpretability, and other domainspecific challenges. We also discuss the implications of these advancements for various stakeholders, including AI practitioners, medical professionals, and interdisciplinary collaborators.

Dynamic Neural Embedding (DNE) introduces a novel approach to language modeling by dynamically adjusting word embeddings based on contextual information, thereby enhancing the model's ability to capture semantic nuances and improving performance across various natural language processing tasks. Comprehensive empirical evaluations demonstrate that DNE achieves higher accuracy and F1 scores on benchmarks such as GLUE and SQuAD compared to established baseline models. Additionally, DNE exhibits robust adaptability across diverse domains, maintaining high performance with minimal fine-tuning, and effectively leverages larger datasets to enhance performance, indicating strong scalability with increasing data volumes. These findings underscore the potential of DNE to advance the field of language modeling, offering a promising direction for future research and development in natural language processing.

Large language models (LLMs) for recommendation, aka LLM4Rec, has attracted increasing attention due to their advanced knowledge and powerful reasoning ability. However, it raises concerns about the high computational demands, storage costs, and response latency due to the massive parameters in LLMs, prompting the development of numerous efficiency-oriented solutions. Therefore, we, for the first time, present a comprehensive and systematic review of efficient LLM4Rec solutions, whereby we organize the literature in a hierarchical taxonomy aligning with RSs' pipeline (data processing → model design → learning strategy) and different phases (training → serving). Specifically, we first categorize existing methods into data-, model-, and strategy-level ones, and further differentiate each category based on their contributions to various phases of RSs. Secondly, we provide an in-depth analysis of their solutions and technical contributions, highlighting both their inherent connections and key differences. Finally, we evaluate the pros and cons of each category of methods accordingly, present valuable insights, and suggest future research directions.

Measurement of a qubit (quantum bit) is a key element of quantum computation and is necessary for extracting information from a quantum processor. The physical implementation of a measurement protocol involves monitoring a physical quantity of a sensor, such as a current or a voltage, that is correlated with the state of the qubit to be measured. The monitoring of these sensors used to measure qubits in modern quantum computer architectures is typically divided between RF reflectometry and current/voltage monitoring. There are tradeoffs involved with both approaches for the reliable and accurate readout of qubits when it comes to scalable quantum computing architectures. In this article, we will review the working principles of a selection of sensors that are used to measure the state of qubits in semiconductors. The goal is to introduce electronic engineers with little or no background in quantum computing to the working principles of charge sensors in semiconductor quantum computers, their identification in various physical systems, and an overview of their design tradeoffs when measuring a qubit. This will allow electronic engineers to begin working with physics literature in the field of charge sensing. An overview of the article and the topics that will be discussed is given in Fig. 5.

Traffic congestion is a critical challenge in urban environments, leading to increased delays, fuel consumption, and inefficiencies in logistics. Traditional traffic light control methods, which rely on preset schedules and historical data, often fail to adapt to real-time traffic conditions, resulting in suboptimal performance. Recent advances in vehicular communication and computational methods offer new opportunities to enhance traffic signal control. This paper explores the application of Deep Q-Networks (DQN) to optimize traffic signal timings at a single intersection. By leveraging reinforcement learning (RL), I aim to address the limitations of traditional methods, such as large state spaces and slow convergence, which arise from high-dimensional data. I compare DQN against various baseline approaches, including Random, Periodic Cycle, Greedy algorithms, and Naive Q-learning. My findings show that DQN outperforms these methods in minimizing overall vehicle waiting times and ensuring a fairer distribution of wait times across vehicles. The results highlight the potential of DQN for improving traffic flow, laying the foundation for future work in multi-intersection scenarios and more complex urban environments.

Person re-identification (Re-ID) aims to recognize individuals across multiple camera views, presenting significant challenges due to variations in lighting, occlusions, and viewpoints. Traditional methods, including Principal Component Analysis (PCA) and various feature extraction techniques such as Local Binary Patterns (LBP) and Scale-Invariant Feature Transform (SIFT), often rely on assumptions of consistent external conditions which are seldom met in real-world scenarios. Recent advances leveraging deep learning approaches, such as convolutional neural networks (CNNs) and Siamese neural networks, show promise but encounter limitations with small and medium-sized datasets. This paper presents a novel approach to person Re-ID by implementing a Siamese neural network, designed to improve recognition accuracy by learning discriminative features from pairs of images, irrespective of varying conditions. We also explore the performance of a base model utilizing VGG16 architecture with batch normalization, comparing it to the Siamese network to address its shortcomings in generalization on limited datasets. Our findings aim to provide insights into enhancing Re-ID systems' robustness and efficacy across diverse and challenging surveillance environments.