检索结果-内蒙古大学图书馆

arXiv 2025年

作者： Peng, Dengyun Zhou, Yuhang Chen, Qiguang Liu, Jinhao Chen, Jingjing Qin, Libo Research Center for Social Computing and Interactive Robotics Harbin Institute of Technology China School of Computer Science and Engineering Central South University China Fudan University China

Large Language Models (LLMs) have achieved remarkable success across diverse tasks, largely driven by well-designed prompts. However, crafting and selecting such prompts often requires considerable human effort, significantly limiting its scalability. To mitigate this, recent studies have explored automated prompt optimization as a promising solution. Despite these efforts, existing methods still face critical challenges in robustness, efficiency, and generalization. To systematically address these challenges, we first conduct an empirical analysis to identify the limitations of current reflection-based prompt optimization paradigm. Building on these insights, we propose 7 innovative approaches inspired by traditional deep learning paradigms for prompt optimization (DLPO), seamlessly integrating these concepts into text-based gradient optimization. Through these advancements, we progressively tackle the aforementioned challenges and validate our methods through extensive experimentation. We hope our study not only provides valuable guidance for future research but also offers a comprehensive understanding of the challenges and potential solutions in prompt optimization. Our code is available at https://***/sfasfaffa/DLPO. Copyright © 2025, The Authors. All rights reserved.

关键词： Contrastive Learning

来源：评论

学校读者我要写书评

暂无评论

Cloud-Enhanced Machine Learning: Exploring the Synergy for Intelligent Automation and Optimization 15th

Cloud-Enhanced Machine Learning: Exploring the Synergy for I...

引用

15th International Conference on Soft Computing and Pattern Recognition, SoCPaR 2023 and 14th World Congress on Nature and Biologically Inspired Computing, NaBIC 2023

作者： Kumar, Naween Dash, Yajnaseni Abraham, Ajith Choudhary, Rakesh Kumar Pandey, Shriyansh School of Computer Science Engineering and Technology Bennett University UP Greater Noida India School of Artificial Intelligence Bennett University UP Greater Noida India School of Artificial Intelligence Sai University Chennai India

ISBN: (纸本)9783031810855

The convergence of cloud computing and machine learning (ML) is evolving in the new era of technological innovation. This paper elaborates on the symbiotic relationship between these two fields, highlighting how cloud infrastructure enables scalable, flexible, and cost-efficient ML development and deployment. At the same time, ML enhances cloud services through intelligent automation, optimization, and data-driven insights. This study of synergy presents transformative and revolutionizing opportunities in the industry and opens doors to evolved ML models with the help of cloud environments. Additionally, the paper mentions best practices, challenges, and future trends in integrating ML and cloud technologies, providing a comprehensive guide for students, professionals, and researchers. Also, the pre-existing research on the subject has been reviewed thoroughly and summarized, making this study a guide for researchers. Through performance analyses, we illustrate the practical benefits and certain pitfalls of this integration, highlighting valuable insights for leveraging the full potential of ML and cloud synergy. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

关键词： Artificial Intelligence Cloud Computing Machine learning Model Deployment

来源：评论

学校读者我要写书评

暂无评论

Efficient terahertz radiation generation from air plasma with mid- and near-infrared three-color femtosecond pulses

引用

Optics Express 2025年第11期33卷 22610-22631页

作者： Bintao Fu Zhiwen Zhang Zefu Liu Kang Wang Liwei Song Wenkai Li Vasily A Kostin Alexer A Silaev Longzhen Zhuang Chenhui Lu Yi Liu School of Optical Electrical and Computer Engineering University of Shanghai for Science and Technology Shanghai 200093 China State Key Laboratory of Ultra-Intense Laser Science and Technology Shanghai Institute of Optics and Fine Mechanics Chinese Academy of Sciences Shanghai 201800 China Gaponov-Grekhov Institute of Applied Physics of the Russian Academy of Sciences Nizhny Novgorod 603950 Russia Lobachevsky State University of Nizhny Novgorod Nizhny Novgorod 603022 Russia School of Mechanical and Automotive Engineering Shanghai University of Engineering Science Shanghai 201620 China

We report on the generation of the terahertz (THz) radiation from air plasma driven by a three-color femtosecond field with the fundamental pulse in the near- (1700 nm and 800 nm) and mid-infrared (3900 nm) ranges. With an inline optical setup for three-color pulse generation and relative phase control, we observed that the THz pulse energy can be enhanced up to 6 times compared with the conventional two-color excitation scheme, dependent on the incident fundamental pulse energy. The polarization of the enhanced THz pulse was found to be elliptical, with the second and third harmonics components of the three-color pump also being elliptically polarized. With the 3.9 µm driving pulses, the maximum THz pulse energy reaches 39.4 µJ with a conversion efficiency of about 1%. Based on the plasma current model, we analytically and numerically reproduced the THz energy enhancement, as well as its dependence on incident pulse energy. The THz enhancement can be attributed to the two additional ionization wave mixing channels introduced with the addition of the third harmonic.

关键词： Femtosecond lasers Femtosecond pulses Nonlinear optical crystals Optical fields Sum frequency generation Ultrafast lasers

来源：评论

学校读者我要写书评

暂无评论

Digitalization in Agriengineering 5.0 24th

Digitalization in AgriEngineering 5.0

引用

24th International Symposium for Production Research, ISPR 2024

作者： Aksoy, Serra Demircioglu, Pinar Bogrekci, Ismail Bauer, Jorge Institute of Computer Science Ludwig Maximilian University of Munich (LMU) Munich Germany Institute of Materials Science TUM School of Engineering and Design Technical University of Munich Garching Germany Department of Mechanical Engineering Faculty of Engineering Aydin Adnan Menderes University Aydin Turkey Institute of Production Engineering and Photonic Technologies TU Wien (Vienna University of Technology) Vienna Austria UTN-Universidad Tecnológica Nacional Buenos Aires Argentina

ISBN: (纸本)9783031836107

Agriengineering 5.0 is the next big leap in farming, blending advanced technology with smart engineering to create more efficient and sustainable agricultural practices. With tools like AI, IoT, and robotics, farmers can now manage their crops and resources with precision, making better decisions based on real-time data. Autonomous machines, such as drones and self-driving tractors, reduce manual labor and help optimize water and fertilizer use. Digital twins (virtual farm models) let farmers predict outcomes and adjust their strategies before issues arise. At the same time, blockchain improves transparency, helping consumers know exactly where their food comes from. While Agriengineering 5.0 offers exciting opportunities, challenges like cost, access to technology, and training for farmers still need to be addressed to unlock its full potential. Agriengineering 5.0 is transforming farming by using advanced technologies like robotics, AI, and 6G to automate tasks, monitor crops and livestock in real-time, and make smarter, data-driven decisions. This new approach focuses on precision, adapting to the specific needs of each farm to boost efficiency and sustainability. Tools like big data and natural language processing make it easier for farmers to analyze information and communicate more effectively. However, challenges like overcoming reliance on certain technologies and encouraging farmers to embrace new methods need to be addressed for widespread adoption. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

关键词： Agriengineering 5.0 Artificial intelligence Data-driven decision-making Smart farming Sustainability

来源：评论

学校读者我要写书评

暂无评论

Sub-network Knowledge Injection and Transferable Parameter Updating Strategy for Continual Learning of Vision-and-Language Tasks

IEEE Transactions on Artificial Intelligence

引用

IEEE Transactions on Artificial Intelligence 2025年

作者： Wang, Jiong Zhang, Qing Liu, Jie Nie, Kaifeng Zhang, Ze Song, Linqi Li, Yinqiao Beijing Normal University School of Artificial Intelligence Beijing100091 China School of Artificial Intelligence and Computer Science North China University of Technology Beijing100144 China Capital Normal University Beijing100048 China Capital Normal University School of Information Engineering Beijing100048 China City University of Hong Kong Department of Computer Science 999077 Hong Kong

The external environment of the real world is often filled with various types of multimodal information that are highly dynamic and unpredictable. Therefore, intelligent systems need to continuously learn and retain knowledge. Continual learning (CL) provides a foundation for vision-and-language (VaL) models to adapt to the changes in the real world. Existing research on CL for VaL models mainly focus on preserving memory stability to overcome catastrophic forgetting. However, it is difficult for existing methods to flexibly adapt to the dynamic growth of external information. In this paper, we propose a generic approach that appropriately attenuates and employs old memories in parameter distributions to improve learning plasticity. We propose a new parameter-sharing CL method that combines adaptive parameter update strategies with network pruning to enhance the plasticity of VaL models during the continual learning process. When learning a new task, we utilize the pruned model for learning new tasks. And then we employ the sub-network parameter initialization updating strategy to transfer the most similar important knowledge from learned tasks. Subsequently, to strength the stability of the model, we further update the model parameters, adjusting old memories to better adapt to new task learning. Experiments on a series of VaL tasks have shown that our proposed method outperforms the compared methods in improving model plasticity, and is more stable than existing parameter sharing methods. © 2020 IEEE.

关键词： Multi-task learning

来源：评论

学校读者我要写书评

暂无评论

COO-DuDo: computation overhead optimization methods for dual-domain sparse-view CT reconstruction

引用

Expert Systems with Applications 2025年 286卷

作者： Deng, Zihan Wang, Zhisheng Shan, Yuanlin He, Guohang Du, Tiantian Wang, Shunli Center of Ultra-Precision Optoelectronic Instrument Engineering Harbin Institute of Technology Harbin150080 China Ministry of Industry and Information Technology Harbin150080 China Artificial Intelligence Media Computing Laboratory School of Computer Science and Technology Harbin Engineering University Harbin150080 China Guangxi Key Laboratory of Multimedia Communications and Network Technology School of Computer Electronic and Information Guangxi University Nanning530000 China Faculty of Education University of Hong Kong Hong Kong

Recently, deep learning methods have shown exciting effects in Sparse-view CT reconstruction. The Dual-Domain (DuDo) deep learning method is one of the representative methods, and it can process the information in both the sinogram and image domains. However, the existing DuDo methods do not pay enough attention to the allocation of training costs and strategies for the two domains, which will result in wasted computing overhead or insufficient training in one of the two domains. In this paper, we propose a Computation-Overhead Optimization (COO) DuDo training strategy for sparse-view CT reconstruction, i.e., COO-DuDo. The training ratio of different domains is controlled by calculating their computation overhead, loss, and gradient variation of the loss. To make our COO-DuDo strategy enable sparse-view CT reconstruction better, we adopt a DuDo-Network (COO-DDNet) structure based on two coding-decoding-type subnetworks. The evaluation, animal and clinical experiments have verified the effectiveness of our training strategies and methods. Our research provides a broader perspective for dual-domain image restoration tasks from the perspective of computational overhead. © 2025 Elsevier Ltd

关键词： Deep learning

来源：评论

学校读者我要写书评

暂无评论

DQ-Data2vec: Decoupling Quantization for Multilingual Speech Recognition

arXiv

引用

arXiv 2025年

作者： Shao, Qijie Dong, Linhao Wei, Kun Sun, Sining Xie, Lei School of Computer Science and Engineering Northwestern Polytechnical University Xi'An710072 China Bytedance Speech Beijing Bytedance Technology Co Ltd. Beijing100098 China

Data2vec is a self-supervised learning (SSL) approach that employs a teacher-student architecture for contextual representation learning via masked prediction, demonstrating remarkable performance in monolingual ASR. Previous studies have revealed that data2vec's shallow layers capture speaker and language information, middle layers encode phoneme and word features, while deep layers are responsible for reconstruction. Language and phoneme features are crucial for multilingual ASR. However, data2vec's masked representation generation relies on multi-layer averaging, inevitably coupling these features. To address this limitation, we propose a decoupling quantization based data2vec (DQ-Data2vec) for multilingual ASR, which includes a data2vec backbone and two improved online K-means quantizers. Our core idea is using the K-means quantizer with specified cluster numbers to decouple language and phoneme information for masked prediction. Specifically, in the language quantization, considering that the number of languages is significantly different from other irrelevant features (e.g., speakers), we assign the cluster number to match the number of languages, explicitly decoupling shallow layers' language-related information from irrelevant features. This strategy is also applied to decoupling middle layers' phoneme and word features. In a self-supervised scenario, experiments on the CommonVoice dataset demonstrate that DQ-Data2vec achieves a relative reduction of 9.51% in phoneme error rate (PER) and 11.58% in word error rate (WER) compared to data2vec and UniData2vec. Moreover, in a weakly-supervised scenario incorporating language labels and high-resource language text labels, the relative reduction is 18.09% and 1.55%, respectively. Copyright © 2025, The Authors. All rights reserved.

关键词： Self-supervised learning

来源：评论

学校读者我要写书评

暂无评论

UNet-CrossViT: A Deep Learning Model for Classifying Alzheimer's Disease

UNet-CrossViT: A Deep Learning Model for Classifying Alzheim...

引用

Computing and Information technology (ICCIT-1441), 2020 International Conference on

作者： Anitha Thavasimuthu P. Jesu Jayarain Narmatha C. Department of Computer Science and Engineering Saveetha School of Engineering Saveetha Institute of Medical and Technical Sciences Chennai Tamilnadu India Department of Information Technology Saveetha School of Engineering Saveetha Institute of Medical and Technical Sciences Chennai Tamilnadu India Faculty of Computers and Information Technology University of Tabuk Saudi Arabia

ISBN: (数字)9798350353839

ISBN: (纸本)9798350353846

In recent times, deep-learning (DL) methods are implemented to classify neuroimaging data associated with Alzheimer's disease (AZD), and notable advancements have been achieved. Traditional Machine Learning (ML) has been shown to be helpful in aiding the diagnosis of AZD, however only a limited number of research have explored the capabilities of DL and deep transfer learning (DTL) in this intricate process. In this research, a novel DTL model called Cross attention multi scale Vision Transformer (CrossViT) is proposed to classify the AZD using MRI images. For this classification, a set of MRI images of AZD is collected from the Kaggle website. The collected dataset has 6400 images of MRI related to AZD and its various stages. The MRI images are preprocessed and divided into training and testing sets as 80% and 20%. After preprocessing, the segmentation is performed using the UNet architecture. The segmented images are fed to the proposed CrossViT model for AZD classification. The classification is performed as a binary classification, as it classified the results as either the AZD is present or not. The performances of the research model are assessed based on accuracy, specificity, sensitivity, f1-score, and precision. The research model has attained 99.63% accuracy, 99.51% sensitivity, 99.69% precision, 99.35% specificity, and 99.50% f1-score in training. The model attained 99.21% accuracy, 98.94% sensitivity, 99.25% precision, 99.10% specificity, and 99.04% f1-score in testing. The test results were compared with the current models for validation. The research model has outperformed all the compared model in comparison.

关键词： Training Image segmentation computer vision Sensitivity Accuracy Magnetic resonance imaging Computational modeling Transformers Alzheimer's disease Testing

来源：评论

学校读者我要写书评

暂无评论

A Lightweight Hybrid CNN-Fuzzy Logic Approach for Real Time On-Device Document Classification 25

A Lightweight Hybrid CNN-Fuzzy Logic Approach for Real Time ...

引用

Proceedings of the 18th Innovations in Software engineering Conference

作者： Shashank Mouli Satapathy Akila Victor Kartik Kumar Gounder Vishal Agrawal Riddhi Snehal Panchal Amith Kumar Mahale School of Computer Science and Engineering Vellore Institute of Technology Vellore Vellore Tamil Nadu India Samsung R&D Institute Bangalore Bengaluru Karnataka India

来源：评论

学校读者我要写书评

暂无评论

Computation Overhead Optimization Dual-Domain Network with Multi-Level Cross-Domain Connections for Sparse-view CT Reconstruction

Computation Overhead Optimization Dual-Domain Network with M...

引用

Acoustics, Speech, and Signal Processing Workshops (ICASSPW), IEEE International Conference on

作者： Zihan Deng Zhisheng Wang Shanyuan Lin Guohang He Demin Jiang Shunli Wang Ultra Precision Optoelectronic Instrument Engineering Research Institute Harbin Institute of Technology Harbin China School of Computer Science and Technology Harbin Engineering University Harbin China Guangxi Key Laboratory of Multimedia Communications and Network Technology Guangxi University Nanning China

ISBN: (数字)9798331519315

ISBN: (纸本)9798331519322

To reduce the harmful effects of radiation dose on the human body, the sparse scanning method is widely used. Recently, deep learning methods have been widely used in rather challenging reconstruction tasks such as sparse angles, and the dual domain method has a natural advantage because it can process the information in both the sine domain and the image domain. However, the existing two-domain methods do not pay enough attention to the allocation of training costs and strategies for the two domains. To solve this problem, in this paper, we design a Computation Overhead Optimization Dual-Domain Network with Multi-Level Cross-Domain Connections for Sparse View CT Reconstruction. We carries out feature extraction and image recovery for projection and image respectively through two Transformer-based subnets of coding-decoding type. In particular, in the two domains, we use the multilevel cross-domain connection method to connect the coding layers of the same scale in the two subnetworks, and adopt the two-channel method for up-sampling .The experimental results on clinical data validate the effectiveness of our method and training strategy. From the subjective and objective evaluation indexes of reconstruction results, our method can outperform some of the most advanced methods in the past.

关键词： Training Deep learning Computed tomography Signal processing Transformers Resource management Speech processing Positron emission tomography Image reconstruction Optimization

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：