检索结果-内蒙古大学图书馆

IEEE International conference on Big Data and Cloud Computing (BdCloud)

作者： Huan Wang Lijuan Wang Jun Shen School of Cyber Engineering Xidian University Xi'an China School of Computing & Information Technology University of Wollongong Wollongong Australia

Federated learning (FL) strives to enable collaborative training of deep models on the distributed clients of different data without centrally aggregating raw data and hence improving data privacy. Nevertheless, a central challenge in training classification models in the federated system is learning with non-IID data. Most of the existing work is dedicated to eliminating the heterogeneous influence of non-IID data in a federated system. However, in many real-world FL applications, the co-occurrence of data heterogeneity and long-tailed distribution is unavoidable. The universal class distribution is long-tailed, causing them to become easily biased towards head classes, which severely harms the global model performance. In this work, we also discovered an intriguing fact that the classifier logit vector (i.e., pre-softmax output) introduces a heterogeneity drift during the learning process of local training and global optimization, which harms the convergence as well as model performance. Therefore, motivated by the above finding, we propose a novel logit calibration FL method to solve the joint problem of non-IID and long-tailed data in federated learning, called Federated Learning with Logit Calibration (FedLC). First, we presented a method to mitigate the local update drift by calculating the Wasserstein distance among adjacent client logits and then aggregating similar clients to regulate local training. Second, based on the model ensemble, a new distillation method with logit calibration and class weighting was proposed by exploiting the diversity of local models trained on heterogeneous data, which effectively alleviates the global drift problem under long-tailed distribution. Finally, we evaluated FedLC using a highly non-IID and long-tailed experimental setting, comprehensive experiments on several benchmark datasets demonstrated that FedLC achieved superior performance compared with state-of-the-art FL methods, which fully illustrated the effectiveness of

关键词： Training Data privacy Federated learning distributed databases Collaboration Data models Calibration

来源：评论

学校读者我要写书评

暂无评论

An Efficient Scheduling Algorithm for distributed Heterogeneous Systems with Task Duplication Allowed

An Efficient Scheduling Algorithm for Distributed Heterogene...

引用

IEEE International conference on Big Data and Cloud Computing (BdCloud)

作者： Hao Shi Yixiang Chen Jinyi Xu Software Engineering Institute East China Normal University Shanghai China MOE Engineering Research Center for Software/Hardware Co-Design Technology and Application Shanghai China

Due to the restriction of data transmission speed and bandwidth between processing units, communication delays are significantly impacting the complete time of applications in distributed heterogeneous systems. Most of the existing methods considering communication delays are approximate. The performance of solutions given by these heuristic methods are depends on the input instance. Another common choice is optimization methods, such as genetic algorithms and branch-and-bound search. However, these algorithms solve general situations with low efficiency. This paper proposes an efficient scheduling algorithm with task duplication allowed, called DHSA. Task duplication allows tasks executing multiple times to trade extra computation for communication delays, which is widely used in parallel system scheduling approximately. To address the complexity explosion caused by allowing all-task duplication, we decompose the scheduling into allocation and ordering stages and optimize them separately with solving optimization model and slack-interval aware ordering approach. Extensive experiments show that DHSA gives more optimal solutions within feasible time comparing to two state-of-the-art methods, particularly for applications with medium to high communication load.

关键词： Schedules Scheduling algorithms Computational modeling Optimization methods Scheduling Delays Complexity theory

来源：评论

学校读者我要写书评

暂无评论

Implementation of a Round Robin processing Element for Deep Learning Accelerator 17

Implementation of a Round Robin Processing Element for Deep ...

引用

17th International SoC Design conference (ISOCC)

作者： Lee, Eunchong Lee, Yongseok Lee, Sang-Seol Choi, Byoung-Ho Korea Elect Technol Inst Intelligent Image Proc Res Ctr Gyeonggi Do South Korea

ISBN: (纸本)9781728183312

The deep learning acceleration hardware performance is greatly affected by processing Elements (PEs). In order to apply deep learning accelerators to mobile devices, optimized PE must be designed as ASIC. To improve the performance of PE, we focused on methods of minimizing external memory access and parallelization. As a result, a deep learning accelerator architecture consisting of 512 PEs in parallel is proposed and the results of FPGA implementation is presented.

关键词： Deep Learning processing Element FPGA

来源：评论

学校读者我要写书评

暂无评论

parallel Interactive Networks for Multi-Domain Dialogue State Generation

Parallel Interactive Networks for Multi-Domain Dialogue Stat...

引用

conference on Empirical methods in Natural Language processing (EMNLP)

作者： Chen, Junfan Zhang, Richong Mao, Yongyi Xu, Jie Beihang Univ BDBC Beijing Peoples R China Beihang Univ SKLSDE Beijing Peoples R China Univ Ottawa Sch EECS Ottawa ON Canada Univ Leeds Leeds England

ISBN: (纸本)9781952148606

The dependencies between system and user utterances in the same turn and across different turns are not fully considered in existing multidomain dialogue state tracking (MDST) models. In this study, we argue that the incorporation of these dependencies is crucial for the design of MDST and propose parallel Interactive Networks (PIN) to model these dependencies. Specifically, we integrate an interactive encoder to jointly model the in-turn dependencies and cross-turn dependencies. The slot-level context is introduced to extract more expressive features for different slots. And a distributed copy mechanism is utilized to selectively copy words from historical system utterances or historical user utterances. Empirical studies demonstrated the superiority of the proposed PIN model.

关键词：

来源：评论

学校读者我要写书评

暂无评论

Research on Design of High-Performance Graph Computing System Based on CPU Heterogeneous Architecture

Research on Design of High-Performance Graph Computing Syste...

引用

Electronic Information Technology and Smart Agriculture (ICEITSA), International conference on

作者： Yanjiang Li King's College London London UK

ISBN: (纸本)9781665413015

High performance computing is also known as supercomputing, mainly refers to more data in a short time to help the computer system to complete more computing tasks. Cloud computing, as a parallel distributed computing application concept proposed by the current technological innovation, will use the network to connect a large number of computing resources in practical application, thus forming a computing resource pool with stronger computing capacity and parallel processing capacity, so as to bring more superior computing services to users. From the perspective of overall development, cloud computing, as a new processing mode proposed to achieve high performance computing, is very novel in both application scope and promotion form, which plays a positive role in the future research of computer technology. Therefore, based on the understanding of high performance computing research significance and the central processing unit (CPU) implementation on the basis of the advantages of high performance computing, start from a computer unified device architecture (CUDA) for fast image processing system, in-depth studies are proving CPU heterogeneous architecture as the core of high performance computing system has a certain application advantages.

关键词： Technological innovation Cloud computing High performance computing image processing Graphics processing units Weather forecasting Computer architecture

来源：评论

学校读者我要写书评

暂无评论

Unsupervised Learning-Based Depth Estimation-Aided Visual SLAM Approach

引用

CIRCUITS SYSTEMS AND SIGNAL processing 2020年第2期39卷 543-570页

作者： Geng, Mingyang Shang, Suning Ding, Bo Wang, Huaimin Zhang, Pengfei Natl Univ Def Technol Coll Comp Natl Key Lab Parallel & Distributed Proc Changsha 410073 Peoples R China

Simultaneous localization and map construction (SLAM) tasks have been proven to benefit greatly from the depth information of the environment. In this paper, we first present an unsupervised end-to-end learning framework for the task of monocular depth and camera motion estimation from video sequences. The difference between our work and the existing unsupervised methods is that we not only use image reconstruction for supervising but also exploit the pose estimation method used in traditional SLAM approaches to enhance the supervised signal and add extra training constraints for the task of monocular depth and camera motion estimation. Furthermore, we successfully exploit our unsupervised learning framework to assist the traditional ORB-SLAM system when the initialization module of ORB-SLAM method could not match enough features. Qualitative and quantitative experiments have shown that our unsupervised learning framework performs the depth estimation task superior to the supervised methods and outperforms the previous state-of-the-art unsupervised approach by 13.5% on KITTI dataset. For the pose estimation task, our method performs comparably to the supervised methods that use ground-truth pose data for training. Besides, our unsupervised learning framework can significantly accelerate the initialization process of the traditional ORB-SLAM system and effectively improve the accuracy of environmental mapping in strong lighting and weak texture scenes.

关键词： Monocular depth estimation Pose estimation Unsupervised learning Visual SLAM system

来源：评论

学校读者我要写书评

暂无评论

DrugProtKGE: Weakly Supervised Knowledge Graph Embedding for Highly-Effective Drug-Protein Interaction Representation

DrugProtKGE: Weakly Supervised Knowledge Graph Embedding for...

引用

IEEE International conference on Bioinformatics and Biomedicine (BIBM)

作者： Yanlong Qiu Siqi Wang Xi Yang Xinyuan Qiu Chengkun Wu Yingbo Cui Canqun Yang Institute for Quantum Information & State Key Laboratory of High-Performance Computing College of Computer Science National University of Defense Technology Changsha P. R. China National Supercomputer Center in Tianjin Tianjin P. R. China National Key Laboratory of Parallel and Distributed Computing College of Computer Science National University of Defense Technology Changsha P. R. China Department of Biology and Chemistry College of Science National University of Defense Technology Changsha P. R. China

With the exponential growth of biomedical knowledge in unstructured text repositories such as PubMed, it is imminent to establish a knowledge graph-style, efficient searchable and targeted database that can support the need of information retrieval from researchers and clinicians. To mine knowledge from graph databases, most previous methods view a triple in a graph (see Fig. 1) as the basic processing unit and embed the triplet element (i.e. drugs/chemicals, proteins/genes and their interaction) as separated embedding matrices, which cannot capture the semantic correlation among triple elements. To remedy the loss of semantic correlation caused by disjoint embeddings, we propose a novel approach to learn triple embeddings by combining entities and interactions into a unified representation. Furthermore, traditional methods usually learn triple embeddings from scratch, which cannot take advantage of the rich domain knowledge embedded in pre-trained models, and is also another significant reason for the fact that they cannot distinguish the differences implied by the same entity in the multi-interaction triples. In this paper, we propose a novel fine-tuning based approach to learn better triple embeddings by creating weakly supervised signals from pre-trained knowledge graph embeddings. The method automatically samples triples from knowledge graphs and estimates their pairwise similarity from pre-trained embedding models. The triples are then fed pairwise into a Siamese-like neural architecture, where the triple representation is fine-tuned in the manner bootstrapped by triple similarity scores. Finally, we demonstrate that triple embeddings learned with our method can be readily applied to several downstream applications (e.g. triple classification and triple clustering). We evaluated the proposed method on two open-source drug-protein knowledge graphs constructed from PubMed abstracts, as provided by BioCreative. Our method achieves consistent improvement in both t

关键词：

来源：评论

学校读者我要写书评

暂无评论

Tracking Multi-objects with Anchor-Free Siamese Network 5th

Tracking Multi-objects with Anchor-Free Siamese Network

引用

5th IFIP TC 12 International conference on Intelligence Science, ICIS 2022

作者： Hui, Bingyu Feng, Jie Yao, Quanhe Gu, Jing Jiao, Licheng Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education Xidian University Xi’an710071 China

ISBN: (纸本)9783031149023

Single object tracker based on siamese neural network have become one of the most popular frameworks in this field for its strong discrimination ability and high efficiency. However, when the task switch to multi-object tracking, the development of siamese-network based tracking methods is limited by the huge calculation cost comes from repeatedly feature extract and excessive predefined anchors. To solve these problems, we propose a siamese box adaptive multi-object tracking (SiamBAN-MOT) method with parallel detector and siamese tracker branch, which implementation in an anchor-free manner. Firstly, ResNet-50 is constructed to extract shared features of the current frame. Then, we design a siamese specific feature pyramid networks (S-FPN) to fuse the multi-scale features, which improves detection and tracking performance with the anchor-free architecture. To alleviate the duplicate feature extraction, RoI align is operated to extract all trajectories’ template feature and search region feature in a single feature map at once. After that, anchor-free based Siamese tracking network outputs the tracking result of each trajectory according to its template and search region feature. Meanwhile, current frame’s detection results are obtained from the detector for the target association. Finally, a simple novel IOU-matching scheme is designed to map the tracking results to the detection results so as to refine the tracking results and suppress the drifts caused by siamese tracking network. Through experimental verification, our method achieves competitive results on MOT17. © 2022, IFIP International Federation for Information processing.

关键词： Feature extraction

来源：评论

学校读者我要写书评

暂无评论

image defogging improvement under non-uniform fog concentration

Image defogging improvement under non-uniform fog concentrat...

引用

2020 International conference on Engineering Management of Communication and Technology, EMCTECH 2020

作者： Chen, Dapeng College of Electronic and Information Engineering Shandong University of Science and Technology Qingdao China

ISBN: (纸本)9781665404488

Atmospheric scattering model is the main method of image defogging. Aiming at the defects of the existing defogging algorithms in different aspects, this paper puts forward some improvement measures based on the analysis of the defects of the existing defogging methods. In view of the irrationality of estimation of atmospheric light and atmospheric scattering coefficient, a method of locating atmospheric light region based on edge extraction and a method of selecting atmospheric scattering coefficient based on fog concentration are proposed. The estimation of depth of field and transmittance is improved. Finally, adaptive contrast enhancement is carried out for foggy image. The effectiveness of the proposed method is verified by comparative experiments. © 2020 IEEE.

关键词： image enhancement

来源：评论

学校读者我要写书评

暂无评论

Poetic Expression Through Scenery: Sentimental Chinese Classical Poetry Generation from images 26th

Poetic Expression Through Scenery: Sentimental Chinese Class...

引用

26th International conference on Database Systems for Advanced Applications (DASFAA)

作者： Li, Haotian Zhu, Jiatao Cao, Sichen Li, Xiangyu Zeng, Jiajun Wang, Peng Southeast Univ Sch Artificial Intelligence Nanjing Peoples R China Southeast Univ Sch Comp Sci & Engn Nanjing Peoples R China Southeast Univ Sch Cyber Sci & Engn Nanjing Peoples R China

ISBN: (纸本)9783030731960;9783030731977

Most Chinese poetry generation methods only accept texts or user-specified words as input, which contradicts with the fact that ancient Chinese wrote poems inspired by visions, hearings and feelings. This paper proposes a method to generate sentimental Chinese classical poetry automatically from images based on convolutional neural networks and the language model. First, our method extracts visual information from the image and maps it to initial keywords by two parallel image classification models, then filters and extends these keywords to form a keywords set which is finally input into the poetry generation model to generate poems of different genres. A bi-directional generation algorithm and two fluency checkers are proposed to ensure the diversity and quality of generated poems, respectively. Besides, we constrain the range of optional keywords and define three sentiment-related keywords dictionary to avoid modern words that lead to incoherent content as well as ensure the emotional consistency with given images. Both human and automatic evaluation results demonstrate that our method can reach a better performance on quality and diversity of generated poems.

关键词： Information extraction and summarization Poetry generation Vision-driven modelling Natural language processing

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：