Adaptive Quantization with Mixed-Precision Based on Low-Cost Proxy

Concrete surface crack detection plays a crucial role in infrastructure maintenance and safety. Deep learning-based methods have shown great potential in this task. However, under real-world conditions such as poor image quality, environmental interference, and complex crack patterns, existing models still face challenges in detecting fine cracks and often rely on large training parameters, limiting their practicality in complex environments. To address these issues, this paper proposes a crack detection model based on adaptive feature quantization, which primarily consists of a maximum soft pooling module, an adaptive crack feature quantization module, and a trainable crack post-processing module. Specifically, the maximum soft pooling module improves the continuity and integrity of detected cracks. The adaptive crack feature quantization module enhances the contrast between cracks and background features and strengthens the model’s focus on critical regions through spatial feature fusion. The trainable crack post-processing module incorporates edge-guided post-processing algorithms to correct false predictions and refine segmentation results. Experiments conducted on the Crack500 Road Crack Dataset show that, the proposed model achieves notable improvements in detection accuracy and efficiency, with an average F1-score improvement of 2.81% and a precision gain of 2.20% over the baseline methods. In addition, the model significantly reduces computational cost, achieving a 78.5–88.7% reduction in parameter size and up to 96.8% improvement in inference speed, making it more efficient and deployable for real-world crack detection applications.

Photonic neural networks have emerged as a promising solution to overcome limitations of traditional hardware for neuromorphic computations, offering advantages in bandwidth, latency, and power efficiency. However, their performance is constrained by the limited precision of analog photonic computing, which is affected by inherent noise sources such as thermal and shot noise, and distortions. These effects degrade the photonic neural network performance, reducing the bit resolution achievable in photonic hardware typically to 2–4 bits. Traditional quantization strategies fail to account for these noise contributions, resulting in a substantial accuracy loss during inference. This paper introduces an adaptive quantization method called Adaptive-Quantization Photonic-Aware Neural Network (AQ-PANN) to address the challenges posed by different noise sources in analog photonic hardware. The proposed method uses a learnable step size quantization scheme to achieve high accuracy and stability under varying noise levels, introducing a scheme that unifies quantization step adaptation with noise injection exactly where photonic distortions arise. This design incurs only a minor training-time overhead, as it involves learning a small number of per-layer quantization step sizes and does not affect inference. Experimental evaluations on three commonly used test datasets (MNIST, SVHN, and CIFAR-10) with different bit resolutions show the robustness of AQ-PANN. On MNIST, an accuracy drop of only 2% was observed from low to high noise levels in a 4-bit configuration, while traditional DoReFa quantization suffered a 29% drop. For the SVHN dataset, AQ-PANN obtained a mean accuracy of 92% under high noise with 4-bit quantization, outperforming DoReFa by over 45%. On CIFAR-10, AQ-PANN maintained close to 60% accuracy under high noise in the 4-bit configuration, whereas DoReFa and PACT both collapsed below 40%. These results highlight the effectiveness of AQ-PANN in sustaining model performance across different noise intensities, enabling practical photonic neural network deployment.

In this paper, we proposed a novel watermarking scheme based on adaptive quantization index modulation and singular value decomposition in the hybrid discrete wavelet transform (DWT) and discrete cosine transform (DCT). The secret watermark bits are embedded on the singular values vector of blocks within low frequency subband in host image hybrid DWT-DCT domain. To embed watermark imperceptibly, robustly and securely, we model the adaptive quantization steps by utilizing human visual system (HVS) characteristics and particle swarm optimization (PSO) algorithm. Experimental results demonstrate that the proposed scheme is robust to variety of image processing attacks. In the proposed algorithm the quantized embedding strategy is adopted, so no host image is needed for blind extraction of watermarking image.

An investigation is conducted of the effects of various source statistics and buffer management algorithms for a buffer of fixed size, and an algorithm that provides particularly robust rate-distortion performance for a wide range of stationary sources is described. The performance of this scheme, which is referred to as the adaptive entropy-coded quantizer (AECQ), is compared to that of conventional adaptive pulse code modulation (APCM) operating at the same fixed channel transmission rate. Experimental results indicate that the AECQ can provide considerably improved performance and more robust operation than APCM. The relative sensitivity of the AECQ to scale mismatch is directly related to the rather simple memoryless feedback control strategy employed. This might be remedied by alternative choices of the memoryless feedback control function or by introducing memory. The former is much easier to implement and/or analyze than the latter. >

The development of an adaptive vector quantizer (AVQ) codec chip using Open Architecture Silicon Implementation System (OASIS) tools is described. The goal is development of a 16-kb/s AVQ coder prototype system which can be used as an alternative to the currently used 32-kb/s adaptive delta pulse code modulation (ADPCM). The AVQ semicustom chip layout uses 1.2- mu m CMOS technology, and consists of 157,191 transistors. The AVQ system is being designed with Mentor Graphics CAD tools and is a four-layer PC board encompassing over 60 analog and digital components.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Case law plays a critical role in shaping our understanding of human rights, including the right to adequate housing. However, analyzing large legal databases like HUDOC, which contains over 40,000 cases, is a challenging task that requires automated solutions. This study focuses on detecting cases related to housing—a topic encompassing issues such as eviction, access to adequate housing and etc.—from the HUDOC database. For this, we developed classifiers to identify cases related to both housing and eviction issues. We first constructed a dataset using an unsupervised process refined through manual corrections. Then, we trained the Adaptive Chordal Distance-based Subspace Learning Vector Quantization models. These models achieved classification accuracies of 93% for housing-related cases and 91.5% for eviction-specific cases, matching the performance of transformer-based models while requiring fewer computational resources. Furthermore, they provide interpretability by assigning word-level importance scores, helping legal scholars understand and verify the reasoning behind the model’s predictions. The models identified 2,305 potentially housing-related cases. Manual reviews confirmed that 278 of 340 reviewed cases were indeed relevant. By detecting overlooked cases and enriching legal datasets, this study highlights the utility of NLP methods in facilitating the analysis of human rights case law. This approach supports a deeper exploration of housing rights and eviction-related decisions under the European Court of Human Rights (ECtHR), offering transparency, efficiency, and scalability for legal research.

The Independent Curriculum provides thinking in learning, allowing teachers to adjust the material to the needs of students. However, challenges such as limited resources, teacher readiness, and variations in student learning styles are still challenges in its implementation (Kemendikbud, 2022). This study aims to examine the implementation of adaptive learning modules as a solution to improve learning effectiveness. Adaptive modules that integrate technology and personalization of teaching materials are expected to improve students' understanding in a way that is more appropriate to their individual needs (Mayer, 2009). The approach used in this study includes literature studies and field observations, which show that adaptive modules can significantly improve student learning outcomes. However, the implementation of this adaptive learning module still requires support in the form of adequate teacher training and supporting technological infrastructure so that learning can run effectively and efficiently (Batubara, 2019). This study provides recommendations to strengthen teacher readiness and provide the resources needed in the implementation of adaptive modules.

Aim Vascular anastomosis is an important but complex skill requiring mastery by trainees across multiple specialities, with significant risk of suboptimal outcomes. Simulation allows multiple repetitions of a procedure within controlled environment, allowing training without accompanying patient risk. Current simulation courses on vascular anastomosis are expensive and inaccessible to several junior trainees. A low-cost scaffold was developed for vascular anastomosis training and aimed to assess trainee feedback on utility and acceptability. Method The kit was built using following: Six 27 x 0.9 cm latex balloons, Ferrero Rocher chocolate box, Polypropylene 6.0 surgical sutures, Castroviejo needle holder and Forceps. Sharp scissors were used to make openings in the chocolate box, the balloons were then passed through. The kit’s utility was assessed via trainee feedback (N=16, ranked 1-5 where 1 was Unsatisfactory and 5 was Excellent). Results Trainees were able to perform various types of anastomoses: end-to-end, end-to-side, and patching, using our model. The resistance to needle passage through latex balloons was found to closely resemble that of real arteries and veins. Median rating of different categories in trainee feedback were as followed: Improved understanding 4, Helped practice 5, Realistic 4, Appropriateness for learning 4 and Affordability 5, ranging between 3 and 5. Conclusions Improved skills lead to better outcomes in surgical practice. Steps should be taken to make simulation more accessible. Low-cost models help reduce financial impact on trainees and are suitable for a wide range of institutions, including those with limited resources. Subsequent research should assess the effectiveness of this model across various contexts.

The second order scattering information provided by the Hessian matrix and its inverse plays an important role in both, parametric inversion and uncertainty quantification. On the one hand, for parameter inversion, the Hessian guides the descent direction such that the cost function minimum is reached with less iterations. On the other hand, it provides a posteriori information of the probability distribution of the parameters obtained after full waveform inversion, as a function of the a priori probability distribution information. Nevertheless, the computational cost of the Hessian matrix represents the main obstacle in the state-of-the-art for practical use of this matrix from synthetic or real data. The second order adjoint state theory provides a strategy to compute the exact Hessian matrix, reducing its computational cost, because every column of the matrix can be obtained by performing two forward and two backward propagations. In this paper, we first describe an approach to compute the exact Hessian matrix for the acoustic wave equation with constant density. We then provide an analysis of the use of the Hessian matrix for uncertainty quantification of the full waveform inversion of the velocity model for a synthetic example, using the 2D acoustic and isotropic wave equation operator in time.

This paper addresses the problem of routing and wavelength assignment (RWA) in multifiber WDM networks with limited resources. Given a traffic matrix, the number of fibers per link, and the number of wavelengths a fiber can support, we seek to maximize the carried traffic of connections. We formulate the problem as an integer linear program (ILP), and show that the lightpaths selected by this formulation can indeed be established by properly configuring the optical switches. An upper bound on the carried traffic can be computed by solving the linear programming (LP)-relaxation of the ILP formulation. It is shown that this bound can be also computed exactly, and in polynomial-time, by solving a significantly simplified LP which considers only one wavelength. The bound can, thus, easily scale to an arbitrarily large number of wavelengths. Furthermore, we demonstrate that any instance of the RWA problem is also an instance of the more general maximum coverage problem. This allows us to take a greedy algorithm for maximum coverage and obtain an algorithm which provides solutions for the RWA problem that are guaranteed to be within a factor of (1-(1/e)) of the optimal solution. Each iteration of the greedy algorithm selects a set of lightpaths that realizes, using one wavelength, the maximum number of connection requests not previously realized. Computational results confirm the high efficiency of our proposed algorithm.

The evolution of the Internet has transformed the simplistic Ethernet/IP based packet forwarding into a complex collection of lookup schemes. Depending on the location of the networking equipment (core, provider/customer edge, etc.,) a router/switch may potentially have to support several such complex lookup schemes. However, the hardware resources allocated to perform such operations are limited, especially in single chip implementations. In this paper, we propose techniques to map such complex lookup schemes on to hardware platforms under a limited resource budget and produce a design for a pipelined packet lookup architecture. An Integer Linear Programming (ILP) based technique is introduced to optimally allocate the limited hardware resources for a single lookup scheme. We extend our solution to multiple lookup schemes by proposing techniques to improve resource sharing, which results in a resource planning tool. Field Programmable Gate Array (FPGA) - a natural choice for high-speed packet processing applications - is used as the target platform. By using the proposed techniques, we show that up to 4 complex lookup schemes can be hosted on a single FPGA consuming only 20 Mbit on-chip memory and 750 pins for external memory communication.

Background modeling techniques for embedded computer vision applications must balance accuracy, speed, and power. Basic background modeling techniques run quickly, but their accuracy is not sufficient for computer vision problems involving dynamic background. In contrast, adaptive background modeling techniques are more robust, but run more slowly. Due to its high inherent fine-grain parallelism, robust adaptive background modeling has been implemented on GPUs with significant performance improvements over CPUs. However, these implementations are infeasible in embedded applications due to the high power ratings of the targeted general-purpose GPU platforms. This paper focuses on exploiting fine-grain data parallelism and optimizing memory access patterns to target a low-cost adaptive background modeling algorithm multimodal mean (MMM) to a low-power GPU with thermal design power (TDP) of only 12 watts. The algorithm has comparable accuracy with the Gaussian mixture model (GMM) algorithm, but less computational and memory cost. It achieves a frame rate of 392 fps with a full VGA resolution (640x480) frame on the low-power integrated GPU NVIDIA ION. This is a 20x speed-up of the MMM algorithm compared to the embedded CPU platform Intel Atom of comparable TDP. In addition, the MMM algorithm attains a 5-6x speed up over the GMM implementation on the ION GPU platform.

Purpose Graph neural networks (GNNs) have been widely applied to various recommendation scenarios due to their outstanding performance in capturing and modelling hidden relationships. However, as the performance of a GNN heavily relies on the graph structure, the structure bias can lead to significant unfairness problems in a GNN-based recommender system across different user groups. Furthermore, prior studies have shown that sensitive attributes of nodes can be revealed through the graph structure. In other words, even if a GNN-based recommendation system does not explicitly disclose a user’s sensitive information, it can still be inferred from the imbalance in users’ connections within the graph. This article therefore addresses the unfairness arising from such structural biases by proposing a novel model designed to mitigate these biases in GNN-based recommendation. Here, the fairness requirements refer to preventing the exposure of sensitive attributes during user modeling and avoiding biased recommendations across different demographic groups. Design/methodology/approach To address this issue, we propose a novel fairness-aware graph neural network model, EqStruct-GNN, which aims to mitigate inter-group unfairness at the structural level. The model is composed of three key modules: (1) the Adaptive Behavior Modeling Module, which computes fairness-aware node behavior probability distributions; (2) the Fairness-Aware Structure Construction Module, which builds graph structures based on these distributions and (3) the Fair Representation Learning Module, which generates unbiased node embeddings. Together, these modules enable EqStruct-GNN to construct a fairness-aware graph that reduces structural bias and improves fairness across sensitive groups. Findings Experimental results demonstrate that the proposed method achieves strong recommendation performance while ensuring fairness across multiple datasets. Furthermore, when applied to various backbone models, EqStruct-GNN not only significantly improves fairness metrics but also maintains or even enhances recommendation accuracy. These results highlight the generalization capability and practical effectiveness of the proposed framework. Originality/value Unlike existing methods that focus on representation-level fairness, our model tackles structural bias at its root by building a fairness-aware debiased graph through a novel graph construction mechanism. This structure enables the learning of fair and high-quality node representations by effectively mitigating structural bias while preserving critical interaction information. In addition, the proposed approach is universally adaptable and can be seamlessly applied to various GNN architectures, demonstrating strong generality and practical applicability.

The vision system provides to a humanoid robot the most complete information about the environment, having thus visual information extraction a great influence in the decision process of interaction with the environment. It represents one of the most resources consumer in the system, so it has to be efficiently designed without compromising the reliability of the provided information. Software implementation of an accurate vision system is one of the more distinctive elements in RoboCup competitions, especially in the RoboCup, where normative is becoming more and more restrictive and no hardware modifications can be made. In embedded control systems, management of the available resources is crucial to achieve a good performance and get the most from them. Focusing on the Robocup SPL, Nao robot has the CPU as the main shared and limited resource and, to a lesser degree, wireless network and memory should also be considered. In this work, the design of the vision module and the ability of the Hidalgos Team architecture to cope with limited resources are presented. The goal is to maximise CPU utilisation by boosting the capacities of the vision system, while maintaining control properties such as reactivity and stability. A first strategy emphasises on reducing resources consumed by the vision system through adaptability to control requirements. A second strategy focuses on adapting the execution rates of the modules, and especially of the vision system, to the available resources.

The proliferation in edge computing in consumer electronics has intensified the demand for efficient, privacy-preserving fine-tuning of large language models (LLMs) amid resource heterogeneity, and data regulations. Traditional federated and split learning methods are limited by central coordination, static partitioning, and high overhead. This paper introduces SplitGossip, a novel peer-to-peer (P2P) framework for fine-tuning distilled LLMs on consumer edge devices. It combines reinforcement learning (RL)-based adaptive model splitting, layer-wise Low-Rank Adaptation (LoRA), and compressed gossip synchronization with error-feedback and local differential privacy (DP). SplitGossip employs an RL agent to dynamically partition layers through a multiobjective reward, balancing energy, latency, and accuracy. Local training uses LoRA with gradients compressed through adaptive top-k sparsification and INT8 quantization. Experiments conducted on AWS-simulated edge clusters show that the proposed SplitGossip outperforms baseline methods across all metrics, i.e., F1-score, communication overhead, energy efficiency, and convergence rate.