状态可变数据流(Mutable States Data Flow,MS-DF)是机器学习系统运行时的主要特征,MS-DF可由有向图来表示,其顶点由算子构成,表示机器学习运算逻辑;边代表算子之间的输入输出依赖关系.MS-DF的集群调度是保障机器学习系统高效运行的主...状态可变数据流(Mutable States Data Flow,MS-DF)是机器学习系统运行时的主要特征,MS-DF可由有向图来表示,其顶点由算子构成,表示机器学习运算逻辑;边代表算子之间的输入输出依赖关系.MS-DF的集群调度是保障机器学习系统高效运行的主要工作,如何高效进行MS-DF的集群调度已经成为机器学习的研究热点.其中,机器学习系统(TensorFlow、PyTorch等)作为中间层解耦了机器学习运算逻辑和资源分配(CPU,GPU,FGPA),从而机器学习无需再“独占式”静态绑定资源,而是由机器学习系统运行时动态管理,而算子是该解耦过程的关键要素,这给MS-DF的集群调度带来了新的挑战,这些挑战主要由算子资源需求刻画的准确性、算子调度决策的适应性和算子调度调整的差异性这三方面导致的.首先介绍算子资源需求的感知、协同两个机制,以克服多种算子组合导致其自身资源需求难以准确刻画的挑战;然后,通过决策约束、决策模型和决策求解来介绍算子调度决策,以应对算子状态频繁变化带来的适应性挑战;接着,介绍迁移、伸缩、挂起恢复等算子调度调整策略,以适用于不同算子状态同步方式带来的差异性挑战.最后,基于上述三个挑战,对近年来的集群调度最新研究成果进行归纳和分析,并展望MS-DF的集群调度,指出算子异构资源需求多层次分析及协同刻画、算子复杂调度约束的灵活定义和发现、学习驱动的算子低成本调度调整技术是其主要发展方向.展开更多
r-learning,which is based on e-learning and u-learning,is defined as a learning support system that intelligent robots serve verbal and nonverbal interactions on ubiquitous computing environment.In order to guarantee ...r-learning,which is based on e-learning and u-learning,is defined as a learning support system that intelligent robots serve verbal and nonverbal interactions on ubiquitous computing environment.In order to guarantee the advantages of r-learning contents with no limits of timc and place and with nonverbal interaction which are not in e-learning contents,in recent years,assessment criteria for r-learning contents are urgently rcquired.Therefore,the reliable and valid assessment criteria were developed for nonverbal interaction contents in r-learning,and its detailed research content is as follows.First,assessment criteria for nonverbal interaction in r-learning contents will be specified into gesture,facial expression,semi-verbal message,distance,physical contact and time.Second,the validity of the developed assessment criteria will be proved by statistics.Consequently,the assessment criteria for nonverbal interaction contents will be helpful when choosing the better r-learning content and producing the better r-learning content,and the reliability of school education is improved ultimately.展开更多
文摘状态可变数据流(Mutable States Data Flow,MS-DF)是机器学习系统运行时的主要特征,MS-DF可由有向图来表示,其顶点由算子构成,表示机器学习运算逻辑;边代表算子之间的输入输出依赖关系.MS-DF的集群调度是保障机器学习系统高效运行的主要工作,如何高效进行MS-DF的集群调度已经成为机器学习的研究热点.其中,机器学习系统(TensorFlow、PyTorch等)作为中间层解耦了机器学习运算逻辑和资源分配(CPU,GPU,FGPA),从而机器学习无需再“独占式”静态绑定资源,而是由机器学习系统运行时动态管理,而算子是该解耦过程的关键要素,这给MS-DF的集群调度带来了新的挑战,这些挑战主要由算子资源需求刻画的准确性、算子调度决策的适应性和算子调度调整的差异性这三方面导致的.首先介绍算子资源需求的感知、协同两个机制,以克服多种算子组合导致其自身资源需求难以准确刻画的挑战;然后,通过决策约束、决策模型和决策求解来介绍算子调度决策,以应对算子状态频繁变化带来的适应性挑战;接着,介绍迁移、伸缩、挂起恢复等算子调度调整策略,以适用于不同算子状态同步方式带来的差异性挑战.最后,基于上述三个挑战,对近年来的集群调度最新研究成果进行归纳和分析,并展望MS-DF的集群调度,指出算子异构资源需求多层次分析及协同刻画、算子复杂调度约束的灵活定义和发现、学习驱动的算子低成本调度调整技术是其主要发展方向.
基金Project(2011)supported by the research grant of the Chungbuk National University,South Korea
文摘r-learning,which is based on e-learning and u-learning,is defined as a learning support system that intelligent robots serve verbal and nonverbal interactions on ubiquitous computing environment.In order to guarantee the advantages of r-learning contents with no limits of timc and place and with nonverbal interaction which are not in e-learning contents,in recent years,assessment criteria for r-learning contents are urgently rcquired.Therefore,the reliable and valid assessment criteria were developed for nonverbal interaction contents in r-learning,and its detailed research content is as follows.First,assessment criteria for nonverbal interaction in r-learning contents will be specified into gesture,facial expression,semi-verbal message,distance,physical contact and time.Second,the validity of the developed assessment criteria will be proved by statistics.Consequently,the assessment criteria for nonverbal interaction contents will be helpful when choosing the better r-learning content and producing the better r-learning content,and the reliability of school education is improved ultimately.
