检索结果-内蒙古大学图书馆

21st International computer conference on Wavelet Active Media Technology and Information Processing, ICCWAMTIP 2024

作者： Li, Xu Li, Jianping Khoury College of Computer Science Northeastern university BostonMA02115 United States School of Computer Science and Engineering University of Electronic Science and Technology of China Chengdu610054 China

ISBN: (纸本)9798331519254

This study investigates the impact of various feature enhancement methods on the accuracy of a deep learning model used for the classification of wild animal. We specifically compare three attention-based mechanisms, Convolutional Block Attention Module (CBAM)., Dual Attention Network (DANet), and Bottleneck Attention Module (BAM), each enhancing feature focus through different attention structures. Using the Animal Image Dataset from Kaggle, we evaluate the classification performance across 90 animal species. The CBAM achieves the highest test accuracy of 92.2 %, while DANet got 91.58% and the accuracy for BAM is 88.5%. In addition, a model applied all three methods sequentially was also examined, its redundancy and overfitting lead to a slight decrease in accuracy, which is 89.73%. Our findings suggest that the CBAM's sequential attention mechanism highlight the important features in the most effective way, making it a suitable choice for complex multi-class image classification tasks. © 2024 IEEE.

关键词： Invertebrates

来源：评论

学校读者我要写书评

暂无评论

high-performance FFT Code Generation via MLIR Linalg Dialect and SIMD Micro-Kernels

High-Performance FFT Code Generation via MLIR Linalg Dialect...

引用

2024 IEEE International conference on Cluster Computing, CLUSTER 2024

作者： He, Yifei Markidis, Stefano KTH Royal Institute of Technology Stockholm Sweden

ISBN: (纸本)9798350358711

Fast Fourier Transform (FFT) libraries are an indispensable and critical component of any high-performance Computing (HPC) software stack. They are used in many applications, from Partial Differential Equation (PDE) solvers to signal spectral analysis and deep-learning. To design and develop the next generation of HPC FFT libraries, it is essential to leverage modern compiler infrastructures, such as Multi-Level Intermediate Representation (MLIR) and the Low-Level Virtual Machine (LLVM), alongside advanced computer architecture features, including SIMD (Single Instruction, Multiple Data) instructions. In this work, we introduce FFTc 2.0, an MLIR-based domain-specific compilation framework for FFT. We extend the MLIR Linalg dialect with FFT-specific operations to harness high-level tensor-based abstractions for FFT. These abstractions are ideal for formulating various FFT algorithms and facilitating formula rewriting for FFT decomposition and cache-friendly optimization. We employ micro-kernels to increase the performance, particularly in response to the limited support for complex arithmetic in MLIR and the general-purpose compiler LLVM. These micro-kernels are designed to implement small-size FFT kernels for integration into larger FFTs. They utilize SIMD-friendly data layouts for complex arithmetic and feature an optimized memory access pattern, enabling performance enhancements that are not achievable with standard compiler implementations. Our method achieves performance levels comparable to or surpass those of widely used FFT HPC libraries, such as FFTW, FFTE, and Spiral FFT. Additionally, it establishes a robust software infrastructure that facilitates further optimizations and supports additional hardware backends. © 2024 IEEE.

关键词： program compilers

来源：评论

学校读者我要写书评

暂无评论

Dataset Augmentation for Robust Spiking Neural Networks 4

Dataset Augmentation for Robust Spiking Neural Networks

引用

4th IEEE International conference on Automatic Computing and Self-Organizing Systems Companion, ACSOS-C 2023

作者： Baietto, Anthony Stewart, Christopher Bihl, Trevor J. The Ohio State University Computer Science and Engineering Columbus United States Air Force Research Laboratory Wright-Patterson Air Force Base United States

ISBN: (纸本)9798350337464

In spiking neural networks (SNNs), neurons are connected in layers but, unlike artificial neural networks (ANNs), they transmit output signals only after their input signals exceed the activation threshold. By eliding unnecessary transmissions, processors designed for SNNs can consume much less power than processors designed for ANNs, making SNNs a promising architecture for energy-constrained datacenters and Internet of Things (IoT) devices. However, training SNNs to perform machine learning tasks as well as ANNs is challenging because backpropagation, a widely used technique to train ANNs, cannot infer the changing subset of transmitting neurons and the duration of their transmissions for each input. State-of-the-art SNN platforms provide platform-specific, mechanistic models to characterize neuron activations;however, these models are often heavily tied to the specific spike distribution. In this paper, we show that SNNs trained on state-of-the-art platforms perform poorly when presented with different spike distributions. We present a platform-agnostic approach that automatically learns neuron activations from observations. Precisely, we use established approximations, combined with a generative adversarial network (GAN) to augment the training dataset with broader spike distribution data. Our approach achieved 54.57 % accuracy on the CIFAR-10 dataset with an average 1.80% improvement in accuracy over existing state-of-the-art SNNs when evaluated on differing spike distributions. These preliminary results validate our approach and lay the groundwork for future research into strengthening SNN models. © 2023 IEEE.

关键词： Generative adversarial networks

来源：评论

学校读者我要写书评

暂无评论

Effectiveness of Consumer-Grade EEG Devices in Brainwaves Signal Processing: A Review

Effectiveness of Consumer-Grade EEG Devices in Brainwaves Si...

引用

2024 IEEE Region 10 conference, TENCON 2024

作者： Pigtain, Rhou Ian A. Oppus, Carlos M. Agustin, Luisito L. Legarda, Jon M. Ateneo de Manila University Department of Electronics Computer and Communications Engineering Quezon City Philippines Deusto Institute of Technology Univesidad de Deusto Bilbao Spain

ISBN: (纸本)9798350350821

Electroencephalography (EEG) is a versatile tool for neuroscience research and medical applications. While high-end research-grade EEG devices offer superior accuracy and performance, they can be costly and complex. This study evaluates the suitability of consumer-grade EEG devices for brainwave signal processing tasks. Two popular consumer-grade EEG devices, the Emotiv Epoc X and the NeuroSky Mindwave Mobile, were compared. Their performance was assessed in three recognition tasks: concentration, relaxation, and eye blink. The Emotiv Epoc X consistently outperformed the NeuroSky Mindwave Mobile in all tasks. Notably, the Epoc X achieved a recognition accuracy of over 75 percent for eye blinks, while the Mindwave Mobile fell below 50 percent. These findings suggest that the Emotiv Epoc X is a more promising option for brainwave signal processing in research and medical applications. While consumer-grade devices may not always match the performance of research-grade equipment, they offer a more accessible and cost-effective alternative for certain applications. © 2024 IEEE.

关键词： Electroencephalography

来源：评论

学校读者我要写书评

暂无评论

Semi-Supervised Medical Image Segmentation with Multimodal Data Augmentation and Self-Supervised Learning 21

Semi-Supervised Medical Image Segmentation with Multimodal D...

引用

21st International computer conference on Wavelet Active Media Technology and Information Processing, ICCWAMTIP 2024

作者： Decahng, Duan School of Computer Science and Engineering University of Electronic Science and Technology of China Chengdu611731 China

ISBN: (数字)9798331519254

ISBN: (纸本)9798331519254

Medical image segmentation is a critical task in clinical diagnosis and treatment. However, the high cost of acquiring annotated data and the scarcity of labeled datasets have made semi-supervised learning methods a focal point of research in this field. In this paper, we propose a semi-supervised medical image segmentation method that combines multimodal data augmentation, self-supervised pretraining, and pseudo-label generation and correction strategies. By leveraging complementary information from different modalities such as CT and MRI, along with self-supervised pretraining on unlabeled data, our model significantly improves segmentation performance under limited labeled data. Moreover, a confidence-based pseudo-label generation and correction strategy further enhances the label quality during the training process. Experimental results demonstrate that the proposed method outperforms traditional fully-supervised and semi-supervised approaches across multiple medical image segmentation datasets, with notable improvements in segmentation accuracy, especially in scenarios with scarce labeled data. This method provides an effective solution for medical image segmentation tasks and shows great potential for widespread application, particularly in medical image analysis with limited annotated data. © 2024 IEEE.

关键词： Self-supervised learning

来源：评论

学校读者我要写书评

暂无评论

Research on Spiking Neural Network with Additional Time Dimension for Sequencial Modeling 21

Research on Spiking Neural Network with Additional Time Dime...

引用

21st International computer conference on Wavelet Active Media Technology and Information Processing, ICCWAMTIP 2024

作者： Zhang, Enqi School of Computer Science and Engineering University of Electronic Science and Technology of China Chengdu China

ISBN: (纸本)9798331519254

Spiking neural networks have gained significant traction in image and sequence tasks due to low power consumption. However, a comprehensive theoretical understanding of how information is stored within spike trains remains underexplored. This study analyzes the fundamental mechanisms of information storage and transmission in spike trains from a theoretical standpoint and introduces a new framework: the spiking neural network model with multiple time dimensions. Given the equivalence of this model's nature to artificial neural networks with quantized activation values, we evaluated the performance of these equivalent ANNs in sequential tasks. The experimental results establish a theoretical foundation that could guide the future development of more efficient and interpretable spiking neural networks for sequential tasks. © 2024 IEEE.

关键词： Neural network models

来源：评论

学校读者我要写书评

暂无评论

Domain Adapted Object Detection with Intermediate Domain Generation 21

Domain Adapted Object Detection with Intermediate Domain Gen...

引用

21st International computer conference on Wavelet Active Media Technology and Information Processing, ICCWAMTIP 2024

作者： Yuhang, He Yangyang, Sun Zhijun, Xue Shijiao, Han School of Computer Science and Engineering University of Electronic Science and Technology of China Chengdu China Beijing Institute of Control and Electronic Technology Beijing China

ISBN: (纸本)9798331519254

Autonomous driving detection technology in real-world road scenarios faces numerous challenges, including variable weather conditions and complex road environments. Therefore, developing an object detection model with robust domain adaptation is crucial. In this paper, we investigate the use of image style transfer techniques to leverage target domain images for enhancing model performance. Experimental results show that our proposed approach significantly improves detection efficacy. Notably, our method outperforms the Oracle results in tasks such as transitioning from Cityscapes to Foggy Cityscapes, highlighting its effectiveness in addressing domain adaptation challenges. © 2024 IEEE.

关键词： Image enhancement

来源：评论

学校读者我要写书评

暂无评论

Development of a Vision-based Automated Hole Assembly System with Quality Inspection

引用

MANUFACTURING LETTERS 2023年 35卷 64-73页

作者： Kim, Doowon TabkhPaz, Majid Park, Simon S. Lee, Jihyun Univ Calgary Mech & Mfg Engn 2500 Univ Dr NW Calgary AB T2N 1N4 Canada GN Corp Inc 2873 Kingsview Blvd SE Airdrie AB T4A 0E1 Canada

This study proposes a novel mechatronics system capable of automating a peg-in-hole assembly process and inspecting the quality of the assembly with vision. The proposed mechatronics system integrates a customized peg insertion tool, a new assembly mechanism, and a control algorithm to efficiently insert pegs into holes with a tolerance of 200 mu m. The system improves assembly performance by utilizing dual cameras and several computer vision techniques, including Contrasted Limited Adaptive Histogram Equalization, Canny Edge Detector, Hough Circle Transform, and YOLOv5. In addition, a real-time statistical quality inspection method is proposed and compared with machine learning-based inspection approaches. Various surface textures and materials of the cylindrical workpiece are used to assess the robustness of the proposed inspection methods. Experimental results demonstrate that the proposed system and quality inspection methods exhibit high performance and adaptability. (c) 2023 The Authors. Published by ELSEVIER Ltd.

关键词： Manufacturing Process Automation computer Vision Hole Assembly Automation Quality Inspection

来源：评论

学校读者我要写书评

暂无评论

Ant colony optimization based multi-objective task scheduling algorithm for supercomputing Internet workflow 3

Ant colony optimization based multi-objective task schedulin...

引用

3rd International conference on high performance Computing and Communication Engineering, HPCCE 2023

作者： Fu, Peng Wang, Zhihao Guo, Meng Shandong Jinan250013 China Shandong Provincial Key Laboratory of Computer Networks Shandong Fundamental Research Center for Computer Science Shandong Jinan250013 China

ISBN: (数字)9781510674653

ISBN: (纸本)9781510674646

Supercomputing Internet is based on supercomputers and high-speed Internet, achieving the interconnection of data and computing resources between supercomputing centers in various regions, solving the problem of uneven distribution of computing resources, and providing diversified computing services. Workflow tasks consist of a large number of related tasks, and task scheduling of workflows is crucial in the supercomputing Internet. Good algorithms help provide users with higher quality services. Traditional heuristic methods usually focus on shortening the makespan and tend to prioritize the use of resources with stronger computing power. However, this method is prone to falling into local optima and increases cost overhead. In this article, we model the makespan of workflows and the cost of supercomputing resources as multi-objective optimization problems and propose a multi-objective ant colony algorithm (MOACO). The experimental results show that the algorithm effectively reduces makespan and cost, and fully utilizes available resources. © 2024 SPIE.

关键词： Heuristic methods

来源：评论

学校读者我要写书评

暂无评论

MALAGA - MultidimensionAL Big DAta Analytics over Massive Graph DAta

MALAGA - MultidimensionAL Big DAta Analytics over Massive Gr...

引用

2024 IEEE International conference on Big Data, BigData 2024

作者： Cuzzocrea, Alfredo Hajian, Mojtaba Hafsaoui, Abderraouf University of Calabria IDEA Lab Rende Italy University of Paris City Dept. of Computer Science Paris France

ISBN: (纸本)9798350362480

Focusing on the main research context represented by the issue of supporting big data analytics over big graph data, this paper introduces and experimentally assesses the framework MALAGA (MultidimensionAL Big DAta Analytics over Massive Graph DAta). MALAGA incorporates several innovations, including OLAP analysis of big graph data, columnar-OLAP methodologies, and Apache Hive extensions. A comprehensive experimental assessment and analysis of the framework's performance is finally presented and discussed, by significantly integrating the conceptual contributions of our research. © 2024 IEEE.

关键词： Big Data Big Data Analytics Big Graph Data Massive Graph Data Multidimensional Big Data Analytics

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：