检索结果-内蒙古大学图书馆

Speculative Decoding with CTC-based Draft Model for LLM Inference Acceleration

学校读者我要写书评

暂无评论

arXiv 2024年

作者： Wen, Zhuofan Gui, Shangtong Feng, Yang Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences China State Key Lab of Processors Institute of Computing Technology Chinese Academy of Sciences China Key Laboratory of Al Safety Chinese Academy of Sciences China University of Chinese Academy of Sciences Beijing China

Inference acceleration of large language models (LLMs) has been put forward in many application scenarios and speculative decoding has shown its advantage in addressing inference acceleration. Speculative decoding usually introduces a draft model to assist the base LLM where the draft model produces drafts and the base LLM verifies the draft for acceptance or rejection. In this framework, the final inference speed is decided by the decoding speed of the draft model and the acceptance rate of the draft provided by the draft model. Currently the widely used draft models usually generate draft tokens for the next several positions in a non-autoregressive way without considering the correlations between draft tokens. Therefore, it has a high decoding speed but an unsatisfactory acceptance rate. In this paper, we focus on how to improve the performance of the draft model and aim to accelerate inference via a high acceptance rate. To this end, we propose a CTC-based draft model which strengthens the correlations between draft tokens during the draft phase, thereby generating higher-quality draft candidate sequences. Experiment results show that compared to strong baselines, the proposed method can achieve a higher acceptance rate and hence a faster inference speed. Copyright © 2024, The Authors. All rights reserved.

关键词： Decoding

A Cross-direction Task Decoupling Network for Small Logo Detection

学校读者我要写书评

暂无评论

arXiv 2023年

作者： Hou, Sujuan Li, Xingzhuo Min, Weiqing Li, Jiacheng Wang, Jing Zheng, Yuanjie Jiang, Shuqiang School of Information Science and Engineering Shandong Normal University Jinan China Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences Beijing China

关键词： Convolution

Optically transparent transmissive orbital angular momentum (OAM) metasurface with high transmittance and broadband characteristics

学校读者我要写书评

暂无评论

Optics Express 2025年第12期33卷 24385-24395页

作者： Yong Cai Yuying Dai Yue Ma Gang Wang Guanghui Xu Xingang Ren Hui Wang Xiaohu Zhao Xianliang Wu Information Materials and Intelligent Sensing Laboratory of Anhui Province Anhui University Hefei 230039 China Key Laboratory of Electromagnetic Environmental Sensing of Anhui Higher Education Institutes Anhui University Hefei 230039 China wang_toward@*** ghxu86@*** Key Laboratory of Intelligent Computing & Signal Processing Ministry of Education Institute of Physical Science and Information Technology Anhui University Hefei 230039 China xgren@*** Hefei Normal University Hefei 230009 China

This work proposes a high-performance broadband orbital angular momentum (OAM) beam-generating metasurface based on optically transparent media. A broadband polarization-converting transmissive metasurface, which operates in the 18 - 24 (28.57% fractional bandwidth) GHz range, has been realized by collaboratively designing dual C-split ring resonator structures and bidirectional lattice-patterned surface networks. The array structure demonstrates typical helical phase distribution features in the K-band, and experimental results verify simulation design. Innovatively employing a composite structure of quartz glass substrate and patterned conductive thin films, the design achieves compatible optimization of 77% visible-light transmittance and microwave broadband performance. This approach breaks through the mutually exclusive limitations of traditional metasurfaces’ optical/electromagnetic properties, providing a highly integrated solution for transparent antenna radomes in satellite communication systems and multi-mode beam multiplexing applications.

关键词： Orbital angular momentum multiplexing Phase plates Ring resonators Spatial light modulators Spiral phase Thin films

A Cell Condition-Sensitive Frequency Segmentation Method Based on the Sub-Band Instantaneous Energy Spectrum of Aluminum Electrolysis Cell Voltage

学校读者我要写书评

暂无评论

Engineering 2021年第9期7卷 1282-1292页

作者： Zhaohui Zeng Weihua Gui Xiaofang Chen Yongfang Xie Hongliang Zhang Yubo Sun School of Automation Central South UniversityChangsha 410083China Key Laboratory of Intelligent Computing&Information Processing Ministry of EducationXiangtan UniversityXiangtan 411105China School of Metallurgy and Environment Central South UniversityChangsha 410083China

Cell voltage is a widely used signal that can be measured online from an industrial aluminum electrolysis cell.A variety of parameters for the analysis and control of industrial cells are calculated using the cell *** this paper,the frequency segmentation of cell voltage is used as the basis for designing filters to obtain these *** on the qualitative analysis of the cell voltage,the sub-band instantaneous energy spectrum(SIEP)is first proposed,which is then used to quantitatively represent the characteristics of the designated frequency bands of the cell voltage under various cell ***,a cell condition-sensitive frequency segmentation method is *** proposed frequency segmentation method divides the effective frequency band into the[0,0.001]Hz band of lowfrequency signals and the[0.001,0.050]Hz band of low-frequency noise,and subdivides the lowfrequency noise into the[0.001,0.010]Hz band of metal pad abnormal rolling and the[0.01,0.05]Hz band of sub-low-frequency *** with the instantaneous energy spectrum based on empirical mode decomposition,the SIEP more finely represents the law of energy change with time in any designated frequency band within the effective frequency band of the cell *** proposed frequency segmentation method is more sensitive to cell condition changes and can obtain more elaborate details of online cell condition information,thus providing a more reliable and accurate online basis for cell condition monitoring and control decisions.

关键词： Sub-band instantaneous energy spectrum Cell condition-sensitive frequency band Frequency segmentation Metal pad abnormal rolling Aluminum electrolysis

SafeCFG: Redirecting Harmful Classifier-Free Guidance for Safe Generation

学校读者我要写书评

暂无评论

arXiv 2024年

作者： Pan, Jiadong Gao, Hongcheng Li, Liang Zha, Zheng-Jun Huang, Qingming Luo, Jiebo Key Laboratory of Intelligent Information Processing Institute of Computing Technology CAS China University of Chinese Academy of Sciences China University of Science and Technology of China China University of Rochester United States

Diffusion models (DMs) have demonstrated exceptional performance in text-to-image (T2I) tasks, leading to their widespread use. With the introduction of classifier-free guidance (CFG), the quality of images generated by DMs is improved. However, DMs can generate more harmful images by maliciously guiding the image generation process through CFG. Some safe guidance methods aim to mitigate the risk of generating harmful images but often reduce the quality of clean image generation. To address this issue, we introduce the Harmful Guidance Redirector (HGR), which redirects harmful CFG direction while preserving clean CFG direction during image generation, transforming CFG into SafeCFG and achieving high safety and quality generation. We train HGR to redirect multiple harmful CFG directions simultaneously, demonstrating its ability to eliminate various harmful elements while preserving high-quality generation. Additionally, we find that HGR can detect image harmfulness, allowing for unsupervised fine-tuning of safe diffusion models without pre-defined clean or harmful labels. Experimental results show that by incorporating HGR, images generated by diffusion models achieve both high quality and strong safety, and safe DMs trained through unsupervised methods according to the harmfulness detected by HGR also exhibit good safety performance. The codes will be publicly available. Copyright © 2024, The Authors. All rights reserved.

关键词：

Multi-Layer Visual Perception for No-Reference Image Quality Assessment 6

学校读者我要写书评

暂无评论

Multi-Layer Visual Perception for No-Reference Image Quality...

6th IEEE International Conference on Electronic information and Communication Technology, ICEICT 2023

作者： Qi, Junwei Wang, Yingzhen Wang, Qingchun Li, Yingsong College of Computer Science and Technology Harbin Engineering University Harbin150001 China College of Information and Communication Engineering Harbin Engineering University Harbin150001 China Shanghai Yanding Information Technology Co. Ltd. Shanghai201210 China Anhui University Key Laboratory of Intelligent Computing and Signal Processing Ministry of Education Anhui Hefei230601 China

ISBN: (纸本)9798350399059

In this work, we propose a no-reference image quality evaluation approach, aiming to solve the problem that the traditional convolutional neural network is insufficient to express the global information of the image. Therefore, in order to make the captured image features have the coherence of global context information, we combine a Transformer structure widely used in NLP (natural language processing) with the traditional CNN to realize the attention mechanism in the human visual perception characteristics. Firstly, the self-attention mechanism of Transformer is used to learn the global representation of the image from the multi-layer features extracted from different levels of CNN. Additionally, to forecast the ultimate image quality score, the global and local attributes are combined. Finally, the proposed algorithm is evaluated on three public datasets commonly used in CSIQ [1], TID2013 [2] and LIVE [3]. Compared with eight traditional no-reference image quality evaluation algorithms and five methods based on deep learning, the proposed algorithm achieves competitive results. © 2023 IEEE.

关键词： Convolutional neural networks

Reinforcement Learning with Token-level Feedback for Controllable Text Generation

学校读者我要写书评

暂无评论

arXiv 2024年

作者： Li, Wendi Wei, Wei Xu, Kaihe Xie, Wenfeng Chen, Dangyang Cheng, Yu Cognitive Computing and Intelligent Information Processing Laboratory School of Computer Science and Technology Huazhong University of Science and Technology China China Ping An Property & Casualty Insurance company of China China The Chinese University of Hong Kong Hong Kong

To meet the requirements of real-world applications, it is essential to control generations of large language models (LLMs). Prior research has tried to introduce reinforcement learning (RL) into controllable text generation while most existing methods suffer from overfitting issues (finetuning-based methods) or semantic collapse (post-processing methods). However, current RL methods are generally guided by coarse-grained (sentence/paragraph-level) feedback, which may lead to suboptimal performance owing to semantic twists or progressions within sentences. To tackle that, we propose a novel reinforcement learning algorithm named TOLE which formulates TOken-LEvel rewards for controllable text generation, and employs a "first-quantize-then-noise" paradigm to enhance the robustness of the RL algorithm. Furthermore, TOLE can be flexibly extended to multiple constraints with little computational expense. Experimental results show that our algorithm can achieve superior performance on both single-attribute and multi-attribute control tasks. We have released our codes at https://***/WindyLee0822/CTG. Copyright © 2024, The Authors. All rights reserved.

关键词： Reinforcement learning

Investigation on a quantum algorithm for linear differential equations

学校读者我要写书评

暂无评论

arXiv 2024年

作者： Dong, Xiaojing Peng, Yizhe Tang, Qili Yang, Yin Yu, Yue Hunan Key Laboratory for Computation and Simulation in Science and Engineering Key Laboratory of Intelligent Computing and Information Processing of Ministry of Education National Center for Applied Mathematics in Hunan School of Mathematics and Computational Science Xiangtan University Hunan Xiangtan411105 China

Ref. [BCOW17] introduced a pioneering quantum approach (coined BCOW algorithm) for solving linear differential equations with optimal error tolerance. Originally designed for a specific class of diagonalizable linear differential equations, the algorithm was extended by Krovi in [Kro23] to encompass broader classes, including non-diagonalizable and even singular matrices. Despite the common misconception, the original algorithm is indeed applicable to non-diagonalizable matrices, with diagonalisation primarily serving for theoretical analyses to establish bounds on condition number and solution error. By leveraging basic estimates from [Kro23], we derive bounds comparable to those outlined in the Krovi algorithm, thereby reinstating the advantages of the BCOW approach. Furthermore, we extend the BCOW algorithm to address time-dependent linear differential equations by transforming non-autonomous systems into higher-dimensional autonomous ones, a technique also applicable for the Krovi algorithm. © 2024, CC BY-NC-ND.

关键词： Consensus algorithm

An Enhanced Deep Reinforcement Learning-Based Global Router for VLSI Design

学校读者我要写书评

暂无评论

Wireless Communications and Mobile computing 2023年第1期2023卷

作者： Xu, Saijuan Yang, Liliang Liu, Genggeng Department of Information Engineering Fujian Business University Fuzhou China College of Computer and Data Science Fuzhou University Fuzhou China Key Laboratory of Network Computing and Intelligent Information Processing Fuzhou University Fuzhou China

Global routing is a crucial step in the design of Very Large-Scale Integration (VLSI) circuits. However, most of the existing methods are heuristic algorithms, which cannot conjointly optimize the subproblems of global routing, resulting in congestion and overflow. In response to this challenge, an enhanced Deep Reinforcement Learning- (DRL-) based global router has been proposed, which comprises the following effective strategies. First, to avoid the overestimation problem generated by Q-learning, the proposed global router adopts the Double Deep Q-Network (DDQN) model. The DDQN-based global router has better performance in wire length optimization and convergence. Second, to avoid the agent from learning redundant information, an action elimination method is added to the action selection part, which significantly enhances the convergence performance of the training process. Third, to avoid the unfair allocation problem of routing resources in serial training, concurrent training is proposed to enhance the routability. Fourth, to reduce wire length and disperse routing resources, a new reward function is proposed to guide the agent to learn better routing solutions regarding wire length and congestion standard deviation. Experimental results demonstrate that the proposed algorithm outperforms others in several important performance metrics, including wire length, convergence performance, routability, and congestion standard deviation. In conclusion, the proposed enhanced DRL-based global router is a promising approach for solving the global routing problem in VLSI design, which can achieve superior performance compared to the heuristic method and DRL-based global router. © 2023 Saijuan Xu et al.

关键词： Reinforcement learning