个性化联邦学习因其在应对数据异质性和隐私保护方面的优势而备受关注。现有算法专注于平衡全局信息和个性化信息之间的矛盾,忽视了全局信息中的不同标签信息带来的干扰,尤其在维护单一全局头部的算法中,容易出现标签间特征冲突导致的...个性化联邦学习因其在应对数据异质性和隐私保护方面的优势而备受关注。现有算法专注于平衡全局信息和个性化信息之间的矛盾,忽视了全局信息中的不同标签信息带来的干扰,尤其在维护单一全局头部的算法中,容易出现标签间特征冲突导致的收敛困难。为此,提出一种新的算法——全局多头部联邦学习(federated learning with global multi-head,FedGMH)算法,该算法在服务器创建多个全局头部,每个头部专门处理一种标签信息,而客户端下载与本地标签相关的全局头部,从而避免无关标签信息的干扰。此外,FedGMH引入参数级聚合机制:评估头部参数重要性,并将关键参数更新为全局多头部的加权参数,以加快收敛速度并且提高准确率。在3个视觉数据集上的大量实验表明,FedGMH优于现有的先进算法。展开更多
Sensor nodes in a wireless sensor network （WSN） are typically powered by batteries, thus the energy is constrained. It is our design goal to efficiently utilize the energy of each sensor node to extend its lifetime,...Sensor nodes in a wireless sensor network （WSN） are typically powered by batteries, thus the energy is constrained. It is our design goal to efficiently utilize the energy of each sensor node to extend its lifetime, so as to prolong the lifetime of the whole WSN. In this paper, we propose a path-based data aggregation scheme （PBDAS） for grid-based wireless sensor networks. In order to extend the lifetime of a WSN, we construct a grid infrastructure by partitioning the whole sensor field into a grid of cells. Each cell has a head responsible for aggregating its own data with the data sensed by the others in the same cell and then transmitting out. In order to efficiently and rapidly transmit the data to the base station （BS）, we link each cell head to form a chain. Each cell head on the chain takes turn becoming the chain leader responsible for transmitting data to the BS. Aggregated data moves from head to head along the chain, and finally the chain leader transmits to the BS. In PBDAS, only the cell heads need to transmit data toward the BS. Therefore, the data transmissions to the BS substantially decrease. Besides, the cell heads and chain leader are designated in turn according to the energy level so that the energy depletion of nodes is evenly distributed. Simulation results show that the proposed PBDAS extends the lifetime of sensor nodes, so as to make the lifetime of the whole network longer.展开更多
Head and neck squamous cell carcinoma (HNSCC) is the sixth most common cause of cancer mortality in the world and the 5th most commonly occurring cancer (Siegel, R. 2014). In the last few decades a growing interest fo...Head and neck squamous cell carcinoma (HNSCC) is the sixth most common cause of cancer mortality in the world and the 5th most commonly occurring cancer (Siegel, R. 2014). In the last few decades a growing interest for the emerging data from both tumor biology and multimodality treatment in HNSCC has been developed. A huge number of new markers need to be managed with bio-informatics systems to elaborate and correlate clinical and molecular data. Data mining algorithms are a promising medical application. We used this technology to correlate blood samples with clinical outcome in 120 patients treated with chemoradiation for locally advanced HNSCC. Our results did not find a significant correlation because of the sample exiguity but they show the potential of this tool.展开更多
【目的】城市功能区是城市规划和人类活动共同作用、相互影响的结果,其准确识别对于优化配置公共资源和高效组织商业活动具有重要意义。目前,许多研究利用新兴的社会感知大数据进行城市功能区识别,但往往未能挖掘这些数据中蕴含的深层...【目的】城市功能区是城市规划和人类活动共同作用、相互影响的结果,其准确识别对于优化配置公共资源和高效组织商业活动具有重要意义。目前,许多研究利用新兴的社会感知大数据进行城市功能区识别,但往往未能挖掘这些数据中蕴含的深层次特征,或者未能充分捕捉和利用不同特征之间的相互关系和关联性,导致识别精度较低。【方法】针对这些问题,本研究提出了一种融合区域嵌入表示的城市功能区识别框架。该方法基于手机定位数据和兴趣点数据(Point of Interest,POI),采用Node2vec算法提取工作日与周末6个时段的区域间空间交互特征,并利用GloVe模型提取区域的语义特征。随后,通过多头注意力机制进行特征融合,并结合部分人工标注的功能区进行分类识别,在福州市三环以内地区进行了实证研究。【结果】实验结果表明,本方法生成的区域表示特征具有较高区分度,能够有效识别6类功能区,总体精度(OA)为81%,Kappa系数为0.77。【结论】与DTW_KNN和Word2Vec方法相比,精度分别提高了30%和20%,能够充分挖掘具有全局性质的空间交互特征和语义特征。此外,消融实验进一步表明,与单一数据源或简单融合方法相比,本方法在捕捉区域内部和区域间复杂关系的同时,对重要特征赋予更高的权重,使得模型的整体OA值相较于单源数据提高了约18%和6%,相较于简单融合方法提高了约13%,尤其在住宅区和混合区的识别方面表现出了显著优势。展开更多
针对脑电数据不足和数据异构导致运动想象解码性能下降,提出一种融合双服务器结构和多头注意力机制的联邦学习模型,实现异构脑电数据的运动想象多任务识别。通过欧几里得对齐,减少客户端数据分布差异。设计双服务器架构解决数据异构,服...针对脑电数据不足和数据异构导致运动想象解码性能下降,提出一种融合双服务器结构和多头注意力机制的联邦学习模型,实现异构脑电数据的运动想象多任务识别。通过欧几里得对齐,减少客户端数据分布差异。设计双服务器架构解决数据异构,服务器一选择并共享本地模型的最佳特征,将其作为全局共享资源改进客户端更新策略,解决客户端漂移问题;服务器二聚合本地模型参数,并进行全局参数微调训练,增强模型适应不同场景脑电数据异构的能力,提高全局模型的通用性。利用Transfomer多头自注意力,提高运动想象脑电的特征表示和模型学习能力。在BCI IV 2a脑机接口竞赛数据集上,该模型与联邦学习基准模型相比,脑电解码平均准确率提升了21.05百分点,Kappa值提升了0.283。在两个脑电数据集进行不同用户、不同环境和设备的跨数据集测试,脑电解码分别获得了71.13%和86.63%的平均准确率以及0.615和0.822的Kappa系数。结果表明:该模型在多用户、多设备且数据高度异构的场景下,运动想象脑电识别能够获得较好的性能,具有较强的泛化性。展开更多
文摘个性化联邦学习因其在应对数据异质性和隐私保护方面的优势而备受关注。现有算法专注于平衡全局信息和个性化信息之间的矛盾,忽视了全局信息中的不同标签信息带来的干扰,尤其在维护单一全局头部的算法中,容易出现标签间特征冲突导致的收敛困难。为此,提出一种新的算法——全局多头部联邦学习(federated learning with global multi-head,FedGMH)算法,该算法在服务器创建多个全局头部,每个头部专门处理一种标签信息,而客户端下载与本地标签相关的全局头部,从而避免无关标签信息的干扰。此外,FedGMH引入参数级聚合机制:评估头部参数重要性,并将关键参数更新为全局多头部的加权参数,以加快收敛速度并且提高准确率。在3个视觉数据集上的大量实验表明,FedGMH优于现有的先进算法。
基金supported by the NSC under Grant No.NSC-101-2221-E-239-032 and NSC-102-2221-E-239-020
文摘Sensor nodes in a wireless sensor network （WSN） are typically powered by batteries, thus the energy is constrained. It is our design goal to efficiently utilize the energy of each sensor node to extend its lifetime, so as to prolong the lifetime of the whole WSN. In this paper, we propose a path-based data aggregation scheme （PBDAS） for grid-based wireless sensor networks. In order to extend the lifetime of a WSN, we construct a grid infrastructure by partitioning the whole sensor field into a grid of cells. Each cell has a head responsible for aggregating its own data with the data sensed by the others in the same cell and then transmitting out. In order to efficiently and rapidly transmit the data to the base station （BS）, we link each cell head to form a chain. Each cell head on the chain takes turn becoming the chain leader responsible for transmitting data to the BS. Aggregated data moves from head to head along the chain, and finally the chain leader transmits to the BS. In PBDAS, only the cell heads need to transmit data toward the BS. Therefore, the data transmissions to the BS substantially decrease. Besides, the cell heads and chain leader are designated in turn according to the energy level so that the energy depletion of nodes is evenly distributed. Simulation results show that the proposed PBDAS extends the lifetime of sensor nodes, so as to make the lifetime of the whole network longer.
文摘Head and neck squamous cell carcinoma (HNSCC) is the sixth most common cause of cancer mortality in the world and the 5th most commonly occurring cancer (Siegel, R. 2014). In the last few decades a growing interest for the emerging data from both tumor biology and multimodality treatment in HNSCC has been developed. A huge number of new markers need to be managed with bio-informatics systems to elaborate and correlate clinical and molecular data. Data mining algorithms are a promising medical application. We used this technology to correlate blood samples with clinical outcome in 120 patients treated with chemoradiation for locally advanced HNSCC. Our results did not find a significant correlation because of the sample exiguity but they show the potential of this tool.
文摘【目的】城市功能区是城市规划和人类活动共同作用、相互影响的结果,其准确识别对于优化配置公共资源和高效组织商业活动具有重要意义。目前,许多研究利用新兴的社会感知大数据进行城市功能区识别,但往往未能挖掘这些数据中蕴含的深层次特征,或者未能充分捕捉和利用不同特征之间的相互关系和关联性,导致识别精度较低。【方法】针对这些问题,本研究提出了一种融合区域嵌入表示的城市功能区识别框架。该方法基于手机定位数据和兴趣点数据(Point of Interest,POI),采用Node2vec算法提取工作日与周末6个时段的区域间空间交互特征,并利用GloVe模型提取区域的语义特征。随后,通过多头注意力机制进行特征融合,并结合部分人工标注的功能区进行分类识别,在福州市三环以内地区进行了实证研究。【结果】实验结果表明,本方法生成的区域表示特征具有较高区分度,能够有效识别6类功能区,总体精度(OA)为81%,Kappa系数为0.77。【结论】与DTW_KNN和Word2Vec方法相比,精度分别提高了30%和20%,能够充分挖掘具有全局性质的空间交互特征和语义特征。此外,消融实验进一步表明,与单一数据源或简单融合方法相比,本方法在捕捉区域内部和区域间复杂关系的同时,对重要特征赋予更高的权重,使得模型的整体OA值相较于单源数据提高了约18%和6%,相较于简单融合方法提高了约13%,尤其在住宅区和混合区的识别方面表现出了显著优势。
文摘针对脑电数据不足和数据异构导致运动想象解码性能下降,提出一种融合双服务器结构和多头注意力机制的联邦学习模型,实现异构脑电数据的运动想象多任务识别。通过欧几里得对齐,减少客户端数据分布差异。设计双服务器架构解决数据异构,服务器一选择并共享本地模型的最佳特征,将其作为全局共享资源改进客户端更新策略,解决客户端漂移问题;服务器二聚合本地模型参数,并进行全局参数微调训练,增强模型适应不同场景脑电数据异构的能力,提高全局模型的通用性。利用Transfomer多头自注意力,提高运动想象脑电的特征表示和模型学习能力。在BCI IV 2a脑机接口竞赛数据集上,该模型与联邦学习基准模型相比,脑电解码平均准确率提升了21.05百分点,Kappa值提升了0.283。在两个脑电数据集进行不同用户、不同环境和设备的跨数据集测试,脑电解码分别获得了71.13%和86.63%的平均准确率以及0.615和0.822的Kappa系数。结果表明:该模型在多用户、多设备且数据高度异构的场景下,运动想象脑电识别能够获得较好的性能,具有较强的泛化性。
