检索结果-内蒙古大学图书馆

27th European Conference on Artificial Intelligence, ECAI 2024

作者： Ye, Feiyang Lin, Baijiong Cao, Xiaofeng Zhang, Yu Tsang, Ivor W. Department of Computer Science and Engineering Southern University of Science and Technology China Australian Artificial Intelligence Institute University of Technology Sydney Australia The Hong Kong University of Science and Technology Guangzhou China School of Artificial Intelligence Jilin University China Shanghai Artificial Intelligence Laboratory China Centre for Frontier AI Research Agency for Science Technology and Research Singapore Institute of High Performance Computing Agency for Science Technology and Research Singapore School of Computer Science and Engineering Nanyang Technological University Singapore

ISBN: (纸本)9781643685489

In this paper, we study the Multi-Objective Bi-Level Optimization (MOBLO) problem, where the upper-level subproblem is a multi-objective optimization problem and the lower-level subproblem is for scalar optimization. Existing gradient-based MOBLO algorithms need to compute the Hessian matrix, causing the computational inefficient problem. To address this, we propose an efficient first-order multi-gradient method for MOBLO, called FORUM. Specifically, we reformulate MOBLO problems as a constrained multi-objective optimization (MOO) problem via the value-function approach. Then we propose a novel multi-gradient aggregation method to solve the challenging constrained MOO problem. Theoretically, we provide the complexity analysis to show the efficiency of the proposed method and a non-asymptotic convergence result. Empirically, extensive experiments demonstrate the effectiveness and efficiency of the proposed FORUM method in different learning problems. In particular, it achieves state-of-the-art performance on three multi-task learning benchmark datasets. © 2024 The Authors.

关键词： Optimization algorithms

来源：评论

学校读者我要写书评

暂无评论

Combined Effect of Concept Drift and Class Imbalance on Model performance During Stream Classification

引用

computers, Materials & Continua 2023年第4期75卷 1827-1845页

作者： Abdul Sattar Palli Jafreezal Jaafar Manzoor Ahmed Hashmani Heitor Murilo Gomes Aeshah Alsughayyir Abdul Rehman Gilal Department of Computer and Information Sciences Universiti Teknologi PETRONAS(UTP)Seri Iskandar32610Malaysia Centre for Research in Data Science UTPPerak32610Malaysia High Performance Cloud Computing Centre(HPC3) UTPPerak32610Malaysia School of Engineering and Computer Science Victoria University of WellingtonWellington6012New Zealand AI Institute University of Waikato WellingtonHamilton3240New Zealand Anti-Narcotics Force Ministry of Narcotics ControlIslamabad46000Pakistan College of Computer Science and Engineering Taibah UniversityMadinah42353Saudi Arabia

Every application in a smart city environment like the smart grid,health monitoring, security, and surveillance generates non-stationary datastreams. Due to such nature, the statistical properties of data changes overtime, leading to class imbalance and concept drift issues. Both these issuescause model performance degradation. Most of the current work has beenfocused on developing an ensemble strategy by training a new classifier on thelatest data to resolve the issue. These techniques suffer while training the newclassifier if the data is imbalanced. Also, the class imbalance ratio may changegreatly from one input stream to another, making the problem more *** existing solutions proposed for addressing the combined issue of classimbalance and concept drift are lacking in understating of correlation of oneproblem with the other. This work studies the association between conceptdrift and class imbalance ratio and then demonstrates how changes in classimbalance ratio along with concept drift affect the classifier’s *** analyzed the effect of both the issues on minority and majority classesindividually. To do this, we conducted experiments on benchmark datasetsusing state-of-the-art classifiers especially designed for data stream ***, recall, F1 score, and geometric mean were used to measure theperformance. Our findings show that when both class imbalance and conceptdrift problems occur together the performance can decrease up to 15%. Ourresults also show that the increase in the imbalance ratio can cause a 10% to15% decrease in the precision scores of both minority and majority *** study findings may help in designing intelligent and adaptive solutionsthat can cope with the challenges of non-stationary data streams like conceptdrift and class imbalance.

关键词： Classification data streams class imbalance concept drift class imbalance ratio

来源：评论

学校读者我要写书评

暂无评论

Reinforcement Learning for Automated Pragma-Based Loop Optimization

Reinforcement Learning for Automated Pragma-Based Loop Optim...

引用

IEEE Information Technology and Mechatronics engineering Conference (ITOEC)

作者： Jiaxuan Li Qilong Zheng School of Computer Science University of Science and Technology of China (USTC) Hefei China National High Performance Computing Center University of Science and Technology of China (USTC) Hefei China

ISBN: (数字)9798331529482

ISBN: (纸本)9798331529499

Loops often dominate the execution time in high-performance computing, effective loop optimization is critical for overall performance. We propose a reinforcement learning–based framework that automatically discovers and composes transformations—including tiling, fusion, interchange, and unrolling—and evaluate the framework on a subset of Polybench benchmarks. Compared to the Polly compiler baseline, our approach achieves an average speedup of 2.46×, peaking at 7× on the jacobi-1d kernel, while also consistently outperforming a global greedy scheduling algorithm. By adaptively combining multiple transformations, the RL-based method exploits deeper synergies with minimal overhead once trained, thus alleviating the repeated manual tuning and hardware-specific adjustments required by conventional techniques.

关键词： Productivity Mechatronics Scheduling algorithms high performance computing Reinforcement learning Manuals Benchmark testing Kernel Optimization Tuning

来源：评论

学校读者我要写书评

暂无评论

Effective Deep Learning Data Augmentation Techniques for Diabetic Retinopathy Classification

Effective Deep Learning Data Augmentation Techniques for Dia...

引用

2022 International Conference on Machine Learning and Data engineering, ICMLDE 2022

作者： Patil, Mahesh S Chickerur, Satyadhyan Abhimalya, C. Naik, Anishma Kumari, Nidhi Maurya, Shashank School of Computer Science and Engineering K L e Technological University Karnataka Hubballi580031 India Center for High Performance Computing K L e Technological University Karnataka Hubballi580031 India

Diabetic retinopathy(DR) is an eye disorder that affects 80-85% of diabetic patients, and there is a need to detect it in earlier stages to take precautions. Currently, testing for DR is a time-consuming task and requires qualified human resources. Automating DR detection is time-saving and cost-effective, benefiting both doctors and patients. The proposed work automates the DR image detection using deep learning(DL) by using transfer learning. The proposed work provides better results with multi-class classification and generalization than the previous works. i.e a ResNet-50 DL model is trained with the APTOS dataset and tested using the EyePACS dataset. The low accuracy in multi-class classification and data imbalance problems are overcome by using appropriate image pre-processing and augmentation techniques, which have been found by conducting several experiments. Further, the performance of the DL model is enhanced using Test Time Augmentation(TTA). Finally, the authors are able to achieve a multi-class accuracy of 97.87% and Quadratic weighted kappa(QWK) score of 0.985. © 2023 The Authors. Published by Elsevier B.V.

关键词： Classification (of information)

来源：评论

学校读者我要写书评

暂无评论

Study of Data and Model parallelism in Distributed Deep learning for Diabetic retinopathy classification

Study of Data and Model parallelism in Distributed Deep lear...

引用

2022 International Conference on Machine Learning and Data engineering, ICMLDE 2022

作者： Patil, Mahesh S. Chickerur, Satyadhyan School of Computer Science and Engineering K L e Technological University Karnataka Hubballi580031 India Center for High Performance Computing K L e Technological University Karnataka Hubballi580031 India

Distributed deep learning(DDL) is an area of research in Artificial intelligence (AI), where the training time of deep learning(DL) models can be drastically reduced by using multiple accelerators. Most of the previous research works use the data parallelism strategy of DDL and report its performance. But, other parallelism strategies in DDL need to be explored and report its performance modeling for particular applications and software-hardware stacks. Such works can help other researchers make better decisions in designing an efficient DDL algorithm. This work presents a comparative study of data parallelism(DP) and model parallelism(MP) for DDL. The diabetic retinopathy(DR) image classification is used as a case study, and DDL algorithms are implemented using Mesh-Tensor Flow on NVidia's DGX-1. As indicated by the results, the MP strategy provides faster convergence and higher validation accuracy of 62.13% compared to DP strategies validation accuracy of 55.72%. Also, the MP strategy surpasses the DP's highest accuracy in lesser epochs with a speed-up of 1.38X in terms of time duration. Overall, this unique study provides a performance comparison between DP and MP strategies of DDL algorithms being applied for a particular application, dataset, and software-hardware stack. © 2023 The Authors. Published by Elsevier B.V.

关键词： Deep learning

来源：评论

学校读者我要写书评

暂无评论

An improved reconfigurable logic in resistive random access memory

引用

Integration 2022年 87卷 169-175页

作者： Zhao, Yi Chen, Hui Liu, Peng Wu, Jigang Luo, Dongxiang The Center for High Performance Computing and Data Science School of Computer Science and Technology Guangdong University of Technology Guangdong510006 China School of Chemistry and Chemical Engineering/Institute of Clean Energy and Materials/Guangzhou Key Laboratory for Clean Energy and Materials/Huangpu Hydrogen Innovation Center Guangzhou University Guangzhou510006 China Institute of Semiconductors South China Normal University Guangzhou510631 China

Emerging non-volatile devices have shown great potential in computing in-memory (CIM). This work proposes a logic design method based on the complementary resistance switching (CRS) structure, which is connected by two anti-serially memristors. Three logics including AND, OR, and NIMP can be implemented in one step. Moreover, XOR logic can be implemented in two steps. Based on the proposed design, a 1-bit full adder is implemented with the fewer number of memristors and operation steps compared with existing works. The feasibility and correctness of our design are verified by the SPICE simulations with the Voltage Threshold Adaptive Memristor (VTEAM) model. In addition, we evaluate the ISCAS’85 benchmark circuit based on the proposed logic. Compared with existing methods, the performance and area overhead have been improved by the proposed method. © 2022 Elsevier B.V.

关键词： Memristors

来源：评论

学校读者我要写书评

暂无评论

Pipeline Parallelism in Distributed Deep Learning for Diabetic Retinopathy Classification 4th

Pipeline Parallelism in Distributed Deep Learning for Diabet...

引用

4th International Conference on Innovative Data Communication Technologies and Application, ICIDCA 2022

作者： Patil, Shashank Anand Patil, Mahesh S. Giraddi, Shantala Chickerur, Satyadhyan Boormane, Vinayak M. Gamanagatti, Ganesh School of Computer Science and Engineering K L e Technological University Karnataka Hubballi580031 India Center for High Performance Computing K L e Technological University Karnataka Hubballi580031 India

Deep learning (DL) has been used widely and successfully in a variety of industries. It is challenging for researchers to carry out their experiments to find the optimal learning model because training the deep learning model is a time-consuming operation. Distributed deep learning offers a technique for rapid training of DL models (DDL). It is difficult to create and implement DDL, nevertheless, so it is important to report on the performance analysis of different DDL algorithms for the benefit of the research community. The authors of this proposed study use two GPUs to design, construct, and analyse a pipeline parallelism in DDL and then present an analysis of its performance. As a case study for DDL implementation, the authors employ diabetic retinopathy classification. 0.834 is obtained as the validation F1 score. The performance of DDL is reported and examined by the authors, and it is demonstrated to improve training effectiveness in terms of both time and speed. © 2023 The Authors. Published by Elsevier B.V.

关键词： Graphics processing unit

来源：评论

学校读者我要写书评

暂无评论

Complex Permeability Determination for EE Core Based on Flexible PCB Fixture and Magnetic Flux Path Length Analysis 2

Complex Permeability Determination for EE Core Based on Flex...

引用

2nd China Power Supply Society Electromagnetic Compatibility Conference, CPEMC 2024

作者： Jie, Huamin Zhao, Zhenyu Chang, Yongqi Wang, Changdong Jiang, Shiqi Gao, Richard Xian-Ke See, Kye Yak School of Electrical and Electronic Engineering Nanyang Technological University Singapore Singapore Department of Electrical and Computer Engineering National Univeristy of Singapore Singapore Singapore Department of Control Science and Engineering Harbin Institute of Technology Harbin China School of Electronics and Information Engineering Harbin Institute of Technology Harbin China School of Electrical Engineering and Automation Harbin Institute of Technology Harbin China Institute of High Performance Computing A*STAR Singapore Singapore

ISBN: (纸本)9798331509552

The complex permeability of magnetic cores acts as a vital factor in electromagnetic interference (EMI) filtering choke design and optimization. Many efforts have been reported for toroid cores, but fewer target those with other shapes, such as EE, EI, UU, and UI cores. Owing to their commonly used in planar EMI filters, determining the complex permeability of them is of significance. Targeting EE cores, this paper presents a flexible printed circuit board (PCB)-based fixture adapter tailored designed for performing the measurement, while the corresponding core permeability can be calculated through magnetic flux path length analysis. The experiments validate the effectiveness and accuracy of the proposed methods, spanning a wide frequency range up to 120 MHz. ©2024 IEEE.

关键词： Magnetic flux

来源：评论

学校读者我要写书评

暂无评论

An Investigation into Power Aware Aspects of Rendering 3D Models on Multi-Core Processors

An Investigation into Power Aware Aspects of Rendering 3D Mo...

引用

2022 International Conference on Machine Learning and Data engineering, ICMLDE 2022

作者： Konnurmath, Guruprasad Chickerur, Satyadhyan School of Computer Science and Engineering KLE Technological University Hubballi580031 India Centre for High Performance Computing K L e Technological University Hubballi580031 India

Graphics Processing Units (GPUs) are designed to process numerically intensive applications and specially customized with additional computational power to achieve efficient rendering of 3D applications. Shaders are basically the computer programs that run on graphics rendering pipeline and inform about how to process and render each pixel, to perform efficient rendering of 3D applications. Shaders included in our Geometric complexity, texture resolution, per-pixel lighting. These shaders expose GPUs to suffer more power utilization during rendering process. In this paper, initially GPU power usage is investigated at run-time for rendering configuration to define shader parameters on a 3D scene and then power prediction model is proposed at different shader calls on GPU to generate power-aware dynamically reconfigurable rendering model. In the implementation process we integrate the TensorFlow management library routines and dynamic voltage and frequency scheduling (DVFS) mechanism to improve efficiency in rendering process and reach optimal power savings. To analyze the power efficiency of GPUs, shader parameters such as geometric complexity, texture resolution, per-pixel lighting, filtering that expose more towards power consumption are experimented with common GPU workloads during rendering process, then frequency and voltage is altered based on the framerate monitoring. Also in addition to this the frame rate is monitored and maintained in such a way that GPU frequency is adjusted in parallel if the framerate is within the acceptable range so that unnecessary power usage is reduced. During the experimental tests a lower limit of 30 frames per second and upper limit of 60 frames per second is configured. The GPU frequency is reduced if the existing framerate is greater than upper configured limit and frequency range is increased if framerate is below the configured lower limit. Compared with the previous state of art, results show power savings around 40% and imp

关键词： Graphics processing unit

来源：评论

学校读者我要写书评

暂无评论

SMART: Towards Pre-trained Missing-Aware Model for Patient Health Status Prediction 38

SMART: Towards Pre-trained Missing-Aware Model for Patient H...

引用

38th Conference on Neural Information Processing Systems, NeurIPS 2024

作者： Yu, Zhihao Chu, Xu Jin, Yujie Wang, Yasha Zhao, Junfeng School of Computer Science Peking University China Center on Frontiers of Computing Studies Peking University Beijing China National Research and Engineering Center of Software Engineering Peking University China Key Laboratory of High Confidence Software Technologies Ministry of Education Beijing China China Nanhu Laboratory Jiaxing China

Electronic health record (EHR) data has emerged as a valuable resource for analyzing patient health status. However, the prevalence of missing data in EHR poses significant challenges to existing methods, leading to spurious correlations and suboptimal predictions. While various imputation techniques have been developed to address this issue, they often obsess difficult-to-interpolate details and may introduce additional noise when making clinical predictions. To tackle this problem, we propose SMART, a Self-Supervised Missing-Aware RepresenTation Learning approach for patient health status prediction, which encodes missing information via missing-aware temporal and variable attentions and learns to impute missing values through a novel self-supervised pre-training approach which reconstructs missing data representations in the latent space rather than in input space as usual. By adopting elaborated attentions and focusing on learning higher-order representations, SMART promotes better generalization and robustness to missing data. We validate the effectiveness of SMART through extensive experiments on six EHR tasks, demonstrating its superiority over state-of-the-art methods. Our code is available at https://***/yzhHoward/SMART. © 2024 Neural information processing systems foundation. All rights reserved.

关键词：

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：