Artificial Intelligence has enabled scientists to rapidly sift through and analyze massive collections of images, helping to identify objects worthy of closer study such as supernovae, pulsars, and quasars, and allowing us to classify stars, label galaxies, and other astronomical systems. Stellar classification serves as a cornerstone of astronomy, providing a framework for understanding and characterizing the diversity of celestial objects. This classification is based on the spectral properties. AI algorithms, specifically machine learning and deep learning models, have been designed to learn from data and make predictions or classifications based on patterns and trends. These models of spectral and stellar classification are trained on large datasets of known celestial objects, enabling them to recognize and categorize new objects based on their spectral signatures. Here we present a novel Fine Tunned Deep Convolutional Neural Network of 1D Separable Convolutional blocks for stellar classification based on spectral characteristics using SSDS-17 data captured by the Sloan Digital Sky Survey, where the class imbalance is assessed using the SMOTE balancing technique. The results obtained during the performance evaluation confirmed the reliability of the proposed architecture of StellarNet in multi-class stellar classification, obtaining remarkable values of around 97% and 99% for accuracy and AUC score respectively. The proposed StellarNet architecture can be applied to enable real-time classification of captured spectral characteristic data, facilitating automation of the task and providing labeled data for study and future research.

This paper introduces an innovative approach to task completion analysis in multi-agent systems using large language models (LLMs) [3]. Unlike traditional methods, which rely on binary indicators of task completion, our approach emphasizes iterative, context-driven task distribution and execution [7]. Tasks are dynamically redistributed by the Centralized Control Unit (CCU), which operates as a coordinator rather than a final decision-maker on task completion. The thematic context of the first incoming task serves as a guiding framework, allowing agents to iteratively refine their operations and share progress through language-based outputs [1]. The system leverages LLM-generated insights to assess task states, recommend further actions, and dynamically adapt task assignments. Task completion is not determined by static metrics but through linguistic confirmations and iterative feedback loops. This adaptive framework ensures unresolved tasks continue to circulate until a comprehensive resolution is achieved, enabling seamless management of evolving and interdependent tasks in multi-agent environments. The proposed approach offers significant advantages in dynamic environments by fostering adaptability, mitigating bottlenecks, and providing a scalable solution for managing complex, context-dependent tasks across various domains.

Semi-supervised learning (SSL) enables the accurate segmentation of medical images with limited available labeled data. However, its performance usually lags fully supervised methods that require the whole dataset to be labeled. We propose a novel SSL framework that narrows the gap between SSL and fully supervised approaches significantly, while using less than a quarter of labeled data. Our approach is driven by a knowledge exchange process between two networks based on a novel certainty-guided contrastive learning strategy that mitigates the impact of inaccurate pseudo labels and of class imbalance. Building on these, we employ a cross supervised contrastive learning across multiple scales that is able to learn hierarchical features reflecting interrelationships both within and across slices and cases. The computational efficiency of our contrastive learning is boosted by novel sampling strategies that select few representative samples for contrasting, as well as a negative memory bank that increases diversity and eliminates the dependence on batch size. We perform an extensive evaluation on three challenging benchmarks, and the experimental results show that our approach achieves state-ofthe art results. We also show it yields improved accuracy when combined with diverse SSL frameworks, and conduct a detailed ablation study showing the benefits of different components of our model. Our code will be made available publicly following acceptance.

High-density surface EMG (HD-sEMG) is gaining attention because of its non-invasive nature and high spatial resolution. However, wearable HD-sEMG measurements with over 128 channels face challenges in system integration and reliable network-free inter-device synchronization. This article presents a wireless distributed wearable HD-sEMG acquisition system named Oct-HD, which supports up to eight acquisition modules (512 channels) with microsecond-level synchronization without requiring a network. The full-channel impedance detection ensures reliable electrode-skin contact and signal acquisition. The system also includes a self-locking base station that stores, charges, and configures the modules. Synchronization validation shows a longterm inter-module offline error within 2 ms (0.139 ppm) under vibrations and extreme temperatures, demonstrating reliability for network-free outdoor and open-space monitoring. Comparative experiments with a commercial system (Sessantaquattro) involving ten hand postures and 17 subjects were conducted, as hand gesture recognition is one of the most common applications in the field of electromyography. Results showed a significant performance improvement (p < 10 −5) of Oct-HD over the state-of-theart system in signal quality and anti-interference capacity. Gesture classification accuracy increased significantly for both the dynamic task (p = 0.008) and the maintenance task (p = 0.003). This highlights the effectiveness of Oct-HD in enhancing applications in prosthetic control and human-computer interaction. The Oct-HD system offers a notable advancement in wireless HD-sEMG acquisition, offering superior channel capacity, synchronization precision, signal quality, and anti-interference capacity compared to existing systems. The network-free microsecond-level synchronization and enhanced signal performance provide greater monitoring flexibility across various muscle groups, paving the way for broader applications in human-machine interaction.

We investigate the problem of reconfigurable intelligent surface (RIS)-aided monostatic sensing of multiple mobile targets under line-of-sight (LoS) blockage using a dual-functional radar-communications base station. For the purpose of target detection and delay/Doppler/angle estimation, we derive a detector based on the generalized likelihood ratio test (GLRT), which entails a high-dimensional parameter search with high complexity due to angle-Doppler coupling. To circumvent the coupling effect, we introduce a repetitive RIS phase profile design accompanied by a multi-target detection/estimation algorithm based on orthogonal matching pursuit where each iteration solves the single-target GLRT problem and outputs the corresponding parameter estimates. Moreover, as a low-complexity alternative, we introduce a novel subspace-based algorithm based on twodimensional unitary ESPRIT to retrieve target parameters in a non-iterative manner. Simulation results verify the effectiveness of the proposed algorithms, demonstrating their convergence to theoretical bounds and revealing significant improvements over state-of-the-art benchmarks in complexity and performance.

Snow over sea ice affects the overall radiative balance of the polar region by modulating the planetary albedo and influencing sea ice growth. In satellite altimetry, snow is a major source of uncertainty in sea ice thickness estimates, especially with Ku-band radar altimeters. This is because conventional algorithms assume the primary scattering occurs at the snow/ice interface which may not hold for all snowpacks. In particular, complex snow morphological features, including wet snow and/or ice lenses, can shift the main scattering horizon toward the snow surface. These snow features may be particularly prominent over Antarctic sea ice. This study investigates how the properties of the snowpack in the Weddell Sea influence Ku-and Ka-band altimeter waveforms, particularly in the context of extreme conditions and complex snow stratigraphy. Using an adapted waveform model, we deconvolute radar echoes into contributions from the snow surface, snow volume, and ice surface, employing both least-squares fitting and Convolutional Neural Networks (CNN). Our findings challenge previous default assumptions, revealing that snow-volume scattering is as significant as that from the ice surface. To better characterize the snow scattering effects, we propose an adapted threshold first-maximum retracker algorithm (TFMRA) for CryoSat-2 radar freeboard retrieval, identifying a 70% threshold for ice floe retracking as the most accurate based on Operation IceBridge data. The algorithm, applied in CRYO2ICE campaigns and validated against snow buoy measurements, highlights the critical role of snow in altimeter retrievals. The result is particularly important given the recent unprecedented Antarctic sea ice loss and its application for future satellite missions like ESA's CRISTAL.

The risk of cascading failures in power electronic systems is mitigated by detecting damaged semiconductor devices before energizing power circuits. Existing fault diagnostic approaches require the energization of the power circuit and/or specialized sensing and conditioning arrangements. In this paper, a technique to detect faulty MOSFET switching cells, prior to power-loop energization and using no additional hardware, is proposed. The proposed method exploits a gate driver-induced residual voltage (GIRV), appearing across the de-energized DC terminals of a switching cell, to distinguish between healthy and damaged cells and thus, identify failures. To quantify the GIRV, the operation of a de-energized switching cell is analyzed in the time domain. Short-circuit, open-circuit, and gate failure conditions are then analyzed and the implementation of the GIRV-based failure detection scheme outlined. The proposed scheme can be implemented using only the MOSFET gate drivers and a DC bus voltage sensor. The analysis of the GIRV is verified in circuit simulations and experiments. Further, the proposed failure detection scheme is experimentally validated on existing hardware by modifying the software start-up routines.

This paper proposes a near-field Euler rotation-based technique to transform the far-field calculation of arbitrary conformal arrays into the pattern-multiplication form. Additionally, the calculation is greatly accelerated by a three-dimensional virtual equivalent source expansion (3-D VESE) technique and a layer-wise two-dimensional fast Fourier transform (2-D FFT) technique. Computational complexity analysis and several numerical examples validate the advantages of the proposed solution in terms of the efficiency and accuracy.

We consider the effectiveness of multi-objective counterfactual explanations (MOCE) in helping individuals learn tactics, or rules of thumb, to apply when required to select a course of action in a specific context. In this setting, a counterfactual explanation compares one course of action against another. A MOCE presents this comparison by highlighting how the two options differ across a range of objectives or metrics. We conduct a study in which participants are presented with various scenarios, alongside courses of action that could be implemented in those scenarios. Counterfactual explanations, including those involving multiple objectives, are used to identify the positive and negative aspects of the provided options. Participants were then required to identify the best course of action in a range of contexts. Participants trained with MOCE outperformed those given no explanations in seven of eight scenarios and those given single-objective explanations (SOCE) in four. SOCEs gave participants an aggregated outcome (expected rewards) without breaking these into specific objectives. MOCE improved tactic learning, but participants provided with SOCE or no explanation performed better in multi-tactic scenarios. These findings suggest that MOCE enhances tactical decision-making, but further research is needed for multi-tactic integration.

This paper investigates the effects of non-ideal Digital-to-Analog Converters (DACs) on the modulated signal in Filter Bank Multi-Carrier (FBMC) transmitters. The mathematical model of the DAC for FBMC signals is described and evaluated. First, a closed-form expression for the optimal clipping threshold is given, considering uniform quantization in the DACs as a function of the bit resolution. Then, a further optimization for the placement of the quantization levels is proposed, considering non-uniform quantization to maximize the Signal-to-Distortion and Quantization Ratio (SDQR) at the output of the DACs. Finally, an upper limit for the SDQR is also presented using a Lloyd-Max quantizer.

In design and optimization of intensive care unit (ICU) networks, one common practice is to prioritize the treatment for patients of higher emergency levels, while ensuring fairness to other patients by guaranteeing a certain Quality of Service (QoS) level. One common approach to realize such priority arrangement is bed reservation policy, which designates a certain number of last occupied beds in each hospital to be exclusively used by certain patient classes. In this paper, we propose an approach that can significantly improve the computational efficiency in obtaining the optimal reservation thresholds for each patient class given their respective requirements, in a non-hierarchical ICU model (where the external emergency patients can possibly be allocated to any ICU hospital) which has been shown to be computationally challenging in performance evaluation and optimization. Specifically, we apply the Information Exchange Surrogate Approximation (IESA) to analytically approximate the key QoS metrics under given reservation thresholds, and the integer Particle Swarm Optimization (PSO) algorithm to search for the optimal threshold based on the approximation results by IESA. We demonstrate numerically, with the real data from ICUs in Hong Kong, that IESA approximation can obtain reasonably accurate results for QoS metrics, and thus lead to accurate optimal reservation thresholds. In addition, our proposed approach combing IESA and PSO can significantly reduce the computation time by more than four orders of magnitude, compared to the state-of-the-art evaluation and optimization methods in existing research for similar problems, especially for ICU networks with practical sizes.

The inconsistency between training strategies and task objectives is a principal constraint impeding DensePose development. DensePose aims to minimize the surface distance between the estimations and the ground truth, thereby establishing a correspondence between two-dimensional images and threedimensional humans. However, the discrete IUV optimization strategy employed by the existing pipeline deviates from the goal of minimizing surface distance, thus yielding sub-optimal results. To solve this problem, we propose Geodesic-consistent RCNN (GC RCNN), which models the intrinsic interdependence within the IUV to enhance human surface understanding and facilitate surface distance optimization. GC RCNN incorporates the adaptively semantic enhancement module (ASEM) and the geodesic-consistent loss (GCL) into the traditional DensePose estimation framework. ASME enhances human surface semantics by extracting distinct IUV features and dynamically modeling the feature-level IUV interdependence. GCL extracts the surface distance information by quantifying the degree of deviation between the estimated IUV distribution and the ground truth. The distribution deviation is utilized to refine the IUV optimization process, thereby aiding in minimizing surface distance. Empirical analyses conducted on the DensePose-COCO dataset validate the superior performance of GC RCNN, which surpasses the base model by a margin of up to 3.3% AP.

This study explores optimizing large language models using the Low-Rank Adaptation (LoRA) technique, focusing on how different values of the r parameter affect model performance. We started by fine-tuning Gemma 2B v1.1 Instruct for mental health support while maintaining its original capabilities. Using a subset of the SQuAD dataset, we evaluated the model with various r values and found that performance differences were minimal. Next, we fine-tuned the model with a conversational dataset tailored for mental health and conducted qualitative assessments. These assessments revealed significant improvements in the quality and relevance of conversations in the mental health domain. In conclusion, our study found that LoRA is an efficient method for fine-tuning large models, with an optimal r value of 16, striking a balance between adaptation and preserving the base model's capabilities.

This article presents a method of sickle cell detection from microscopic images. We extract five attribute values from the connected components of an image, and train machine learning classifiers to recognize the sickle cells. Four classifiers were experimented with and the vest one was the K-Nearest neighbor classifier with 97.3% accuracy. The other classifiers are the Neutral network, Decision tree and Naïve Bayesian classifiers which resulted in accuracy rates of between 89-96.3%. This method is applicable for use in low cost computers since it is computationally cheap. The findings of this research can be considered as a screening method for diagnosing sickle cell aneamia.

This paper introduces a cyber-physical test bench for metropolitan area digital substations, which serves as a platform for studying advanced cyber-attacks and evaluating intrusion detection systems. The test bench provides a hardwarein-the-loop environment with a high Technology Readiness Level (TRL). Additionally, the paper demonstrates the testbed's capabilities by testing and validating an AI-based intrusion detection system developed by an industrial company. The demonstration includes testing the cybersecurity tool against a wide range of cyber-attacks, such as false data injection, packet replay, and time desynchronization, on relevant substation automation protocols commonly used in EU power grid infrastructures like IEC-60870-5-104 and IEC61850. The results of the demonstration indicate that the test bench effectively simulates realistic impacts on substation operation when subjected to attacks, thus providing the opportunity to validate the anomaly detection solution.

Specific emitter identification (SEI) is a unique physical-layer security technology that plays a crucial role in protecting wireless communication systems from various security threats. Although SEI based on artificial neural network models has achieved good identification performance, its performance degrades when labeled samples are limited. To address this issue, this letter proposes a few-shot SEI method based on a contrastive masked learning framework. This method combines contrastive learning and masked learning to enhance the model’s representation capability, and it consists of an encoder, a signal decoder, a feature decoder, and a momentum encoder. Simulation experiments on the open-source datasets LoRa and ADS-B show that the proposed method outperforms other SEI methods.

Short-packet communication (SPC) and unmanned aerial vehicles (UAVs) are anticipated to play crucial roles in the development of 5G-and-beyond wireless networks and the Internet of Things (IoT). In this paper, we propose a secure SPC system, where a UAV serves as a mobile decode-and-forward (DF) relay, periodically receiving and relaying small data packets from a remote IoT device to its receiver in two hops with strict latency requirements, in the presence of an eavesdropper. This system requires careful optimization of important design parameters, such as the coding blocklengths of both hops, transmit powers, and the UAV’s trajectory. While the overall optimization problem is nonconvex, we tackle it by applying a block successive convex approximation (BSCA) approach to divide the original problem into three subproblems and solve them separately. Then, an overall iterative algorithm is proposed to obtain the final design with guaranteed convergence. Our proposed low-complexity algorithm incorporates robust trajectory design and resource management to optimize the effective average secrecy throughput of the communication system over the course of the UAV-relay’s mission. Simulation results demonstrate significant performance improvements compared to various benchmark schemes and provide useful design insights on the coding blocklengths and transmit powers along the trajectory of the UAV.

The demand for efficient edge machine learning (ML) applications has driven interest in Microcontroller Unit (MCU)-based TinyML solutions, especially with the rise of RISC-V’s Instruction Set Architecture. While MCUs are power- efficient, their limited resources challenge the deployment of complex ML models. Mixed-Precision Quantization (MPQ) can achieve the best trade-off between model size, energy consump- tion, and accuracy, by using different weights and activations precision across model layers. However, MCU-class processors often lack hardware support for MPQ. We present an end-to-end flow from training to hardware deployment designed to efficiently run Mixed-Precision Quantized Neural Networks (MP-QNNs) on MCU-class processors. Central to our approach is STAR- MAC, a precision-scalable Multiply-and-Accumulate unit that supports flexible MPQ operations on 16-, 8-, and 4-bit integer data. STAR-MAC combines two subword-parallel techniques, Sum-Together and Sum-Apart, within a unified multiplier architecture that effectively reconfigures for Fully-Connected, 2D Convolution, and Depth-wise layers. We integrate STAR-MAC into the low-power RISC-V Ibex core and validate our flow using a Field-Programmable Gate Array-based System-on-Chip setup. Inference results over the four MLPerf Tiny MP-QNN models using our modified TensorFlow Lite for Microcontrollers (TFLM) deployment flow, show an average latency reduction of 68% with respect to their 8-bit counterparts using the standard TFLM runtime, with a reduced flatbuffer size of 27% on average, and with little to no accuracy drop. Synthesis on 28-nm CMOS technology indicates limited area and power overhead (+11.2%, +8.5%) over the original Ibex. We open-source our framework to encourage further development of MP-QNNs on RISC-V MCUs.