Search Results (496)

We propose a novel hybrid deep learning framework that synergistically integrates Convolutional Neural Networks (CNNs), Spiking Neural Networks (SNNs), and Generative Adversarial Networks (GANs) for robust and accurate classification of high-resolution frontal and sagittal human gait video sequences—capturing both lower-limb kinematics and upper-body posture—from subjects with Knee Osteoarthritis (KOA), Parkinson’s Disease (PD), and healthy Normal (NM) controls, classified into three disease-type categories. Our approach first employs a tailored CNN backbone to extract rich spatial features from fixed-length clips (e.g., 16 frames resized to 128 × 128 px), which are then temporally encoded and processed by an SNN layer to capture dynamic gait patterns. To address class imbalance and enhance generalization, a conditional GAN augments rare severity classes with realistic synthetic gait sequences. Evaluated on the controlled, marker-based KOA-PD-NM laboratory public dataset, our model achieves an overall accuracy of 99.47%, a sensitivity of 98.4%, a specificity of 99.0%, and an F1-score of 98.6%, outperforming baseline CNN, SNN, and CNN–SNN configurations by over 2.5% in accuracy and 3.1% in F1-score. Ablation studies confirm that GAN-based augmentation yields a 1.9% accuracy gain, while the SNN layer provides critical temporal robustness. Our findings demonstrate that this CNN–SNN–GAN paradigm offers a powerful, computationally efficient solution for high-precision, gait-based disease classification, achieving a 48.4% reduction in FLOPs (1.82 GFLOPs to 0.94 GFLOPs) and 9.2% lower average power consumption (68.4 W to 62.1 W) on Kaggle P100 GPU compared to CNN-only baselines. The hybrid model demonstrates significant potential for energy savings on neuromorphic hardware, with an estimated 13.2% reduction in energy per inference based on FLOP-based analysis, positioning it favorably for deployment in resource-constrained clinical environments and edge computing scenarios. Full article

This paper proposes and implements a method to efficiently parallelize constraint solving in rigid body simulation using GPUs. Rigid body simulation is widely used in robot development, computer games, movies, and other fields, and there is a growing need for faster computation. As current computers are reaching their limits in terms of scale-up, such as clock frequency improvements, performance improvements are being sought through scale-out, which increases parallelism. However, rigid body simulation is difficult to parallelize efficiently due to its characteristics. This is because, unlike fluid or molecular physics simulations, where each particle or lattice can be independently extracted and processed, rigid bodies can interact with a large number of distant objects depending on the instance. This characteristic causes significant load imbalance, making it difficult to evenly distribute computational resources using simple methods such as spatial partitioning. Therefore, this paper proposes and implements a computational method that enables high-speed computation of large-scale scenes by hierarchically clustering rigid bodies based on their number and associating the hierarchy with the hardware structure of GPUs. In addition, to effectively utilize parallel computing resources, we considered a more relaxed parallelization condition for the conventional Gauss–Seidel block parallelization method and demonstrated that convergence is guaranteed. We investigated how speed and convergence performance change depending on how much computational cost is allocated to each hierarchy and discussed the desirable parameter settings. By conducting experiments comparing our method with several widely used software packages, we demonstrated that our approach enables calculations at speeds previously unattainable with existing techniques, while leveraging GPU computational resources to handle multiple rigid bodies simultaneously without significantly compromising accuracy. Full article

Learning from Demonstration (LfD) enables robots to acquire manipulation skills by observing human actions. However, existing methods often face challenges such as high computational cost, limited generalizability, and a loss of key interaction details. This study presents a compact representation for interaction recognition in LfD that encodes human–object interactions using 2D wrist trajectories and 3D object poses. A lightweight extraction pipeline combines MediaPipe-based wrist tracking with FoundationPose-based 6-DoF object estimation to obtain these trajectories directly from RGB-D video without specialized sensors or heavy preprocessing. Experiments on the GRAB and FPHA datasets show that the representation effectively captures task-relevant interactions, achieving 94.6% accuracy on GRAB and 96.0% on FPHA with well-calibrated probability predictions. Both Bidirectional Long Short-Term Memory (Bi-LSTM) with attention and Transformer architectures deliver consistent performance, confirming robustness and generalizability. The method achieves sub-second inference, a memory footprint under 1 GB, and reliable operation on both GPU and CPU platforms, enabling deployment on edge devices such as NVIDIA Jetson. By bridging pose-based and object-centric paradigms, this approach offers a compact and efficient foundation for scalable robot learning while preserving essential spatiotemporal dynamics. Full article

With the rapid development of autonomous vehicles and intelligent transportation systems, vehicle detection and distance estimation have become critical technologies for ensuring driving safety. However, real-world in-vehicle environments impose strict constraints on computational resources, making it impractical to deploy high-end GPUs. This implies that even highly accurate algorithms, if unable to run in real time on embedded platforms, cannot fully meet practical application demands. Although existing deep learning-based detection and stereo vision methods achieve state-of-the-art accuracy on public datasets, they often rely heavily on massive computational power and large-scale annotated data. Their high computational requirements and limited cross-scenario generalization capabilities restrict their feasibility in real-time vehicle-mounted applications. On the other hand, traditional algorithms such as Semi-Global Block Matching (SGBM) are advantageous in terms of computational efficiency and cross-scenario adaptability, but when used alone, their accuracy and robustness remain insufficient for safety-critical applications. Therefore, the motivation of this study is to develop a stereo vision-based collision warning system that achieves robustness, real-time performance, and computational efficiency. Our method is specifically designed for resource-constrained in-vehicle platforms, integrating a lightweight YOLOv8n detector with SGBM-based depth estimation. This approach enables real-time performance under limited resources, providing a more practical solution compared to conventional deep learning models and offering strong potential for real-world engineering applications. Full article

In advanced semiconductor manufacturing, computational lithography, particularly sub-resolution assist features (SRAFs), is crucial for enhancing the process window. However, conventional SRAF placement methodologies are hampered by a critical trade-off between speed and pattern fidelity, and they largely fail to optimize the complex, curvilinear layouts essential for advanced nodes. This study develops a deep learning framework to replace and drastically accelerate the optical refinement of SRAF shapes. We established a large-scale dataset with coarse, binarized SRAF patterns as inputs. Ground-truth labels were generated via an Level-Set Method (LSM) optimized purely for optical performance. A U-Net convolutional neural network was then trained to learn the mapping from the coarse inputs to the optically optimized outputs. Experimental results demonstrate a dual benefit: the model provides a multi-order-of-magnitude acceleration over traditional CPU-based methods and is significantly faster than modern GPU-accelerated algorithms while achieving a final pattern fidelity highly comparable to the computationally expensive LSM. The U-Net-generated SRAFs exhibit high fidelity to the ground-truth layouts and comparable optical performance. Our findings demonstrate that a data-driven surrogate can serve as an effective alternative to traditional algorithms for SRAF optical refinement. This represents a promising approach to mitigating computational costs in mask synthesis and provides a solid foundation for future integrated optimization solutions. Full article

This study introduces an explainable neuro-symbolic and large language model (LLM)-driven framework for intelligent interpretation of corneal topography and precision surgical decision support. In a prospective cohort of 20 eyes, comprehensive IOLMaster 700 reports were analyzed through a four-stage pipeline: (1) automated extraction of key parameters—including corneal curvature, pachymetry, and axial biometry; (2) mapping of these quantitative features onto a curated corneal disease and refractive-surgery knowledge graph; (3) Bayesian probabilistic inference to evaluate early keratoconus and surgical eligibility; and (4) explainable multi-model LLM reporting, employing DeepSeek and GPT-4.0, to generate bilingual physician- and patient-facing narratives. By transforming complex imaging data into transparent reasoning chains, the pipeline delivered case-level outputs within ~95 ± 12 s. When benchmarked against independent evaluations by two senior corneal specialists, the framework achieved 92 ± 4% sensitivity, 94 ± 5% specificity, 93 ± 4% accuracy, and an AUC of 0.95 ± 0.03 for early keratoconus detection, alongside an F1 score of 0.90 ± 0.04 for refractive surgery eligibility. The generated bilingual reports were rated ≥4.8/5 for logical clarity, clinical usefulness, and comprehensibility, with representative cases fully concordant with expert judgment. Comparative benchmarking against baseline CNN and ViT models demonstrated superior diagnostic accuracy (AUC = 0.95 ± 0.03 vs. 0.88 and 0.90, p < 0.05), confirming the added value of the neuro-symbolic reasoning layer. All analyses were executed on a workstation equipped with an NVIDIA RTX 4090 GPU and implemented in Python 3.10/PyTorch 2.2.1 for full reproducibility. By explicitly coupling symbolic medical knowledge with advanced language models and embedding explainable artificial intelligence (XAI) principles throughout data processing, reasoning, and reporting, this framework provides a transparent, rapid, and clinically actionable AI solution. The approach holds significant promise for improving early ectatic disease detection and supporting individualized refractive surgery planning in routine ophthalmic practice. Full article

3D object detection and tracking technology are increasingly being adopted in unmanned ground vehicles, as robust perception systems significantly improve the obstacle avoidance performance of a UGV. However, most existing algorithms depend heavily on computationally intensive point cloud neural networks, rendering them unsuitable for resource-constrained platforms. In this work, we propose an efficient 3D object detection and tracking method specially designed for deployment on low-cost vehicle platforms. For the detection phase, our method integrates an image-based 2D detector with data fusion techniques to coarsely extract object point clouds, followed by an unsupervised learning approach to isolate objects from noisy point cloud data. For the tracking process, we propose a multi-target tracking algorithm based on multi-level data association. This method introduces an additional data association step to handle targets that fail in 3D detection, thereby effectively reducing the impact of detection errors on tracking performance. Moreover, our method enhances association precision between detection outputs and existing trajectories through the integration of 2D and 3D information, thereby further mitigating the adverse effects of detection inaccuracies. By adopting unsupervised learning as an alternative to complex neural networks, our approach demonstrates strong compatibility with both low-resolution LiDAR and GPU-free computing platforms. Experiments on the KITTI benchmark demonstrate that our tracking framework achieves significant computational efficiency gains while maintaining detection accuracy. Furthermore, experimental evaluations on the real-world UGV platform demonstrated the deployment feasibility of our approach. Full article

Wearable devices can be developed using hardware platforms such as Application Specific Integrated Circuits (ASICs), Graphics Processing Units (GPUs), Digital Signal Processors (DSPs), Micro controller Units (MCUs), or Field Programmable Gate Arrays (FPGAs), each with distinct advantages and limitations. ASICs offer high efficiency but lack flexibility. GPUs excel in parallel processing but consume significant power. DSPs are optimized for signal processing but are limited in versatility. CPUs provide low power consumption but lack computational power. FPGAs are highly flexible, enabling powerful parallel processing at lower energy costs than GPUs but with higher resource demands than ASICs. The combined use of FPGAs and CPUs balances power efficiency and computational capability, making it ideal for wearable systems requiring complex algorithms in far-edge computing, where data processing occurs onboard the device. This approach promotes green electronics, extending battery life and reducing user inconvenience. The primary goal of this work was to develop a versatile framework, similar to existing software development frameworks, but specifically tailored for mixed FPGA/MCU platforms. The framework was validated through a real-world use case, demonstrating significant improvements in execution speed and power consumption. These results confirm its effectiveness in developing green and smart wearable systems. Full article

Interpreting multi-component ¹H NMR spectra is difficult due to peak overlap, concentration variability, and low-abundance signals. We cast mixture identification as a single-pass multi-label task. A compact CNN–Transformer (“Hybrid”) model was trained end-to-end on domain-informed and realistically simulated spectra derived from a 13-component flavor library; the model requires no real mixtures for training. On 16 real formulations, the Hybrid attains micro-F1 = 0.990 and exact-match (subset) accuracy = 0.875, outperforming CNN-only and Transformer-only ablations, while remaining efficient (~0.47 M parameters; ~0.68 ms on GPU, V100). The approach supports abstention and shows robustness to simulated outsiders. Although the evaluation set was small, and the macro-ECE (per-class, 15 bins) was inflated by sparse classes (≈0.70), the micro-averaged Brier is low (0.0179), and temperature scaling had negligible effect (T ≈ 1.0), indicating the good overall probability quality. The pipeline is readily extensible to larger libraries and adjacent applications in food authenticity and targeted metabolomics. Classical chemometric baselines trained on simulation failed to transfer to real measurements (subset accuracy 0.00), while the Hybrid model maintained strong performance. Full article

Speech recognition, as a key driver of artificial intelligence and global communication, has advanced rapidly in major languages, while studies on low-resource languages remain limited. Tongan, a representative Polynesian language, carries significant cultural value. However, Tongan speech recognition faces three main challenges: data scarcity, limited adaptability of transfer learning, and weak dictionary modeling. This study proposes improvements in adaptive transfer learning and NBPE-based dictionary modeling to address these issues. An adaptive transfer learning strategy with layer-wise unfreezing and dynamic learning rate adjustment is introduced, enabling effective adaptation of pretrained models to the target language while improving accuracy and efficiency. In addition, the MEA-AGA is developed by combining the Mind Evolutionary Algorithm (MEA) with the Adaptive Genetic Algorithm (AGA) to optimize the number of byte-pair encoding (NBPE) parameters, thereby enhancing recognition accuracy and speed. The collected Tongan speech data were expanded and preprocessed, after which the experiments were conducted on an NVIDIA RTX 4070 GPU (16 GB) using CUDA 11.8 under the Ubuntu 18.04 operating system. Experimental results show that the proposed method achieved a word error rate (WER) of 26.18% and a word-per-second (WPS) rate of 68, demonstrating clear advantages over baseline methods and confirming its effectiveness for low-resource language applications. Although the proposed approach demonstrates promising performance, this study is still limited by the relatively small corpus size and the early stage of research exploration. Future work will focus on expanding the dataset, refining adaptive transfer strategies, and enhancing cross-lingual generalization to further improve the robustness and scalability of the model. Full article

The risk of wave overtopping is amplifying under sea-level rise and increased frequency of extreme coastal events. Conventional empirical and physical methods for estimating overtopping characteristics are limited by site-specific assumptions, which underscores the need for robust and efficient approaches. This study develops a multiscale numerical modeling framework that couples the regional ADCIRC–UnSWAN (Advanced CIRCulation and Unstructured Simulating WAves Near-shore) model with DualSPHysics (SPH) model to simulate overtopping responses under varying sea states. ADCIRC-UnSWAN provides regional-scale hydrodynamic and wave forcing, which is nested into localized SPH model to resolve wave-structure interactions. The proposed framework accurately reproduces overtopping responses including water thickness and velocity while leveraging GPU acceleration for computational efficiency. The model outputs are further analyzed to classify overtopping hazard levels and perform probabilistic pedestrian risk as sessments that account for uncertainties in wave characteristics and human vulnerability. The results supports the development of early warning systems and provide a foundation for dynamic hazard level updates in real or near-real time, contributing to improved coastal risk governance under uncertainties. Full article

This study proposes a 3D reconstruction method for LIDAR building point clouds using geometric primitive constrained optimization. It addresses challenges such as low accuracy, high complexity, and slow modeling. This new algorithm studies the reconstruction of point clouds at the level of geometric primitives and is an incremental joint optimization method based on the GPU rendering pipeline. Firstly, the building point cloud collected by the LIDAR laser scanner was preprocessed, and an initial building mesh model was constructed by the fast triangulation method. Secondly, based on the geometric characteristics of the building, geometric primitive constrained optimization rules were generated to optimize the initial mesh model (regular surface optimization, basis spline surface optimization, junction area optimization, etc.). And a view-dependent parallel algorithm was designed to optimize the calculation. Finally, the effectiveness of this approach was validated by comparing and analyzing the experimental results of different buildings’ point cloud data. This algorithm does not require data training and is suitable for outdoor surveying and mapping engineering operations. It has good controllability and adaptability, and the entire pipeline is interpretable. The obtained results can be used for serious applications, such as Building Information Modeling (BIM), Computer-Aided Design (CAD), etc. Full article

Belief Propagation (BP) is a fundamental heuristic for solving Constraint Optimization Problems (COPs), yet its practical applicability is constrained by slow convergence and instability in loopy factor graphs. While Damped BP (DBP) improves convergence by using manually tuned damping factors, its reliance on labor-intensive hyperparameter optimization limits scalability. Deep Attentive BP (DABP) addresses this by automating damping through recurrent neural networks (RNNs), but introduces significant memory overhead and sequential computation bottlenecks. To reduce memory usage and accelerate deep belief propagation, this paper introduces Fast Deep Belief Propagation (FDBP), a deep learning framework that improves COP solving through online self-supervised learning and graphics processing unit (GPU) acceleration. FDBP decouples the learning of damping factors from BP message passing, inferring all parameters for an entire BP iteration in a single step, and leverages mixed precision to further optimize GPU memory usage. This approach substantially improves both the efficiency and scalability of BP optimization. Extensive evaluations on synthetic and real-world benchmarks highlight the superiority of FDBP, especially for large-scale instances where DABP fails due to memory constraints. Moreover, FDBP achieves an average speedup of 2.87× over DABP with the same restart counts. Because BP for COPs is a mathematically grounded GPU-parallel message-passing framework that bridges applied mathematics, computing, and machine learning, and is widely applicable across science and engineering, our work offers a promising step toward more efficient solutions to these problems. Full article

Staining-based assays are widely used for cell analysis but are invasive, alter physiology, and prevent longitudinal monitoring. Label-free, morphology-based approaches could enable real-time, non-invasive drug testing, yet detection of subtle and dynamic changes has remained difficult. We developed a deep learning framework for stain-free monitoring of leukemia cell cultures using automated bright-field microscopy in a semi-automated culture system (AICE3, LABMaiTE, Augsburg, Germany). YOLOv8 models were trained on images from K562, HL-60, and Kasumi-1 cells, using an NVIDIA DGX A100 GPU for training and tested on GPU and CPU environments for real-time performance. Comparative benchmarking with RT-DETR and interpretability analyses using Eigen-CAM and radiomics (RedTell) was performed. YOLOv8 achieved high accuracy (mAP@0.5 > 98%, precision/sensitivity > 97%), with reproducibility confirmed on an independent dataset from a second laboratory and an AICE3 setup. The model distinguished between morphologically similar leukemia lines and reliably classified untreated versus differentiated K562 cells (hemin-induced erythroid and PMA-induced megakaryocytic; >95% accuracy). Incorporation of decitabine-treated cells demonstrated applicability to drug testing, revealing treatment-specific and intermediate phenotypes. Longitudinal monitoring captured culture- and time-dependent drift, enabling separation of temporal from drug-induced changes. Radiomics highlighted interpretable features such as size, elongation, and texture, but with lower accuracy than the deep learning approach. To our knowledge, this is the first demonstration that deep learning resolves subtle, drug-induced, and time-dependent morphological changes in unstained leukemia cells in real time. This approach provides a robust, accessible framework for label-free longitudinal drug testing and establishes a foundation for future autonomous, feedback-driven platforms in precision oncology. Ultimately, this approach may also contribute to more precise and adaptive clinical decision-making, advancing the field of personalized medicine. Full article

Reference-less evaluation of machine translation, or Quality Estimation (QE), is vital for low-resource language pairs where high-quality references are often unavailable. In this study, we investigate segment-level QE methods comparing encoder-based models such as MonoTransQuest, CometKiwi, and xCOMET with various decoder-based methods (Tower+, ALOPE, and other instruction-fine-tuned language models). Our work primarily focused on utilizing eight low-resource language pairs, involving both English on the source side and the target side of the translation. Results indicate that while fine-tuned encoder-based models remain strong performers across most low-resource language pairs, decoder-based Large Language Models (LLMs) show clear improvements when adapted through instruction tuning. Importantly, the ALOPE framework further enhances LLM performance beyond standard fine-tuning, demonstrating its effectiveness in narrowing the gap with encoder-based approaches and highlighting its potential as a viable strategy for low-resource QE. In addition, our experiments demonstrates that with adaptation techniques such as LoRA (Low Rank Adapters) and quantization, decoder-based QE models can be trained with competitive GPU memory efficiency, though they generally require substantially more disk space than encoder-based models. Our findings highlight the effectiveness of encoder-based models for low-resource QE and suggest that advances in cross-lingual modeling will be key to improving LLM-based QE in the future. Full article