Search Results (795)

Recent advances in large language model (LLM) agents have shown promise for autonomous decision-making in recommender systems. However, existing approaches suffer from two fundamental limitations: flat agent memories that conflate different information modalities and prohibitive computational costs that prevent scaling beyond a few hundred users. We propose Hybrid-GraphRAG, a recommender system that integrates hierarchical agent memory structures, graph-based retrieval-augmented generation (Graph RAG), and knowledge distillation for scalable deployment. Our approach extends agent-based collaborative filtering by structuring agent memories into intrinsic, collaborative, and interaction tiers that disentangle different information types; performing multi-hop retrieval over a dynamically constructed heterogeneous interaction graph to enable relational reasoning; and distilling LLM-generated memory dynamics into efficient graph neural encoders with adaptive gating between full and efficient inference paths. Experiments on Amazon review datasets (CDs and Vinyl, Office Products) demonstrate that Hybrid-GraphRAG achieves recommendation quality comparable to full LLM-based agents while reducing computational cost by 85% and improving NDCG@10 by 12.7% over flat-memory agent baselines. Our results establish a principled bridge between semantic agent reasoning and scalable graph-based recommendation. Full article

Finite-control-set model predictive control provides excellent torque–speed regulation for switched reluctance motor drives but requires an online combinatorial search at every control instant, making low-cost embedded implementation challenging. This article investigates whether offline reinforcement learning can distill policies from multiple classical controllers into a single feedforward policy requiring neither online optimization nor controller gain tuning. A replay buffer is populated with trajectories generated by three demonstrators—hysteresis current control, proportional–integral control with pulse-width modulation, and finite-control-set model predictive control—using a finite-element model of a four-phase 8/6 switched reluctance machine parameterized from measurements of the physical drive. An implicit Q-learning agent then learns a control policy without evaluating actions outside the offline dataset. The central finding is that demonstration diversity governs the stability of offline reinforcement learning on this problem: policies trained from a single demonstrator experience early mode collapse in all fifteen runs, whereas two- or three-demonstrator datasets converge stably in all fifteen. Behavior cloning trained on the identical buffer, split, architecture, and deployed controller provides the reference point for interpreting this result. It matches the offline RL policy on torque quality and improves on its speed regulation, exhibiting none of the seed-to-seed fragility seen at no load while requiring roughly $8\%$ more switching transitions. The stability requirement therefore appears to be a property of the advantage-weighted offline RL objective rather than the control task, and the measured benefit of that objective on this problem is confined to switching effort. We report this rather than claim a broader advantage. The characterization of the distilled controller shows that it generalizes to operating points that are not included in the training dataset, gains nothing systematic beyond approximately 60% of the replay buffer, remains insensitive to $\pm 20\%$ perturbations of all reward weights, and degrades gracefully under measurement noise while the current mask enforces the peak-current constraint throughout. A deployment analysis shows that the 18,432 multiply–accumulate policy meets a $50\mathsf{\mu }$s control period in its existing form at a measured cost of about $2\%$ in torque ripple. All the results are simulation-based on a finite-element model parameterized from a physical machine. Full article

This paper develops a new multi-stage image distillation method that combines two well-known techniques. In the first stage, our method creates a matrix from all training images. In the next stage, it adapts a modified principal component analysis (M-PCA) approach to transform the training matrix. In the third stage, Singular Value Decomposition (SVD) further refines the training-image matrix through low-rank reconstruction and controlled row selection. In the fourth stage, rotation of small $2\times 2$ matrix blocks on the entire left singular matrix is conducted. The upper m (user-selected number) rows of the reconstructed matrix are selected and transformed back to images, which we call distilled images. This dataset is significantly smaller yet retains the critical information needed for accurate classification. We validated the novelty and the advantages of the new method by applying the Baseline, ResNet50V2, and ConvNetD4 CNNs and the public image databases Digit-MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, BloodMNIST, and Tiny ImageNet. Experimental results show that by distilling 50 images per class from CIFAR-10, CIFAR-100, and Tiny ImageNet, the proposed method achieves superior test accuracies of 75.59%, 56.27%, and 31.75%, respectively, when evaluated on ResNet and ConvNetD4. Full article

No-reference image quality assessment (NR-IQA) metrics are widely used to evaluate and train image and video processing algorithms, but the growing reliance on deep neural networks makes these metrics vulnerable to adversarial attacks. Among such attacks, transferable black-box methods are particularly relevant in real-world scenarios where the attacker cannot interact with the target metric to generate adversarial examples and must rely on a substitute white-box model. In this paper, we examine adversarial transferability across modern NR-IQA metrics and introduce two complementary techniques, Consensus-Guided Distillation (CGD) and Local Perceptual Attack (LPA), which together form a two-stage transferable attack pipeline. CGD distills an ensemble of NR-IQA metrics into a single white-box substitute, excluding training samples for which the teacher metrics show high disagreement, thereby improving transferability and reducing the computational cost of both training and attack generation. LPA optimizes adversarial examples over random multi-scale image partitions, encouraging perturbations to exploit local distortion-sensitive features shared across NR-IQA metrics. Experiments using 15 NR-IQA metrics and four datasets show that distillation substantially improves black-box transferability, while LPA consistently outperforms existing transferable attacks against NR-IQA metrics, both in increasing predicted quality scores and in reducing correlations with subjective quality scores. We also find that attack transferability depends strongly on the type of distortions present in the image, indicating that adversarial perturbations can conceal certain visual degradations from NR-IQA metrics. Full article

Two-stage oriented detectors achieve high localization accuracy in synthetic aperture radar (SAR) ship detection, but their large backbones, feature pyramids, proposal modules, and heavy region of interest (RoI) heads hinder deployment. Existing lightweight SAR ship detectors typically use one-stage frameworks that lack proposal-level refinement for precise rotated localization. This paper presents a dual-teacher knowledge distillation and pruning (DTKDP) framework for lightweight oriented SAR ship detection. DTKDP introduces learnable gates into convolutional, normalization, and linear layers to prune convolutional channels and RoI-head neurons. Rotated proposal alignment (RPA) distills teacher and student predictions in a shared teacher-generated rotated proposal space, while a dual-teacher scheme combines classification and regression guidance from a homogeneous main teacher with complementary classification cues from a heterogeneous auxiliary teacher. Experiments on the SAR Ship Detection Dataset (SSDD) and Rotated Ship Detection Dataset in SAR Images (RSDD-SAR) show that DTKDP reduces the parameters of Oriented Region-based Convolutional Neural Network (Oriented R-CNN) and RoI Transformer equipped with ResNet-50 backbones by 87.5–91.8% and their floating-point operations (FLOPs) by 75.6–79.9%. In terms of average precision (AP) and mean average precision (mAP), the resulting Oriented R-CNN-slim and RoI Transformer-slim retain accuracy close to their full-scale counterparts. Relative changes across AP₅₀, AP₇₅, mAP_50:75, and mAP_50:95 range from a 2.38% decrease to a 0.65% improvement. Compared with RTMDet-tiny, they improve all four metrics on both datasets by 0.52–27.55% and consistently surpass representative distillation methods, demonstrating a favorable accuracy–efficiency trade-off. Full article

Near-duplicate diagram detection in electronic documents is challenging because diagram identity depends on graphical structure, spatial composition, and textual labels, while reused images may undergo compression, cropping, rotation, photometric changes, or perspective distortion. This study proposes a cascaded multimodal framework combining perceptual hashing, Siamese Residual Network with 18 layers (Siamese ResNet18), Distillation with No Labels, ver. 2 (DINOv2) visual representations, and a text-similarity classifier. A controlled benchmark was constructed from Artificial Intelligence 2D Diagram Dataset (AI2D) using Light, Medium, and Hard transformations, with source-grouped splitting by base_id to prevent leakage across training, validation, and test sets. Perceptual hashing achieved Area Under the Receiver Operating Characteristic Curve (ROC-AUC) = 0.7483, while Siamese ResNet18 increased ROC-AUC to 0.8599. DINOv2 provided the strongest visual performance, achieving Accuracy = 0.9933, F1-score = 0.9933, ROC-AUC = 0.9992, and Average Precision = 0.9994; F1-score remained 0.9901 for Hard transformations. Visual fusion increased ROC-AUC to 0.9995, and full multimodal fusion reached ROC-AUC = 0.9998. At an early-exit threshold of 0.95, 27.6% of pairs were resolved at the hashing level. These results support the coarse-to-fine design on the constructed AI2D-derived benchmark. A targeted hard-negative stress test revealed substantially higher false-positive rates under deliberately matched spatial layouts, with an overall False Positive Rate (FPR) of 0.48 for DINOv2 and 0.16 for full multimodal fusion. Generalization to naturally reused or redrawn diagrams, larger and more diverse hard-negative collections, and Optical Character Recognition (OCR)-derived text remains to be evaluated. Full article

Non-contact underwater fish counting supports stocking-density estimation and management in intelligent aquaculture. This study evaluated a 900-image annotated TV5/IOC in-domain dataset derived from the right-view stream of one source video. For each of three seed-controlled splits, 720 images were used for training and 180 for validation. All controlled configurations used the same raw single-view (Raw-SV) supervised and inference input. High-density-aware density modeling (HDA-DM) applies a mild sample weight only to the density loss of high-count training images. HDA-DM reduced the Main900 three-seed mean MAE from 4.024 ± 0.578 to 3.739 ± 0.414 (7.07%) and reduced the Hard200 challenging-case mean MAE from 9.799 ± 1.792 to 8.863 ± 0.304 (9.56%). The improvement was not identical in every seed. Cross-view density distillation (CVDD), evaluated as a training-only ablation, did not further improve HDA-DM: the combined configuration obtained 3.987 ± 0.635 on Main900 and 9.666 ± 1.373 on Hard200. A loss-scope audit further showed that the original density-only HDA-DM formulation outperformed weighting density, count, and shape supervision together. HDA-DM introduces no additional inference-stage parameters or operations. Because the evidence is limited to three seeds and one acquisition source, all comparisons are descriptive and no formal statistical-significance or external-generalization claim is made. Full article

Compact deep models are attractive for medical image classification, but conventional knowledge distillation mainly transfers class-level predictions and may not adequately preserve lesion-relevant spatial cues. To address this limitation, we propose LaHKD, a lesion-aware hierarchical knowledge distillation framework for parameter-efficient medical image classification. LaHKD enables a lightweight student to learn hierarchical semantic representations together with lesion-focused guidance from a stronger teacher. We evaluate LaHKD on HAM10000 dermoscopic lesion classification and a brain tumor MRI classification benchmark. Across both datasets, LaHKD improves compact-student classification performance, with the clearest lesion-focused spatial benefits observed on HAM10000, where lesion morphology is central to diagnosis and direct lesion supervision is available. On the magnetic resonance imaging (MRI) benchmark, localization analysis is limited to an auxiliary recovered-mask subset and is therefore interpreted as exploratory; under this setting, consistent localization gains are not observed. Overall, LaHKD provides an effective framework for compact medical image classification, with spatial benefits most clearly supported in tasks with reliable lesion supervision. Full article

Hyperspectral image (HSI) classification is critical for remote sensing but faces challenges in balancing accuracy and inference efficiency. Existing graph-based knowledge distillation (KD) methods are confined to single geometric spaces, ignoring the hierarchical semantics of land-cover categories. Cross-geometry distillation compresses multiple geometries into a single Euclidean student, forcing one geometry to collapse into the other. In this paper, we propose Dual-Space Knowledge Distillation (DSKD), a novel dual-student dual-space KD framework integrating Euclidean (GCN) and Hyperbolic (HGCN) teachers to jointly train a native MLP student and a native HNN student. With cross-geometric feature bridging (CGFB) and output distribution cohesion (ODC), the two students mutually learn each other’s complementary geometry, so DSKD captures complementary spatial-spectral and hierarchical features while enabling graph-free inference without message passing. Extensive experiments on four HSI datasets and four general graph benchmarks demonstrate that DSKD outperforms single-space distillation and single-student cross-geometry baselines in most settings, confirming its effectiveness and generalization capability across diverse graph-structured data. Full article

Deep neural networks have achieved strong performance in medical image classification, but their deployment may be constrained by the computational cost of high-capacity models. Knowledge distillation (KD) addresses this problem by transferring knowledge from a teacher to a lightweight student. However, the reliability of teacher supervision may vary across samples, potentially introducing noisy guidance and local conflicts with ground-truth supervision. We propose Selective Confidence-guided Projection-based Encoding (SCOPE), a conflict-aware KD framework comprising Selective Relation Alignment (SRA) and Gradient Conflict Resolution (GCR). SRA constructs reliability-aware relational supervision by combining teacher-derived relations with dataset-specific auxiliary priors, whereas GCR removes distillation-gradient components that conflict with the classification objective. Experiments on nine medical image datasets and multiple teacher–student architectures demonstrate competitive predictive performance, improved training stability, and low computational overhead. Full article

Accurate detection of blueberries at different maturity stages supports orchard monitoring and harvest planning, but remains difficult under complex field conditions because fruits are small, densely clustered, frequently occluded, and show subtle colour transitions across immature, colour-turning, and ripe stages. A key technical challenge for lightweight detectors is to preserve fine spatial information for small fruits while maintaining sufficient contextual representation for dense clusters and visually ambiguous maturity stages without substantially increasing inference complexity. To address this gap, this study develops a lightweight detection framework that jointly targets small-object perception, contextual feature enhancement, and training-stage semantic supervision. A field dataset containing 4680 images and 60,573 ground-truth bounding boxes was constructed. The P3/P4/P5 detection structure of YOLOv12n was redesigned as a P2/P3/P4 architecture to strengthen high-resolution feature representation for small objects. A lightweight feature-enhancement module based on single-head self-attention, termed SHSA2C2f, was introduced to improve contextual representation in dense fruit clusters and visually ambiguous maturity stages. DINOv3 (ViT-S/16)-guided asymmetric knowledge distillation was further evaluated as a training-only semantic supervision strategy that introduces no additional inference cost. Across three independent runs, the accuracy-oriented M4 model achieved 92.68 ± 0.04% mAP@0.5 and 86.38 ± 0.10% mAP@0.5:0.95 with 0.79 M parameters, while its colour-turning AP50 reached 89.78 ± 0.28%. For the same student architecture, introducing the asymmetric distillation strategy increased recall from 85.00 ± 0.52% to 85.81 ± 1.12%, while slightly reducing mAP@0.5:0.95 from 86.38 ± 0.10% to 86.19 ± 0.17%, indicating a recall-oriented performance trade-off. After TensorRT FP16 conversion, M4 retained 92.75% mAP@0.5 and 86.24% mAP@0.5:0.95 and ran at 15.9 FPS on a Jetson Orin Nano, supporting its feasibility for edge-device inference under the evaluated conditions. Full article

A Healthcare Recommender System (HRS) is a personalized decision-support system designed to recommend healthcare-related services, providers, information, advice, diagnoses, treatments, or lifestyle tips to users based on users’ preferences, characteristics, or individualized health data. Healthcare recommendation systems rely on multiple data sources. Among these sources, user reviews and comments on online healthcare platforms and social media serve as a valuable source to provide direct and often experience-based information about users’ satisfaction and opinions regarding healthcare services. Therefore, this study focuses on exploiting sentiment information extracted from user-generated reviews to enhance healthcare service recommendations. We propose a novel framework that integrates sentiment analysis on the Yelp dataset using DistilBERT, a lightweight transformer-based language model. The proposed framework incorporates Neural Collaborative Filtering (NCF) for the recommendation process. It utilizes Singular Value Decomposition (SVD) to address sparsity issues in user–item interaction data, thereby maintaining reliable performance even with limited data availability. The proposed approach achieved a Mean Absolute Error (MAE) of 0.34, a Root Mean Square Error (RMSE) of 0.66, and an Area Under the Curve (AUC) of 0.92. It also demonstrated strong ranking performance, achieving a Recall@10 of 0.74, demonstrating its effectiveness and accuracy in recommendation tasks. Compared with the rating-based, without-SVD, and without-NCF variants, which achieved RMSE values of 0.97, 1.80, and 1.14, respectively, the proposed model consistently achieved better performance. These results highlight the contribution of sentiment analysis, SVD-based interaction augmentation, and NCF to the recommendation performance. These results are promising and confirm the potential of the proposed approach for improving the reliability and performance of recommender systems. Full article

Bottom detection and tracking from multibeam echosounder water-column data constitute a fundamental step to 3D seabed mapping. Traditional methods based on amplitude thresholding and phase detection rely heavily on heuristic assumptions and are prone to failure in the presence of interference in water-column data. To address these limitations, we propose a novel deep convolutional architecture that models full-swath spatial context, and couple it with a parallelized sliding-window pipeline for end-to-end inference on high-resolution backscatter signals. The proposed model was trained and evaluated on a combined dataset of public deep-water multibeam water-column data acquired by Kongsberg EM302 and EM710 systems. The bottom detection results demonstrate that in deep-water scenarios, our method obtained the correct bottom position, whereas the traditional methods yielded inaccurate or no detection results. Furthermore, applying knowledge distillation yields a lightweight model that achieves a high inference speed on an embedded edge device. The proposed method substantially suppresses interference-induced false detections and provides a robust, real-time solution for marine bathymetry. Full article

Remaining useful life (RUL) prediction of wind-turbine bearings is challenged by nonstationary wind loads, multistage degradation, substantial lifetime dispersion, and strict deployment constraints. Conventional single-task regressors apply a unified feature-to-RUL mapping over the entire life cycle and therefore struggle to characterize stage transitions and bearing-specific degradation progress. To address these issues, this paper proposes a stage-aware multi-task RUL prediction method with causal degradation-prior fusion and collaborative distillation. During training, a high-capacity reference representation path transfers inter-sample relational structures and task-level degradation knowledge to a compact feature encoding path, while only the compact path is retained for inference. Based on the compact representation, a multi-task module jointly performs four-stage classification, stage-conditioned RUL regression, and continuous remaining-life estimation; predicted stage probabilities softly fuse the stage-conditioned outputs. A causal prior-fusion module further integrates a bearing-specific healthy-state anchor, causally identified first prediction time, cumulative damage, and the stage-aware prediction to adapt the RUL trajectory to individual degradation processes. Experiments on the IEEE PHM 2012 and XJTU-SY datasets demonstrate that the proposed method provides accurate and robust RUL prediction across different bearing degradation processes. Moreover, the compact inference path maintains efficient implementation, supporting its potential use in practical wind-turbine condition-monitoring applications. Full article

The integration of Tiny Machine Learning (TinyML) into human activity recognition (HAR) represents a paradigm shift in artificial intelligence, enabling real-time, efficient, and privacy-preserving analysis on resource-constrained edge devices. This paper presents a comprehensive review of TinyML for HAR, covering foundational concepts, methodologies, and applications. the review examines the state of the art of existing works and approaches that combine TinyML and HAR, providing a detailed review and comparison of models, algorithms, and frameworks. This comparison sheds light on the effectiveness and limitations of different methodologies. Key contributions include a systematic taxonomy of HAR systems leveraging TinyML, a detailed analysis of optimization techniques like pruning, quantization, and knowledge distillation, and insights into state-of-the-art frameworks and datasets. Challenges such as scalability, energy efficiency, and generalization to diverse environments are critically examined, alongside solutions like federated learning, multimodal data fusion, the integration of generative AI, neuromorphic hardware, 5G/6G and Internet Of Things (IoT) connectivity, which are highlighted as transformative enablers for advancing HAR applications. This review serves as a foundational resource for researchers and practitioners aiming to harness TinyML’s potential in activity recognition systems. Full article