To better complete various missions, it is necessary to plan an optimal trajectory or provide the optimal control law for the multirole missile according to the actual situation, including launch conditions and target...To better complete various missions, it is necessary to plan an optimal trajectory or provide the optimal control law for the multirole missile according to the actual situation, including launch conditions and target location. Since trajectory optimization struggles to meet real-time requirements, the emergence of data-based generation methods has become a significant focus in contemporary research. However, due to the large differences in the characteristics of the optimal control laws caused by the diversity of tasks, it is difficult to achieve good prediction results by modeling all data with one single model.Therefore, the modeling idea of the mixture of experts(MoE) is adopted. Firstly, the K-means clustering algorithm is used to partition the sample data set, and the corresponding neural network classification model is established as the gate switch of MoE. Then, the expert models, i.e., the mappings from the generation conditions to the optimal control law represented by the results of principal component analysis(PCA), are represented by Kriging models. Finally, multiple rounds of accuracy evaluation, sample supplementation, and model updating are conducted to improve the generation accuracy. The Monte Carlo simulation shows that the accuracy of the proposed model reaches 96% and the generation efficiency meets the real-time requirement.展开更多
近年来,大模型推动自然语言处理、机器视觉等众多领域取得前所未有的进展.混合专家(mixture of experts,MoE)凭借在模型参数扩展、计算成本控制和复杂任务处理等方面的独特优势成为大模型的主流架构之一.然而,随着参数规模的持续增长,...近年来,大模型推动自然语言处理、机器视觉等众多领域取得前所未有的进展.混合专家(mixture of experts,MoE)凭借在模型参数扩展、计算成本控制和复杂任务处理等方面的独特优势成为大模型的主流架构之一.然而,随着参数规模的持续增长,系统的执行效率和可扩展能力愈发难以满足需求,亟待解决.系统优化方法是解决这一挑战的有效途径,日益成为研究热点.故综述大模型时代MoE系统优化技术的研究现状,首先介绍MoE大模型的发展现状,并分析其在系统端面临的性能瓶颈;然后从内存占用、通信延迟、计算效率和并行扩展4个系统核心维度对最新的研究进展进行全面梳理和深入分析,并对其中涉及的关键技术、适用场景和待优化方向进行详细对比阐述;最后总结MoE系统优化的研究现状,并展望未来研究方向.展开更多
有监督异常检测因其精准的工业异常检测能力而广泛应用于布匹质量检测。现有的统一架构的异常检测方法,因其单一的特征适配能力,不能对多样化的,所以度较高的布匹瑕疵进行有效地区分,因此在布匹的多类别的异常检测中性能会显著下降。为...有监督异常检测因其精准的工业异常检测能力而广泛应用于布匹质量检测。现有的统一架构的异常检测方法,因其单一的特征适配能力,不能对多样化的,所以度较高的布匹瑕疵进行有效地区分,因此在布匹的多类别的异常检测中性能会显著下降。为此提出一种基于混合区域匹配专家适配方法(Mixture of Region Experts),通过Mixture of Adapter Experts模块来区别化不同类别的布匹瑕疵特征,使用Align and Differencing模块对齐模板图特征和瑕疵特征来进一步加强异常区域的划分,从而有效提高了模型分辨复杂多类型的布匹瑕疵的能力。同时,模型进一步集成成分检测任务,在完成瑕疵定位的基础上实现异常成分的语义识别。实验结果表明,SAM-MR在布匹纤维材质和缺陷检测任务上取得了优于现有方法的性能,定性、定量分析及消融实验验证了所提出方法在多任务预测中的有效性。展开更多
为解决不同人员相同操作的个体差异以及同一人员不同时间相同操作差异的问题,提出一种基于混合专家系统(mixture of experts,MoE)和长短期记忆神经网络(long short-term memory,LSTM)的倒闸操作识别方法MoE-LSTM。基于MoE对LSTM进行集成...为解决不同人员相同操作的个体差异以及同一人员不同时间相同操作差异的问题,提出一种基于混合专家系统(mixture of experts,MoE)和长短期记忆神经网络(long short-term memory,LSTM)的倒闸操作识别方法MoE-LSTM。基于MoE对LSTM进行集成,学习不同来源数据的特征分布。采集加速度动作数据构建倒闸操作数据集,基于滑动窗口对动作序列进行切分;将动作序列输入到MoE-LSTM中,由不同LSTM独立学习不同动作的时序依赖;通过门控网络选择对当前输入分类较好的LSTM的输出作为动作识别结果。仿真结果表明:不同LSTM对来自不同时空的动作数据都有擅长分类的特征空间。展开更多
基金Defense Industrial Technology Development Program (JCKY2020204B016)National Natural Science Foundation of China (92471206)。
文摘To better complete various missions, it is necessary to plan an optimal trajectory or provide the optimal control law for the multirole missile according to the actual situation, including launch conditions and target location. Since trajectory optimization struggles to meet real-time requirements, the emergence of data-based generation methods has become a significant focus in contemporary research. However, due to the large differences in the characteristics of the optimal control laws caused by the diversity of tasks, it is difficult to achieve good prediction results by modeling all data with one single model.Therefore, the modeling idea of the mixture of experts(MoE) is adopted. Firstly, the K-means clustering algorithm is used to partition the sample data set, and the corresponding neural network classification model is established as the gate switch of MoE. Then, the expert models, i.e., the mappings from the generation conditions to the optimal control law represented by the results of principal component analysis(PCA), are represented by Kriging models. Finally, multiple rounds of accuracy evaluation, sample supplementation, and model updating are conducted to improve the generation accuracy. The Monte Carlo simulation shows that the accuracy of the proposed model reaches 96% and the generation efficiency meets the real-time requirement.
文摘近年来,大模型推动自然语言处理、机器视觉等众多领域取得前所未有的进展.混合专家(mixture of experts,MoE)凭借在模型参数扩展、计算成本控制和复杂任务处理等方面的独特优势成为大模型的主流架构之一.然而,随着参数规模的持续增长,系统的执行效率和可扩展能力愈发难以满足需求,亟待解决.系统优化方法是解决这一挑战的有效途径,日益成为研究热点.故综述大模型时代MoE系统优化技术的研究现状,首先介绍MoE大模型的发展现状,并分析其在系统端面临的性能瓶颈;然后从内存占用、通信延迟、计算效率和并行扩展4个系统核心维度对最新的研究进展进行全面梳理和深入分析,并对其中涉及的关键技术、适用场景和待优化方向进行详细对比阐述;最后总结MoE系统优化的研究现状,并展望未来研究方向.
文摘有监督异常检测因其精准的工业异常检测能力而广泛应用于布匹质量检测。现有的统一架构的异常检测方法,因其单一的特征适配能力,不能对多样化的,所以度较高的布匹瑕疵进行有效地区分,因此在布匹的多类别的异常检测中性能会显著下降。为此提出一种基于混合区域匹配专家适配方法(Mixture of Region Experts),通过Mixture of Adapter Experts模块来区别化不同类别的布匹瑕疵特征,使用Align and Differencing模块对齐模板图特征和瑕疵特征来进一步加强异常区域的划分,从而有效提高了模型分辨复杂多类型的布匹瑕疵的能力。同时,模型进一步集成成分检测任务,在完成瑕疵定位的基础上实现异常成分的语义识别。实验结果表明,SAM-MR在布匹纤维材质和缺陷检测任务上取得了优于现有方法的性能,定性、定量分析及消融实验验证了所提出方法在多任务预测中的有效性。
文摘为解决不同人员相同操作的个体差异以及同一人员不同时间相同操作差异的问题,提出一种基于混合专家系统(mixture of experts,MoE)和长短期记忆神经网络(long short-term memory,LSTM)的倒闸操作识别方法MoE-LSTM。基于MoE对LSTM进行集成,学习不同来源数据的特征分布。采集加速度动作数据构建倒闸操作数据集,基于滑动窗口对动作序列进行切分;将动作序列输入到MoE-LSTM中,由不同LSTM独立学习不同动作的时序依赖;通过门控网络选择对当前输入分类较好的LSTM的输出作为动作识别结果。仿真结果表明:不同LSTM对来自不同时空的动作数据都有擅长分类的特征空间。
