检索结果-内蒙古大学图书馆

MBS: A High-Precision Approximation Method for Softmax and Efficient Hardware Implementation

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I-REGULAR PAPERS 2025年

作者： Wu, Yuanchen Xie, Zhiheng Pan, Hongbing Wang, Yuxuan Nanjing Univ Sch Elect Sci & Engn Nanjing 210023 Peoples R China

The softmax function needs to be frequently used in the multi-head attention layer of Transformer networks. Compared to DNNs and other networks, Transformers have higher computational complexity, requiring higher accuracy and hardware performance for softmax function calculations. Therefore, we propose mixed-base softmax (MBS) for the first time for the approximation of the softmax function. This method combines exponential functions with bases of 2 and 4, which is advantageous for hardware implementation. MBS has a high similarity to the softmax function and demonstrates advanced performance during inference in Transformer network. Through algorithm transformation and hardware optimization, we have designed a low-complexity and highly parallel hardware architecture, which only occupies few additional hardware resources compared to base-2 softmax but achieves higher accuracy. Experimental results show that, under TSMC 90nm CMOS technology at the frequency of 0.5 GHz, our design can achieve the efficiency of 236.18 Gps/(mm(2)& sdot;mW) with the area of 4234 mu m(2). Furthermore, MBS exhibits higher computational accuracy and inference precision compared with base-2 softmax.

关键词： Hardware Accuracy Transformers Computational modeling Computer architecture Approximation methods Approximation algorithms Linear approximation inference algorithms Frequency modulation Softmax function mixed-base softmax transformer neural network high-efficiency hardware

来源：评论

学校读者我要写书评

暂无评论

A Single-Stage Photovoltaic Module Defect Detection Method Based on Optimized YOLOv8

引用

IEEE ACCESS 2025年 13卷 27805-27817页

作者： Gao, Yihong Pang, Chengxin Zeng, Xinhua Jiang, Pengyi Shanghai Univ Elect Power Shanghai 200000 Peoples R China Fudan Univ Acad Engn & Technol Shanghai 200433 Peoples R China NYU New York NY 10012 USA

Defect detection in photovoltaic (PV) modules presents significant challenges. In the presence of inconspicuous or small-scale defects, downsampling operations can cause features to vanish, and computational complexity increases as the model deepens, resulting in reduced detection accuracy, higher latency, and challenges in deploying the model on edge devices with limited computational resources. In response, this article proposes a single-stage object detection model based on YOLOv8, the PSA-PVdetector (PSA-det). The core innovation of PSA-det is the novel Partial Spatial Attention (PSA) mechanism, which integrates Partial Convolution (PConv) with Spatial Attention (SA) to optimize feature extraction and reduce computational overhead. For small-scale defect detection, a Multi-Channel Feature Fusion (MCFF) detection head is designed to enable finer-grained center prediction. Additionally, ShapeIoU is employed as the bounding box regression metric, enhancing the classical IoU by incorporating bounding box shape, thereby improving defect localization accuracy. Experimental results demonstrate that PSA-det achieves an mAP50 of 87.2% on the Panel-2 dataset and 72.0% on the Solar dataset. On the Nvidia Jetson Xavier NX DK platform, PSA-det achieves an inference latency as low as 2.6 ms, effectively balancing accuracy and real-time performance.

关键词： Feature extraction Defect detection Head YOLO Convolution Accuracy Inspection Computational modeling inference algorithms Termination of employment YOLOv8 photovoltaic defect detection edge deployment partial convolution spatial attention

来源：评论

学校读者我要写书评

暂无评论

Counterfactual Music Recommendation for Mitigating Popularity Bias

引用

IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS 2025年第2期12卷 851-861页

作者： Yuan, Jidong Gao, Bingyu Wang, Xiaokang Liu, Haiyang Zhang, Lingyin Beijing Jiaotog Univ Key Lab Big Data & Artificial Intelligence Transpo Minist Educ Beijing 100044 Peoples R China Beijing Jiaotog Univ Sch Comp Sci & Technol Beijing 100044 Peoples R China Beijing Univ Posts & Telecommun Sch Econ & Management Beijing 100876 Peoples R China

Music recommendation systems aim to suggest tracks that users may enjoy. However, the accuracy of recommendation results is affected by popularity bias. Previous studies have focused on mitigating the direct effect of single-item popularity in video, news, or e-commerce recommendations, but have overlooked the multisource popularity biases in music recommendations. This article proposes a causal inference-based method to reduce the influence of both track and artist popularity. First, we construct a causal graph that encompasses users, tracks, and artists within the context of music recommendations. Next, we employ matrix factorization in conjunction with counterfactual inference theory to mitigate the popularity effects of artists and tracks, taking into account both the natural direct and indirect effects of these entities on music recommendations. Experimental results evaluated on four music recommendation datasets indicate that our method outperforms other baselines and effectively alleviates the popularity bias of both tracks and artists.

关键词： Recommender systems Music Training inference algorithms Electronic commerce Accuracy Measurement Vectors Transportation Tracking loops Counterfactual inference music recommendation popularity bias

来源：评论

学校读者我要写书评

暂无评论

Intelligence Evaluating Computational Power: A Multi-Factor Method

引用

IEEE ACCESS 2025年 13卷 27398-27415页

作者： Huang, Yuhuan Li, Hang Lu, Dongdong Zhang, Zhe Tie, Wenjie Suzhou Key Lab Microwave Imaging Proc & Applicat T Suzhou 215000 Peoples R China Suzhou Aerosp Informat Res Inst Suzhou 215000 Peoples R China Natl Key Lab Microwave Imaging Beijing 100000 Peoples R China Chinese Acad Sci Aerosp Informat Res Inst Beijing 100000 Peoples R China Univ Chinese Acad Sci Sch Elect Elect & Commun Engn Beijing 100000 Peoples R China

With the rapid development of hardware technologies and the continuous optimization of deep learning tasks, accurately evaluating the computational power of artificial intelligence (AI) systems has become a critical requirement. However, the diversity of hardware architectures and the complexity of deep learning tasks have hindered the establishment of a widely accepted unified evaluation standard. To address this issue, we propose a novel evaluation method that integrates key metrics such as model performance, computational throughput, and energy efficiency, providing a more comprehensive and reliable assessment of AI hardware computational capabilities. Through benchmark tests on servers, statistical analysis of the method, and case studies in cloud, edge, and terminal environments, this approach not only ensures the accuracy of existing evaluation frameworks but also reveals the key characteristics and potential influencing factors of hardware computational power from multiple dimensions.

关键词： Computational modeling Benchmark testing Artificial intelligence Training inference algorithms Deep learning Analytical models Performance evaluation Accuracy Predictive models Deep learning benchmark test AI hardware performance evaluation model performance assessment

来源：评论

学校读者我要写书评

暂无评论

Knowledge-Driven Channel Estimation for Asymmetrical Massive MIMO Systems

引用

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY 2025年第1期74卷 911-924页

作者： Yang, Ruming Xu, Shu Zhu, Zhiming Li, Chunguo Huang, Yongming Yang, Luxi Southeast Univ Sch Informat Sci & Engn Natl Mobile Commun Res Lab Nanjing 210096 Peoples R China Purple Mt Labs Nanjing 211111 Peoples R China

A novel asymmetrical massive multiple-input multiple-output (MIMO) system has recently emerged as a crucial solution for reducing hardware complexity and alleviating the data processing pressure. However, the absence of channel reciprocity in this system presents unique challenges when directly applying traditional channel estimation methods, inevitably leading to performance loss. Deep learning approaches hold promise for achieving improved channel estimation performance by implicitly learning channel features. Unfortunately, deep learning approaches are often designed empirically. It is critical to utilize prior knowledge to develop an efficient deep neural network (DNN), especially for wireless communication systems. This paper explores a knowledge-driven DNN design approach and introduces a deep learning-based channel estimation framework for asymmetrical transceivers. The channel estimation problem is decoupled into channel denoising and information inference problems. Two novel DNNs are proposed to eliminate noise and exploit correlative features for reconstructing the missing channel information, respectively. Extensive simulations demonstrate that our proposed channel estimation framework can significantly eliminate noise effects, even in low signal-to-noise ratio regimes, and outperform traditional estimators and other deep learning-based methods. Moreover, ablation studies also validate the effectiveness of our knowledge-driven network structure design approach.

关键词： Channel estimation Massive MIMO Receiving antennas Radio frequency Transmitting antennas inference algorithms Noise reduction Asymmetrical transceiver channel estimation channel transfer deep learning uplink-to-downlink

来源：评论

学校读者我要写书评

暂无评论

Learning From Crowds Using Graph Neural Networks With Attention Mechanism

引用

IEEE TRANSACTIONS ON BIG DATA 2025年第1期11卷 86-98页

作者： Zhang, Jing Wu, Ming Sun, Zeyi Zhou, Cangqi Southeast Univ Sch Cyber Sci & Engn Nanjing 211189 Jiangsu Peoples R China Southeast Univ Engn Res Ctr Blockchain Applicat Supervis & Manage Minist Educ Nanjing 211189 Jiangsu Peoples R China Hohai Univ Sch Artificial Intelligence & Automat Nanjing 211100 Jiangsu Peoples R China Zhejiang Lab Inst Artificial Intelligence Knowledge Engn Res Ctr Hangzhou 311121 Zhejiang Peoples R China Nanjing Univ Sci & Technol Sch Comp Sci & Engn Nanjing 210094 Jiangsu Peoples R China

Crowdsourcing has been playing an essential role in machine learning since it can obtain a large number of labels in an economical and fast manner for training increasingly complex learning models. However, the application of crowdsourcing learning still faces several challenges such as the low quality of crowd labels and the urgent requirement for learning models adapting to the label noises. There have been many studies focusing on truth inference algorithms to improve the quality of labels obtained by crowdsourcing. Comparably, end-to-end predictive model learning in crowdsourcing scenarios, especially using cutting-edge deep learning techniques, is still in its infant stage. In this paper, we propose a novel graph convolutional network-based framework, namely CGNNAT, which models the correlation of instances by combining the GCN model with an attention mechanism to learn more representative node embeddings for a better understanding of the bias tendency of crowd workers. Furthermore, a specific projection processing layer is employed in CGNNAT to model the reliability of each crowd worker, which makes the model an end-to-end neural network directly trained by noisy crowd labels. Experimental results on several real-world and synthetic datasets show that the proposed CGNNAT outperforms state-of-the-art and classical methods in terms of label prediction.

关键词： Crowdsourcing Noise measurement Data models Biological system modeling Deep learning Task analysis inference algorithms Learning from crowds repeated labeling graph neural networks attention mechanism

来源：评论

学校读者我要写书评

暂无评论

Vehicle Detection and Tracking Based on Improved YOLOv8

引用

IEEE ACCESS 2025年 13卷 24793-24803页

作者： Liu, Yunxiang Shen, Shujun Shanghai Inst Technol Dept Comp Sci & Informat Engn Shanghai 201418 Peoples R China

With the rapid increase in transportation pressure, the demand for efficient traffic recognition and tracking systems is growing. Traditional methods have certain limitations when dealing with complex situations in traffic scenes, such as large weights and insufficient detection accuracies, et al. Therefore, we propose a novel method based on YOLOv8n. Firstly we introduced SCC_Detect based on SCConv on the detection head to reduce the computation of redundant features. Then we replaced the convolutional kernel with a dual convolutional kernel to construct a lightweight deep neural network. Subsequently, the Focaler-EIoU loss function is introduced to improve the accuracy. The BotSORT tracker is embedded in the period of inference, which achieves more accurate and stable recognition and tracking results in the traffic scene. The experimental results show that the proposed model reduces parameters and weight by approximately 36.5% and 25% respectively at the expense of only 0.2% mAP@0.5 compared with YOLOv8n on the UA-DETRAC dataset. In terms of tracking, MOTA, IDF1 and MOTP of the BoTSORT algorithm on the test video were superior to those of DeepSORT and ByteTrack. The accuracy was improved, and the number of lost tracks was reduced. It has a high practical value and application prospects in traffic detection and deployment.

关键词： Feature extraction Accuracy Convolution Roads Target tracking Kernel Convolutional neural networks YOLO Radar tracking inference algorithms Vehicle tracking multi-object track lightweight model YOLOv8 vehicle detection BotSORT

来源：评论

学校读者我要写书评

暂无评论

YOLOV9-CBM: An Improved Fire Detection Algorithm Based on YOLOV9

引用

IEEE ACCESS 2025年 13卷 19612-19623页

作者： Geng, Xin Han, Xiao Cao, Xianghong Su, Yixuan Shu, Dongxue Zhengzhou Univ Light Ind Coll Bldg Environm Engn Zhengzhou 450001 Henan Peoples R China Shanghai Univ Sch Mechatron Engn & Automat Shanghai 200444 Peoples R China

Regarding the current problems of false alarms and missed detections in fire detection, we propose a high-precision fire detection algorithm, YOLOV9-CBM (C3-SE, BiFPN, MPDIoU), by optimizing YOLOV9. Firstly, to tackle the shortage of both quality and quantity in the existing fire datasets, we collected 2,000 fire and smoke images to establish a dataset named CBM-Fire. Secondly, the RepNCSPELAN4 module of the YOLOv9 backbone was replaced with the C3 module containing SE Attention to improve detection efficiency while guaranteeing accuracy. Besides, we transformed the multi-scale fusion network PANet in the baseline algorithm into a bidirectional feature network pyramid BiFPN to facilitate the bidirectional flow of features, enabling the algorithm to fuse information at different scales more effectively. Finally, instead of CIoU losses, we adopted MPDIoU losses in bounding box regression, which improved the accuracy of model regression and classification. Experimental results indicate that compared with YOLOV9, the recall rate of YOLOV9-CBM has increased by 7.6% and the mAP has risen by 3.8%. The revised model demonstrates good generalization performance and robustness. Code and dataset are at https://***/GengHan-123/***.

关键词： Accuracy Feature extraction Attention mechanisms YOLO inference algorithms Detection algorithms Deep learning Biological system modeling Forestry Computer architecture BiFPN C3 fire detection MPDIoU SE YOLOV9

来源：评论

学校读者我要写书评

暂无评论

Real-Time Task Scheduling With Fairness in Digital Twin Systems

引用

IEEE INTERNET OF THINGS JOURNAL 2025年第7期12卷 7846-7862页

作者： Kim, Cheonyong Saad, Walid Han, Jonghun Yu, Tao Sakaguchi, Kei Jung, Minchae Korea Natl Univ Transportat Dept Comp Software Chungju 27469 South Korea Virginia Tech Dept Elect & Comp Engn Arlington VA 22203 USA Korea Railrd Res Inst Smart Elect & Signaling Div Train Control & Commun Res Team Uiwang 16105 South Korea Inst Sci Tokyo Dept Elect & Elect Engn Tokyo 1528550 Japan Sejong Univ Dept Elect & Informat Engn Seoul 05006 South Korea

Digital twin (DT) can help create a digital representation of a physical system, thereby reflecting its real-time status. The digital object, often called cyber twin (CT), facilitates real-time monitoring and control of the physical object, i.e., the so-called physical twin (PT). Owing to this ability, CTs can optimize the PTs and simulate their status, without interrupting the physical world. Given the various CT use cases, one can identify two distinct types of DT tasks: 1) update tasks for PT-CT synchronization and 2) inference tasks for obtaining real-time testing responses. The diverse real-time requirements for update/inference tasks raise the task scheduling problem that has been neglected in previous studies. In this article, the real-time DT task scheduling problem is investigated. In particular, a new approach for evaluating the performance of real-time scheduling of DT tasks is introduced considering the relationship between update/inference tasks and fairness among CTs. Moreover, offline and online DT task scheduling schemes are proposed with the goals of maximizing the DT freshness ratio and minimizing task rejections. In particular, the DT freshness ratio maximization problem is formulated as an offline task scheduling scheme. The proposed offline solution can significantly reduce the solution space without losing optimality. Furthermore, the scheduling policies for achieving the maximal DT freshness ratio are established using which an online scheduling algorithm is designed. Simulation results show that the proposed offline/online schemes increase the DT freshness ratio by at least 16% and 11%, respectively, compared to benchmarks. The results also show that the task rejection ratio of the proposed online algorithm is within 8% of the lower bound.

关键词： Real-time systems Job shop scheduling Testing Optimization Internet of Things Digital twins Synchronization inference algorithms Computational modeling Accuracy Digital twin (DT) fairness offline and online algorithm real-time task scheduling

来源：评论

学校读者我要写书评

暂无评论

ENCODE: Breaking the Trade-Off Between Performance and Efficiency in Long-Term User Behavior Modeling

引用

IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING 2025年第1期37卷 265-277页

作者： Zhou, Wen-Ji Zheng, Yuhang Feng, Yinfu Ye, Yunan Xiao, Rong Chen, Long Yang, Xiaosong Xiao, Jun Alibaba Grp Hangzhou 310000 Peoples R China Hong Kong Univ Sci & Technol Dept Comp Sci & Engn Kowloon Hong Kong 999077 Peoples R China Bournemouth Univ Bournemouth BH12 5BB England Zhejiang Univ Hangzhou 310027 Peoples R China

Long-term user behavior sequences are a goldmine for businesses to explore users' interests to improve Click-Through Rate (CTR). However, it is very challenging to accurately capture users' long-term interests from their long-term behavior sequences and give quick responses from the online serving systems. To meet such requirements, existing methods "inadvertently" destroy two basic requirements in long-term sequence modeling: R1) make full use of the entire sequence to keep the information as much as possible;R2) extract information from the most relevant behaviors to keep high relevance between learned interests and current target items. The performance of online serving systems is significantly affected by incomplete and inaccurate user interest information obtained by existing methods. To this end, we propose an efficient two-stage long-term sequence modeling approach, named as EfficieNt Clustering based twO-stage interest moDEling (ENCODE), consisting of offline extraction stage and online inference stage. It not only meets the aforementioned two basic requirements but also achieves a desirable balance between online service efficiency and precision. Specifically, in the offline extraction stage, ENCODE clusters the entire behavior sequence and extracts accurate interests. To reduce the overhead of the clustering process, we design a metric learning-based dimension reduction algorithm that preserves the relative pairwise distances of behaviors in the new feature space. While in the online inference stage, ENCODE takes the off-the-shelf user interests to predict the associations with target items. Besides, to further ensure the relevance between user interests and target items, we adopt the same relevance metric throughout the whole pipeline of ENCODE. The extensive experiment and comparison with SOTA on both industrial and public datasets have demonstrated the effectiveness and efficiency of our proposed ENCODE.

关键词： Predictive models Measurement Feature extraction Prediction algorithms inference algorithms Electronic commerce Clustering algorithms Knowledge engineering Data models Data mining CTR prediction user behavior modeling metric learning recommendation systems

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：