检索结果-内蒙古大学图书馆

reparameterization convolutional neural networks for handling imbalanced datasets in solar panel fault classification

引用

ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE 2025年 150卷

作者： Guo, Jielong Chong, Chak Fong Abreu, Pedro Henriques Mao, Chao Li, Jiaxuan Lam, Chan-Tong Ng, Benjamin K. Macao Polytech Univ Fac Appl Sci R Luis Gonzaga Gomes Taipa Macao Peoples R China Univ Coimbra Ctr Informat & Syst Dept Informat Engn P-3030290 Coimbra Portugal

Solar photovoltaic technology has grown significantly as a renewable energy, with unmanned aerial vehicles equipped with thermal infrared cameras effectively inspecting solar panels. However, long-distance capture and low-resolution infrared cameras make the targets small, complicating feature extraction. Additionally, the large number of normal photovoltaic modules results in a significant imbalance in the dataset. Furthermore, limited computing resources on unmanned aerial vehicles further challenge real-time fault classification. These factors limit the performance of current fault classification systems for solar panels. The multi-scale and multi-branch reparameterization of convolutional neural networks can improve model performance while reducing computational demands at the deployment stage, making them suitable for practical applications. This study proposes an efficient framework based on reparameterization for infrared solar panel fault classification. We propose a Proportional Balanced Weight asymmetric loss function to address the class imbalance and employ multi-branch, multi-scale convolutional kernels for extracting tiny features from low-resolution images. The designed models were trained with Exponential Moving Average for better performance and reparameterized for efficient deployment. We evaluated the designed models using the Infrared Solar Module dataset. The proposed framework achieved an accuracy of 83.8% for the 12-Class classification task and 74.0% for the 11-Class task, both without data augmentation to enhance generalization. The accuracy improvements of up to 16.4% and F1-Score gains of up to 18.7%. Additionally, we achieved an inference speed that is 3.4 times faster than the training speed, while maintaining high fault classification performance.

关键词： reparameterization Convolutional neural network Solar energy Fault classification Photovoltaic modules

来源：评论

学校读者我要写书评

暂无评论

reparameterization-Based Parameter-Efficient Fine-Tuning Methods for Large Language Models: A Systematic Survey 13th

Reparameterization-Based Parameter-Efficient Fine-Tuning Met...

引用

13th International Conference on Natural Language Processing and Chinese Computing

作者： Chen, Zezhou Liu, Zhaoxiang Wang, Kai Lian, Shiguo China Unicom Ai Innovat Ctr Beijing 100013 Peoples R China China Unicom Unicom Digital Technol Beijing 100013 Peoples R China

ISBN: (纸本)9789819794362;9789819794379

The rapid advancement of Large Language Models (LLMs) has revolutionized both academia and industry, leveraging Transformer architectures and pre-training objectives to achieve unprecedented performance. To fully exploit the potential of LLMs, fine-tuning LLMs on specific downstream tasks is essential. However, traditional full fine-tuning methods pose significant computational challenges, prompting the emergence of Parameter-Efficient Fine-Tuning (PEFT) methods, especially reparameterization-based PEFT methods. In this survey, we delve into reparameterization-based PEFT methods, which aim to fine-tune LLMs with reduced computational costs while preserving their knowledge. We systematically analyze their design principles and divide these methods into six categories. We analyze the training parameter complexity, GPU memory consumption, training time costs, accuracy and limitations of each method. We summarize challenges within the reparameterization-based PEFT methods and propose future directions.

关键词： Large Language Models Parameter-Efficient Fine-Tuning reparameterization

来源：评论

学校读者我要写书评

暂无评论

L3Net: Localized and Layered reparameterization for incremental learning

引用

NEURAL NETWORKS 2025年 188卷 107420页

作者： Luo, Xuandi Zhang, Huaidong Xie, Yi Zhang, Hongrui Xu, Xuemiao He, Shengfeng South China Univ Technol Sch Future Technol Guangzhou Peoples R China South China Univ Technol Guangdong Engn Ctr Large Model & GenAI Technol Sch Comp Sci & Engn Guangzhou Peoples R China Singapore Management Univ Sch Comp & Informat Syst Singapore Singapore

Model-based class incremental learning (CIL) methods aim to address the challenge of catastrophic forgetting by retaining certain parameters and expanding the model architecture. However, retaining too many parameters can lead to an overly complex model, increasing inference overhead. Additionally, compressing these parameters to reduce the model size can result in performance degradation. To tackle these challenges, we propose a novel three-stage CIL framework called Localized and Layered reparameterization for Incremental Learning (L3Net). The rationale behind our approach is to balance model complexity and performance by selectively expanding and optimizing critical components. Specifically, the framework introduces a Localized Dual-path Expansion structure, which allows the model to learn simultaneously from both old and new features by integrating a fusion selector after each convolutional layer. To further minimize potential conflicts between old and new features, we implement the Feature Selectors Gradient Resetting method, which sparsifies the fusion selectors and reduces the influence of redundant old features. Additionally, to address classification bias resulting from class imbalance, we design the Decoupled Balanced Distillation technique and apply Logit Adjustment to more effectively retain knowledge from the rehearsal set. Extensive experiments demonstrate that our L3Net framework outperforms state-of-the-art methods on widely used benchmarks, including CIFAR-100 and ImageNet-100/1000.

关键词： Class incremental learning Knowledge distillation reparameterization

来源：评论

学校读者我要写书评

暂无评论

Improved Lightweight YOLOv8 With DSConv and reparameterization for Continuous Casting Slab Detection on Embedded Device

引用

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT 2025年 74卷

作者： Ju, Hao Fang, Yiming Yang, Hongliang Si, Fengfei Kang, Kesong Yanshan Univ Key Lab Ind Comp Control Engn Qinhuangdao 066004 Hebei Peoples R China Yanshan Univ Key Lab Ind Comp Control Engn Educ Minist Intelligent Control Syst & Intelligent Qinhuangdao 066004 Hebei Peoples R China Yanshan Univ Engn Res Ctr Educ Minist Intelligent Control Syst & Intelligent Qinhuangdao 066004 Hebei Peoples R China Huitang IoT Technol Co Ltd Tangshan 063000 Hebei Peoples R China

The molten steel is transformed into slabs in the process of continuous casting. These slabs are subsequently conveyed to the hot rolling area for further processing or stored in the slab storage area for future utilization. An improved lightweight object detection algorithm called You Only Look Once version 8 (YOLOv8) with DSConv, light Efficient RepGFPN, and CARAFE (DSCLERC-YOLOv8) was proposed in this article for real-time detection of slabs transmitted on the roller table, to be applied in embedded devices. DSCLERC-YOLOv8 utilizes an enhanced C2f with distribution shifting convolution (C2f_DSConv) as the feature extraction module, which enhances the processing speed of the object detection algorithm while maintaining accuracy. Moreover, the algorithm incorporates an improved Lightweight Efficient-reparameterization generalized feature pyramid network (RepGFPN) and content-aware reassembly of features (CARAFE) for feature map upsampling, enhancing the accuracy and efficiency of image prediction. In addition, the article investigates the impact of reducing network stages on the algorithm. For practical applications, a P4 network structure is employed, resulting in a slab detection accuracy rate of 86.91% for mAP50:95. Although this accuracy is slightly lower than YOLOv8n (87.78%) by 0.87%, the significant improvement of this algorithm lies in a 14.7% reduction in network latency, a 77.22% reduction in parameter count, a 66.29% reduction in computational complexity, and a 62.90% reduction in model size. Finally, by implementing model pruning, quantization, and deploying with the TensorRT engine, the algorithm successfully achieves a real-time speed of 76 FPS on Jeston Orin NX. This algorithm offers a dependable solution for the real-time detection of steel slabs during continuous casting and rolling processes in the steel industry.

关键词： Slabs Feature extraction Convolution Accuracy Object detection Real-time systems Computational modeling Steel Transformers Head Distribution shifting convolution (DSConv) embedded devices lightweight You Only Look Once version (YOLOv8) object detection reparameterization slab detection

来源：评论

学校读者我要写书评

暂无评论

Prefix Tuning Using Residual reparameterization

引用

IEEE ACCESS 2025年 13卷 54866-54872页

作者： Jung, Youngjun Hwang, Hyunsun Lee, Changki Kangwon Natl Univ Interdisciplinary Grad Program Med Bigdata Converg Chunchon 24341 South Korea Kangwon Natl Univ Dept Comp Sci & Engn Chunchon 24341 South Korea

Fine-tuning large language models for specific tasks requires updating and storing all parameters, leading to significant computational and storage cost issues. To address these challenges, parameter-efficient learning such as prefix tuning has gained attention. However, prefix tuning can suffer from sensitivity to prefix length. This is considered to be due to different prefix tokens being required for each task. In this paper, we propose improving the robustness and performance of prefix tuning through residual reparameterization. We add residual connections to the prefix module, providing more flexibility to the model. Additionally, we propose a gate mechanism to assign weights to prefix tokens, allowing for focus on more important tokens. Our experiments on the GLUE benchmark and E2E dataset demonstrate that our methods lead to improved and stabilized performance across various prefix lengths. The residual connections enable faster convergence during training, while the gate mechanism helps balance prefix tokens and find more optimized parameters. Our approach shows particular effectiveness when combining residual connections with the gate mechanism, outperforming original prefix tuning, especially with longer prefix lengths, while remaining parameter-efficient. We also provide an analysis of gate weight trends during training, offering insights into how the model uses prefix tokens for different prefix lengths.

关键词： Tuning Logic gates Training Computational modeling Computer architecture Robustness Costs Correlation Natural language generation Large language models Gate mechanism parameter efficient prefix tuning reparameterization residual connection

来源：评论

学校读者我要写书评

暂无评论

Less is more: Efficient Scene Graph Generation with reparameterization

Less is more: Efficient Scene Graph Generation with reparame...

引用

2025 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2025

作者： Hong, Jonghwan Noh, Seonghyeok Ku, Bonhwa Ko, Hanseok Vision Information Processing Engineering Korea University Seoul Korea Republic of Electrical Engineering Korea University Seoul Korea Republic of

ISBN: (纸本)9798350368741

Scene Graph Generation (SGG) aims to identify objects and their relationships in visual scenes but faces two key challenges: high computational overhead, particularly for real-time applications, and the long-tailed distribution of predicates, which biases models toward frequent relationships. To address these challenges, we propose Reparams-SGG, a lightweight and efficient network architecture composed of multi-path residual blocks. This architecture reduces computational overhead by leveraging a reparameterization strategy that minimizes sequential and parallel processing, making it highly efficient during inference. Moreover, we introduce a dynamic focal loss that dynamically adjusts the temperature scale during training to focus learning on rare predicates, promoting progressively unbiased learning. Additionally, we propose a dynamic distribution loss, compensating for learning limitations solely from one-hot distributions under data imbalance conditions. We evaluate our method on the widely-used Visual Genome and the recent PSG dataset. Reparams-SGG achieves competitive performance with significantly fewer parameters than state-of-the-art models, demonstrating its efficiency and suitability for deployment in resource-constrained environments. © 2025 IEEE.

关键词： Efficient scene graph generation reparameterization Unbiased scene graph generation

来源：评论

学校读者我要写书评

暂无评论

reparameterization of static filtration model of aqueous-based drilling fluids for simultaneous estimation of compressible mudcake parameters

引用

POWDER TECHNOLOGY 2021年 386卷 120-135页

作者： Borges, Rodrigo F. O. Oechsler, Bruno F. Oliveira, Beatriz R. Andrade, Lucas D. R. Calcada, Luis A. Scheid, Claudia M. Calado, Veronica Univ Fed Rural Rio de Janeiro BR-465Km7 BR-23897000 Seropedica RJ Brazil Univ Fed Santa Catarina R Eng Agron Andrei Cristian Ferreira 508 BR-88040535 Florianopolis SC Brazil Univ Fed Rio de Janeiro Av Pedro Calmom 550Cidade Univ BR-2194190 Rio De Janeiro RJ Brazil

To mitigate the fluid invasion and the loss of circulation, drilling fluids are designed to provide a thin mudcake with low permeability. Among the many methods of determining the mudcake properties, the absence of proper procedures related to the estimation of mudcake parameters that also consider the experimental uncertainties is noticed. Besides, the non-Newtonian rheological behavior in filtration and its consideration in resistive force modeling is still lacking in the literature. This paper proposes a new methodology of parameter estimation using the particle swarm optimization technique with a modification of the classic model approaches considering the non-Newtonian fluid filtration theory, including the fluid phase rheological behavior. Static filtrations were performed in an HTHP filtration cell with non-Newtonian suspensions with concentrations like those used in ac-tual drilling processes. The internal structure of the mudcakes was studied by SEM images. A reparameterization of the static filtration model was proposed for the simultaneous estimation of precise and uncorrelated parameters: filter resistance, permeability, and compressibility of the mudcakes, considering two different meth-odologies, Method I - classic and not advisable -and Method II - rigorous statistical process and pressure-dependent filtration properties. As shown in this research, the confidence intervals for all parameters obtained by Method II were narrow with significative statistical meaning, at a confidence level of 95%, and close to those obtained by other authors. Also, the reparameterization procedure removed the parametric correlation be-tween the mudcake permeability and compressibility, resulting in parameters with null correlation. The numer-ical solution does not lead to a high computational cost and could be installed in an onboard control device to predict online filtration loss, providing valuable pieces of information in oil drilling processes. (c) 2021 Elsevier B.V.

关键词： Filtration Drilling fluids Parameter estimation reparameterization

来源：评论

学校读者我要写书评

暂无评论

reparameterization Approach to Gradient-Based Inverse Design of Three-Dimensional Nanophotonic Devices

引用

ACS PHOTONICS 2023年第4期10卷 815-823页

作者： Gershnabel, Erez Chen, Mingkun Mao, Chenkai Wang, Evan W. Lalanne, Philippe Fan, Jonathan A. Stanford Univ Dept Elect Engn Stanford CA 94305 USA Ben Gurion Int Airport Israel Aerosp Ind IL-7010000 Lod Israel Univ Bordeaux Inst Opt Lab Photon Numer & Nanosci CNRSGrad Sch F-33400 Talence France

We propose a three-dimensional freeform nano-photonic platform in which wavelength-scale domains comprise basic geometric structures with explicitly defined dimensions, positions, orientations, and minimum feature size constraints. Given a desired wavefront shaping objective, these parameters can be collectively optimized using gradient-based shape optimization with full accounting of near-field interactions between structures. We apply our concept to a variety of metagratings supporting high diffraction efficiencies and polarization control, and we experimentally demonstrate a device with a tailored polarization response as a function of wavelength. The combination of device capability, feature size constraints, and ease of manufacturability enabled by our methodology will facilitate the development of robust, high performance, nanophotonic technologies.

关键词： metasurface metagrating reparameterization freeform optimization feature size control polarization control

来源：评论

学校读者我要写书评

暂无评论

reparameterization Causal Convolutional Network for Automatic Modulation Classification

引用

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY 2024年第6期73卷 8576-8583页

作者： Tang, Ning Wang, Xiaoyu Zhou, Fei Tang, Shengyu Lyu, Yaohui Ocean Univ China Coll Informat Sci & Engn Qingdao 266003 Peoples R China Jilin Univ Coll Comp Sci & Technol Changchun 130061 Peoples R China

With the proliferation of wireless technologies in vehicular networks, robust automatic modulation classification (AMC) has become crucial for optimizing spectrum utilization and maintaining reliability. However, AMC in dynamic vehicular channels poses significant challenges for traditional machine learning techniques. This paper proposes a novel CNN-based approach named reparameterization Causal Convolutional Network (RepCCNet) to achieve highly accurate and noise-robust AMC performance. RepCCNet incorporates causal convolutions and structural reparameterization techniques to extract long-term time-domain features. A bottleneck structure with channel attention dynamically calibrates feature channels, retaining only helpful information. Multi-sample dropout is integrated during training to improve generalization capability. We demonstrate RepCCNet's state-of-the-art classification accuracy on the two widely used datasets, RadioML 2016.10a and RadioML 2018.01a, across varying signal-to-noise ratios. Compared to existing methods, RepCCNet achieves highly competitive results compared to state-of-the-art approaches, utilizing fewer than 40 k parameters. Ablation studies validate the contributions of the proposed architectural innovations. This work represents a significant advancement toward developing deep learning solutions for robust wireless signal classification tasks.

关键词： Automatic modulation classification automatic modulation recognition deep learning reparameterization causal convolutional network

来源：评论

学校读者我要写书评

暂无评论

reparameterization of piecewise rational Bezier curves and its applications

引用

VISUAL COMPUTER 2001年第6期17卷 329-336页

作者： Tokuyama, Y Konno, K RICOH Co Ltd Software Res Ctr Bunkyo Ku Tokyo 1120002 Japan Lattice Technol Inc Internet Graph Div Chiyoda Ku Tokyo 1020074 Japan

A piecewise rational Bezier curve is formed by several rational Bezier curve segments. It can be represented by a rational B-spline curve where the multiplicity of each interior knot equals degree. Although the curve segments are C-1 continuous in three dimensions, they may be C-0 continuous in four dimensions. In this case, the multiplicity of each interior knot cannot be reduced and the B-spline basis function becomes C-0 continuous. Using a surface generation method, such as skinning these kinds of rational B-spline curves to construct an interpolatory surface, may generate surfaces with C-0 continuity. This paper presents a reparameterization method for reducing the multiplicity of each interior knot to make the curve segments C-1 continuous in four dimensions. The reparameterized rational B-spline curve has the same shape and degree as before and also has a standard form. Some applications in skinned surface and ruled surface generation based on the reparameterized curves are shown.

关键词： piecewise rational Bezier curve rational B-spline curve reparameterization skinning ruled

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：