检索结果-内蒙古大学图书馆

AI computing Systems for Large Language Models Training

计算机科学技术学报（英文版） 2025年第1期40卷 6-41页

作者：张振兴文渊博吕涵祺刘畅张蕊李夏青王超杜子东郭崎李玲周学海陈云霁 1School of Computer Science and Technology University of Science and Technology of China Hefei 230026 China 2State Key Laboratory of Processors Institute of Computing Technology Chinese Academy of Sciences Beijing 100190 China 3Cambricon Technologies Beijing 100191 China 4Shanghai Innovation Center for Processor Technologies Shanghai 201210 China 5Intelligent Software Research Center Institute of Software Chinese Academy of Sciences Beijing 100190 China 6University of Chinese Academy of Sciences Beijing 101408 China

In this paper,we present a comprehensive overview of artificial intelligence(AI)computing systems for large language models(LLMs)*** rapid advancement of LLMs in recent years,coupled with the widespread adoption of algorithms and applications such as BERT,ChatGPT,and DeepSeek,has sparked significant interest in this *** classify LLMs into encoder-only,encoder-decoder,and decoder-only models,and briefly analyze their training and infer-ence processes to emphasize their substantial need for computational *** operations depend heavily on AI-specific accelerators like GPUs(graphics processing units),TPUs(tensor processing units),and MLUs(machine learning units).However,as the gap widens between the increasing complexity of LLMs and the current capabilities of accelerators,it becomes essential to adopt heterogeneous computing systems optimized for distributed environments to manage the growing computational and memory requirements of *** delve into the execution and scheduling of LLM algo-rithms,underlining the critical role of distributed computing strategies,memory management enhancements,and boosting computational *** paper clarifies the complex relationship between algorithm design,hardware infrastructure,and software optimization,and provides an in-depth understanding of both the software and hardware infrastructure sup-porting LLMs training,offering insights into the challenges and potential avenues for future development and deployment.

关键词： artificial intelligence(AI)chip large language model(LLM) AI computing system accelerator

来源：评论

学校读者我要写书评

暂无评论

Aware: Adaptive Distributed Training with Computation, Communication and Position Awareness for Deep Learning Model 24

Aware: Adaptive Distributed Training with Computation, Commu...

引用

24th IEEE International Conference on High Performance computing and Communications, 8th IEEE International Conference on Data Science and Systems, 20th IEEE International Conference on Smart City and 8th IEEE International Conference on Dependability in Sensor, Cloud and Big Data Systems and Application, HPCC/DSS/SmartCity/DependSys 2022

作者： Zeng, Yan Yi, Guangzheng Yin, Yuyu Wu, Jiyang Xue, Meiting Zhang, Jilin Wan, Jian Zhang, Yunquan School of Computer Science and Technology Hangzhou Dianzi University Hangzhou China School of Cyberspace Hangzhou Dianzi University Hangzhou China Institute of Computing Technology Chinese Academy of Sciences State Key Laboratory of Computer Architecture Beijing China

ISBN: (纸本)9798350319934

The accuracy of the neural networks can usually be improved by increasing the size of the dataset and the layers or operators of the network, as it has strong composability. But, it makes a challenge to train these models efficiently, due to the limited resources of a single device, such as memory, computational resource and so on. For this challenge, it has become an inevitable trend to automatically implement model parallelism across multiple devices. This paper proposes Aware, an adaptive distributed parallel training method, to search distributed parallel strategy automatically for deep learning model. It firstly proposes an operator fusion strategy with computation and communication awareness to simplify the computational graph of deep learning model. And then, it introduces position-aware graph embedding algorithm to extract the structural features of models, which can make the searched parallel strategies to transplant to other networks with similar structures. On this basis, we use reinforcement learning algorithm to search distributed parallel strategy automatically. This paper makes experiments with neural networks such as Inception V3, NMT, GNM, NasNet and ResNet, and uses CIFAR10 and PTB datasets to compare the two methods of Hierarchical and Placeto. The results show that compared with Placeto, Aware achieves up to 5% reductions in runtime. Aware achieves more than 8% reductions in runtime compared with Hierarchical. Moreover, it has better transplantation and generalization capabilities, and supports pre-training for large-scale network parallel strategy search and accelerates convergence. © 2022 IEEE.

关键词： Deep learning

来源：评论

学校读者我要写书评

暂无评论

Thematic Event Extraction Based on Event-Related Sentence Detection 1

Thematic Event Extraction Based on Event-Related Sentence De...

引用

1st IEEE International Conference on Medical Artificial Intelligence, MedAI 2023

作者： Zhang, Xing Guo, Yucan Jin, Xiaolong Liang, Haozhe Guan, Saiping School of Computer Science and Technology University of Chinese Academy of Sciences China Institute of Computing Technology Chinese Academy of Sciences Cas Key Laboratory of Network Data Science and Technology China Institute of Systems Engineering Academy of Military Science Australia

ISBN: (纸本)9798350358780

Existing works in event extraction typically extract event arguments within the sentence scope. However, besides the sentence level, events may also be naturally presented at the document level. A document-level event usually reflects, to some extent, the theme (i.e., the main content) of the document (e.g., electronic medical records and news articles), which is thus referred to as the thematic event. Thematic Event Extraction (TEE) aims to extract the arguments of thematic events. TEE faces a major challenge, i.e., the sparsity and dispersion of arguments, which means that the arguments of a thematic event are dispersed in different sentences of the document. To overcome this challenge, we propose an Event-related Sentence Detection based TEE model, called ESDTEE, which first detects the sentences related to the thematic event and then extracts the arguments only within these detected sentences using existing models. Extensive experiments with comprehensive analyses demonstrate the effectiveness of ESDTEE. © 2023 IEEE.

关键词： Medical computing

来源：评论

学校读者我要写书评

暂无评论

Divide-Solve-Combine: An Interpretable and Accurate Prompting Framework for Zero-shot Multi-Intent Detection 39

Divide-Solve-Combine: An Interpretable and Accurate Promptin...

引用

39th Annual AAAI Conference on Artificial Intelligence, AAAI 2025

作者： Qin, Libo Chen, Qiguang Zhou, Jingxuan Wang, Jin Fei, Hao Che, Wanxiang Li, Min School of Computer Science and Engineering Central South University China Key Laboratory of Data Intelligence and Advanced Computing in Provincial Universities Soochow University China Research Center for Social Computing and Information Retrieval Harbin Institute of Technology China Yunnan University China National University of Singapore Singapore

ISBN: (纸本)157735897X

Zero-shot multi-intent detection is capable of capturing multiple intents within a single utterance without any training data, which gains increasing attention. Building on the success of large language models (LLM), dominant approaches in the literature explore prompting techniques to enable zero-shot multi-intent detection. While significant advancements have been witnessed, the existing prompting approaches still face two major issues: lacking explicit reasoning and lacking interpretability. Therefore, in this paper, we introduce a Divide-Solve-Combine Prompting (DSCP) to address the above issues. Specifically, DSCP explicitly decomposes multi-intent detection into three components including (1) single-intent division prompting is utilized to decompose an input query into distinct sub-sentences, each containing a single intent;(2) intent-by-intent solution prompting is applied to solve each sub-sentence recurrently;and (3) multi-intent combination prompting is employed for combining each sub-sentence result to obtain the final multi-intent result. By decomposition, DSCP allows the model to track the explicit reasoning process and improve the interpretability. In addition, we propose an interactive divide-solve-combine prompting (Inter-DSCP) to naturally capture the interaction capabilities of large language models. Experimental results on two standard multi-intent benchmarks (i.e., MIXATIS and MIXSNIPS) reveal that both DSCP and Inter-DSCP obtain substantial improvements over baselines, achieving superior performance and higher interpretability. Copyright © 2025, Association for the Advancement of Artificial Intelligence (***). All rights reserved.

关键词：

来源：评论

学校读者我要写书评

暂无评论

A Design of Fetch Target Buffer Implemented on XiangShan Processor

A Design of Fetch Target Buffer Implemented on XiangShan Pro...

引用

2022 International Conference on Cloud computing, Internet of Things, and computer Applications, CICA 2022

作者： Zou, Jiangrui Tang, Dan Cai, Ye Li, Zusong School of Computer and Software Shenzhen University Shenzhen China State Key Laboratory of Computer Architecture Institute of Computing Technology Chinese Academy of Sciences Beijing China Beijing Vcore Technology Co. Ltd. Beijing China

ISBN: (纸本)9781510656673

In the design of modern high-performance processors, branch prediction is a very important component, and good branch prediction can greatly improve the performance of the processor. And Branch Target Buffer (BTB) is an important part of branch prediction. BTB is used to store information about branch instructions. Different from storing each branch as a unit, Glenn et al. proposed a BTB that stores branch information in a unit of fetch block, which they call Fetch Target Buffer (FTB). This design can optimize the timing by limiting the branch prediction width. But they did not give specific implementation details. In this paper, we propose a specific FTB implementation, and it is used in the second version architecture of XiangShan high-performance processor. First, we give the fields in the FTB entry in this design and their corresponding functions. Most BTB entries need to store the following information: branch tag, branch type, target address. In addition to these fields in this paper, FTB will also store other information for managing entries and optimizing branch prediction performance. Secondly, we proposed the management strategy of FTB entries, when to add, modify, and replace. After all the instructions in a fetch block are committed, we need to decide whether this fetch block needs to be written to FTB, and which way to write, how to avoid the problem of multiple hits. Under the 14nm process, compared to the 1.8GHz of the first version of the architecture, the frequency of the new architecture using FTB can be increased to 2GHz. © 2022 SPIE.

关键词： architecture

来源：评论

学校读者我要写书评

暂无评论

Co-ViSu: a Video Super-Resolution Accelerator Exploiting Codec Information Reuse

Co-ViSu: a Video Super-Resolution Accelerator Exploiting Cod...

引用

International Conference on Field Programmable Logic and Applications

作者： Haishuang Fan Jingya Wu Wenyan Lu Xiaowei Li Guihai Yan State Key Laboratory of Computer Architecture Institute of Computing Technology Chinese Academy of Sciences University of Chinese Academy of Sciences

High-resolution (HR) videos have become popular due to the widespread adoption of high-definition displays. Super-resolution (SR) techniques aim to recover HR frames from low-resolution (LR) frames. Recently, deep neural network (DNN)-based SR methods have achieved superior quality compared to traditional methods. FPGA-based SR accelerators have been proposed to optimize performance and power efficiency. However, most accelerators tailored for video SR only accept uncompressed video frames and operate per-frame DNN inference, ignoring the temporal-spatial information in compressed video bitstreams. In contrast, we observe that non-key frames can be directly constructed using codec information and HR key-frames, saving a significant amount of DNN computing. In this paper, we propose a novel compressed video SR flow and a specific FPGA accelerator called Co-ViSu that integrates decoder, SR, and encoder engines. Co-ViSu exploits codec information reuse scheme to skip non-key frame decoding, avoid complex DNN computation and speed up encoding. Our experimental results show that Co-ViSu achieves 3.6x to 9.4x performance, 4.2x energy efficiency gain with only 0.17dB quality loss compared to the traditional flow, and 2.1x throughput than state-of-the-art.

关键词：

来源：评论

学校读者我要写书评

暂无评论

Multi-Grained Feature Pruning for Video-Based Human Pose Estimation

Multi-Grained Feature Pruning for Video-Based Human Pose Est...

引用

2025 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2025

作者： Wang, Zhigang Fan, Shaojing Liu, Zhenguang Wu, Zheqi Wu, Sifan Jiao, Yingying College of Computer Science and Technology Zhejiang Gongshang University Hangzhou China School of Computing National University of Singapore Singapore The State Key Laboratory of Blockchain and Data Security Zhejiang University Hangzhou China Institute of Blockchain and Data Security Hangzhou China College of Computer Science and Technology Jilin University Changchun China College of Computer Science and Technology Zhejiang University of Technology Hangzhou China

Human pose estimation, with its broad applications in action recognition and motion capture, has experienced significant advancements. However, current Transformer-based methods for video pose estimation often face challenges in managing redundant temporal information and achieving fine-grained perception because they only focus on processing low-resolution features. To address these challenges, we propose a novel multi-scale resolution framework that encodes spatio-temporal representations at varying granularities and executes fine-grained perception compensation. Furthermore, we employ a density peaks clustering method to dynamically identify and prioritize tokens that offer important semantic information. This strategy effectively prunes redundant feature tokens, especially those arising from multi-frame features, thereby optimizing computational efficiency without sacrificing semantic richness. Empirically, it sets new benchmarks for both performance and efficiency on three large-scale datasets. Our method achieves a 93.8% improvement in inference speed compared to the baseline, while also enhancing pose estimation accuracy, reaching 87.4 mAP on the PoseTrack2017 dataset. © 2025 IEEE.

关键词： Large datasets

来源：评论

学校读者我要写书评

暂无评论

SMGC-SBERT: A Multi-Feature Fusion Chinese Short Text Similarity Computation Model Based on Optimised SBERT

SMGC-SBERT: A Multi-Feature Fusion Chinese Short Text Simila...

引用

2024 IEEE International Conference on Systems, Man, and Cybernetics, SMC 2024

作者： Zhao, Shuo Gu, Qiliang Zhang, Jianqiang Song, Gongpeng Lu, Qin Key Laboratory of Computing Power Network and Information Security Ministry of Education Shandong Computer Science Center Jinan China Shandong Engineering Research Center of Big Data Applied Technology Faculty of Computer Science and Technology Jinan China Shandong Provincial Key Laboratory of Computer Networks Shandong Fundamental Research Center for Computer Science Jinan China Shandong Branch of China Mobile Communication Group Design Institute Co. Jinan China

ISBN: (纸本)9781665410205

Chinese short text similarity computation stands as a pivotal task within natural language processing, garnering significant attention. However, existing models grapple with limitations in handling intricate semantic relationships, such as the challenge of discerning subtle semantic nuances in text, inadequacies in effectively integrating diverse levels of semantic information, and the struggle to capture polysemous meanings accurately. In addressing these issues, this paper introduces an innovative Chinese short text similarity computation model, SMGC-SBERT. This model addresses shortcomings of the existing models by employing a multi-module fusion strategy, thereby enabling a more precise measurement of semantic similarity between texts. Primarily, the model incorporates SAT embedding to acquire phrase-level semantic information and leverages the MS-BERT model to encode text, improving the model's comprehension of textual polysemy and obtaining richer semantic representation. Then, the fusion of module features, including multi-branch convolutional networks and mix pooling, enables the extraction of textual features from varied levels, bolstering the model's representational capacity. Additionally, to further reduce overfitting risks while improving accuracy and other performance, a multi-layer feature adjustment network is utilized for short text similarity calculation. The final resultant experimental findings showcase the superiority of the SMGC-SBERT model over other neural network models, demonstrating significant advancements across both Chinese-SNLI and CCKS2018_Task3 Chinese short text datasets. © 2024 IEEE.

关键词： Semantics

来源：评论

学校读者我要写书评

暂无评论

A Battery Swapping Service Management System Reliability-Based Design Concerning Station Information Delay

引用

IEEE Transactions on Consumer Electronics 2025年第1期71卷 1063-1078页

作者： Li, Xinyu Cao, Yue Hu, Ziyi Zhang, Xu Liu, Zhi Lei, Xinhai Wuhan University School of Cyber Science and Engineering Wuhan430073 China Wuhan Institute of Technology Hubei Key Laboratory of Intelligent Robot Wuhan430250 China School of Computing Sciences Norwich United Kingdom University of Electro-Communications Department of Computer and Network Engineering Tokyo Japan Wuhan Public Resources Trading Center Wuhan China

Since the long duration associated with traditional plug-in charging modes inevitably causes range anxiety to electric vehicle (EV) drivers, the battery swapping technology has emerged as a promising alternative. It thus becomes imperative to develop a reliable battery swapping service system, efficiently managing the large-scale energy replenishing demands from EVs. Firstly, this paper considers drivers' trip destination to minimize total trip duration, and the expense paid for battery swapping service to save users' service cost. By optimizing the requirement of battery swapping service during peak-price period, the expenditure for energy supplement can be reduced. Secondly, concerning in practice that the transmission delay is a crucial factor impacting the overall network and system reliability, this paper further investigates the impact of station information delay between the global controller and battery swapping stations (BSSs). By characterizing the delay into two types, namely, low and large delays, the scenario based on a known delay type are investigated. Thirdly, delay cases with multiple BSSs are also taken into account and a mixed decision-making method is developed. Finally, simulation results indicate the desirable performance of proposed scheme in satisfying user demand and alleviating impact of station information delay. © 1975-2011 IEEE.

关键词： Decision making

来源：评论

学校读者我要写书评

暂无评论

FERI: Feature Enhancement and Relational Interaction for Image-text Matching 30

FERI: Feature Enhancement and Relational Interaction for Ima...

引用

30th IEEE International Conference on Parallel and Distributed Systems, ICPADS 2024

作者： Zhang, Yu Zhang, Jianqiang Song, Gongpeng Lu, Qin Zhao, Shuo Key Laboratory of Computing Power Network and Information Security Ministry of Education Shandong Computer Science Center Jinan China Shandong Engineering Research Center of Big Data Applied Technology Faculty of Computer Science and Technology Jinan China Shandong Fundamental Research Center for Computer Science Shandong Provincial Key Laboratory of Computer Networks Jinan China Shandong Branch of China Mobile Communication Group Design Institute Co. Jinan China

ISBN: (纸本)9798331515966

Image-text matching is an important problem at the intersection of computer vision and natural language processing. It aims to establish the semantic link between image and text to achieve high-quality semantic alignment between the two modalities. However, the existing methods have the problem that the meaning expressed in the image or the complex narrative in the text cannot be fully understood due to insufficient feature extraction. Moreover, due to the essential modal differences between images and texts, how to effectively and accurately align the semantic contents in images and texts has become the key of research. In order to solve the above problems, this paper proposes a method based on feature enhancement and relationship interaction. When processing images, the proposed method fuses labeled features, region features and location features to represent images. When processing text, a combination of Bi-GRU and self-attention mechanism is used to represent the text. In order to further align the semantic content in images and texts accurately, this paper improves two relational interaction mechanisms by identifying connection relationships and learning association relationships. Thus, the relation enhanced embedding is obtained. Finally, it calculated the similarity of the enhanced embedding to judge the matching degree of the image and text. Extensive experiments on the public datasets Flickr30K and MSCOCO demonstrate the effectiveness of our method. © 2024 IEEE.

关键词： cross-modal retrieval feature enhancement image-text matching relationship interaction

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：