Training deep neural networks(DNNs)requires a significant amount of time and resources to obtain acceptable results,which severely limits its deployment in resource-limited platforms.This paper proposes DarkFPGA,a nov...Training deep neural networks(DNNs)requires a significant amount of time and resources to obtain acceptable results,which severely limits its deployment in resource-limited platforms.This paper proposes DarkFPGA,a novel customizable framework to efficiently accelerate the entire DNN training on a single FPGA platform.First,we explore batch-level parallelism to enable efficient FPGA-based DNN training.Second,we devise a novel hardware architecture optimised by a batch-oriented data pattern and tiling techniques to effectively exploit parallelism.Moreover,an analytical model is developed to determine the optimal design parameters for the DarkFPGA accelerator with respect to a specific network specification and FPGA resource constraints.Our results show that the accelerator is able to perform about 10 times faster than CPU training and about a third of the energy consumption than GPU training using 8-bit integers for training VGG-like networks on the CIFAR dataset for the Maxeler MAX5 platform.展开更多
The concept of legged-robot stability training with a training platform is proposed and a serial-parallel mechanism platform with 6 degrees of freedom is designed for this target. The designed platform is composed of ...The concept of legged-robot stability training with a training platform is proposed and a serial-parallel mechanism platform with 6 degrees of freedom is designed for this target. The designed platform is composed of 4-DOF parallel mechanism with spherical joints and prismatic pairs,and 2-DOF serial mechanism with prismatic pairs. With this design,the platform has advantages of low platform countertop,big workspace,high carrying capacity and high stiffness. On the basis of DOF analysis and computation of space mechanism,weight supporting auxiliary mechanism and raceways-balls supporting mechanism are designed,so as to improve the stiffness of designed large platform and payload capacity of servo motors. And then the whole structure design work of the platform is done. Meanwhile,this paper derives the analytical solutions of forward kinematics, inverse kinematics and inverse dynamics. The error analysis model of position and orientation is established. And then the simulation is done in ADAMS to ensure the correctness and feasibility of this design.展开更多
Railway train energy simulation is an important and popular research topic.Locomotive traction force simulations are a fundamental part of such research.Conventional energy calculation models are not able to consider ...Railway train energy simulation is an important and popular research topic.Locomotive traction force simulations are a fundamental part of such research.Conventional energy calculation models are not able to consider locomotive wheel-rail adhesions,traction adhesion control,and locomotive dynamics.This paper has developed two models to fill this research gap.The first model uses a 2D locomotive model with 27 degrees of freedom and a simplified wheel-rail contact model.The second model uses a 3D locomotive model with 54 degrees of freedom and a fully detailed wheel-rail contact model.Both models were integrated into a longitudinal train dynamics model with the consideration of locomotive adhesion control.Energy consumption simulations using a conventional model(1D model)and the two new models(2D and 3D models)were conducted and compared.The results show that,due to the consideration of wheel-rail adhesion model and traction control in the 3D model,it reports less energy consumption than the 1D model.The maximum difference in energy consumption rate between the 3D model and the 1D model was 12.5%.Due to the consideration of multiple wheel-rail contact points in the 3D model,it reports higher energy consumption than the 2D model.An 8.6%maximum difference in energy consumption rate between the 3D model and the 1D model was reported during curve negotiation.展开更多
超大预训练模型(Pre-trained model,PTM)是人工智能领域近年来迅速崛起的研究方向,在自然语言处理(Natural language processing,NLP)和计算机视觉等多种任务中达到了有史以来的最佳性能,促进了人工智能生成内容(Artificial intelligenc...超大预训练模型(Pre-trained model,PTM)是人工智能领域近年来迅速崛起的研究方向,在自然语言处理(Natural language processing,NLP)和计算机视觉等多种任务中达到了有史以来的最佳性能,促进了人工智能生成内容(Artificial intelligence-generated content,AIGC)的发展和落地.ChatGPT作为当下最火热的PTM,更是以优异的表现获得各界的广泛关注.本文围绕ChatGPT展开.首先概括PTM的基本思想并对其发展历程进行梳理;接着,详细探讨ChatGPT的技术细节,并以平行智能的视角阐述ChatGPT;最后,从技术、范式以及应用等多个方面对PTM的发展趋势进行展望.展开更多
文摘Training deep neural networks(DNNs)requires a significant amount of time and resources to obtain acceptable results,which severely limits its deployment in resource-limited platforms.This paper proposes DarkFPGA,a novel customizable framework to efficiently accelerate the entire DNN training on a single FPGA platform.First,we explore batch-level parallelism to enable efficient FPGA-based DNN training.Second,we devise a novel hardware architecture optimised by a batch-oriented data pattern and tiling techniques to effectively exploit parallelism.Moreover,an analytical model is developed to determine the optimal design parameters for the DarkFPGA accelerator with respect to a specific network specification and FPGA resource constraints.Our results show that the accelerator is able to perform about 10 times faster than CPU training and about a third of the energy consumption than GPU training using 8-bit integers for training VGG-like networks on the CIFAR dataset for the Maxeler MAX5 platform.
基金Sponsored by the National High-Tech Research and Development Program(Grant No.2006AA04Z201)
文摘The concept of legged-robot stability training with a training platform is proposed and a serial-parallel mechanism platform with 6 degrees of freedom is designed for this target. The designed platform is composed of 4-DOF parallel mechanism with spherical joints and prismatic pairs,and 2-DOF serial mechanism with prismatic pairs. With this design,the platform has advantages of low platform countertop,big workspace,high carrying capacity and high stiffness. On the basis of DOF analysis and computation of space mechanism,weight supporting auxiliary mechanism and raceways-balls supporting mechanism are designed,so as to improve the stiffness of designed large platform and payload capacity of servo motors. And then the whole structure design work of the platform is done. Meanwhile,this paper derives the analytical solutions of forward kinematics, inverse kinematics and inverse dynamics. The error analysis model of position and orientation is established. And then the simulation is done in ADAMS to ensure the correctness and feasibility of this design.
基金The editing contribution of Mr.Tim McSweeney(Adjunct Research Fellow,Centre for Railway Engineering)is gratefully acknowledged.
文摘Railway train energy simulation is an important and popular research topic.Locomotive traction force simulations are a fundamental part of such research.Conventional energy calculation models are not able to consider locomotive wheel-rail adhesions,traction adhesion control,and locomotive dynamics.This paper has developed two models to fill this research gap.The first model uses a 2D locomotive model with 27 degrees of freedom and a simplified wheel-rail contact model.The second model uses a 3D locomotive model with 54 degrees of freedom and a fully detailed wheel-rail contact model.Both models were integrated into a longitudinal train dynamics model with the consideration of locomotive adhesion control.Energy consumption simulations using a conventional model(1D model)and the two new models(2D and 3D models)were conducted and compared.The results show that,due to the consideration of wheel-rail adhesion model and traction control in the 3D model,it reports less energy consumption than the 1D model.The maximum difference in energy consumption rate between the 3D model and the 1D model was 12.5%.Due to the consideration of multiple wheel-rail contact points in the 3D model,it reports higher energy consumption than the 2D model.An 8.6%maximum difference in energy consumption rate between the 3D model and the 1D model was reported during curve negotiation.