联邦学习(Federated Learning)是为了解决机器学习中以隐私保护为前提的数据碎片化和隔离问题。各客户端节点在本地训练数据,将训练的模型参数信息上传到中央服务器,由参数服务器聚合参数信息以达到共同训练的目的。由于现实环境中,各...联邦学习(Federated Learning)是为了解决机器学习中以隐私保护为前提的数据碎片化和隔离问题。各客户端节点在本地训练数据,将训练的模型参数信息上传到中央服务器,由参数服务器聚合参数信息以达到共同训练的目的。由于现实环境中,各节点数据之间的分布往往不一致,通过分析非独立同分布数据对联邦学习准确率的影响,来证明传统联邦学习方法得到的模型精度较低。因此,采用多样化抽样策略模拟数据倾斜度分布,提出了基于DBSCAN(Density-Based Spatial Clustering of Applications with Noise)聚类的集群联邦学习算法(DBSCAN Based Cluster Federated Learning,DCFL),解决了联邦学习中不同节点的数据非独立同分布降低了学习准确率的问题。在Mnist和Cifar-10标准数据集上进行了实验,相比传统的联邦学习算法,基于DBSCAN聚类的集群联邦学习算法对模型的准确率有较大的提升。展开更多
Genetic diversity of 299 inbred indica rice varieties, including 33 introduced varieties, applied in Guangdong Province of China were assessed using 20 ILP （intron length polymorphism） and 34 SSR （simple sequence r...Genetic diversity of 299 inbred indica rice varieties, including 33 introduced varieties, applied in Guangdong Province of China were assessed using 20 ILP （intron length polymorphism） and 34 SSR （simple sequence repeat） markers. Totally, 154 loci were screened for the 299 varieties, with the average number of alleles （Na）, rare alleles （Nr）, and polymorphism information content （PIC） scored at 3.4, 0.7 and 0.32, respectively. The Nei＇s genetic distance （GD） was estimated ranging from 0 to 0.7529 with an average of 0.4797. There was no significant difference of Na, Nr, PIC or GDs between the introduced and local varieties. Neighbor-joining （N J） analysis showed that the 299 varieties failed into three main distinct groups, and the 33 introduced varieties were distributed over all the groups or subgroups. Model-based cluster analysis demonstrated that only 73 （24.4%） of the 299 varieties and 7 （21.2%） of the 33 introduced varieties could be distinctly classified into the three groups. Analysis of molecular variance showed that within the groups divided by NJ analysis, the genetic variations revealed by ILP, SSR and these two combined were 7.7%, 5.6% and 6.6%, and within the groups divided by region （Guangdong local and the introduced varieties）, the genetic variables were 2.1%, 4.6%, 5.4%, respectively. These results suggested that the genetic diversity of the 299 inbred rice varieties in Guangdong Province was low, simultaneously relationship among varieties was poor and close in all kind of groups. Hence, it is very necessary to extend the genetic diversity during the breeding and selection practical procedure.展开更多
文摘联邦学习(Federated Learning)是为了解决机器学习中以隐私保护为前提的数据碎片化和隔离问题。各客户端节点在本地训练数据,将训练的模型参数信息上传到中央服务器,由参数服务器聚合参数信息以达到共同训练的目的。由于现实环境中,各节点数据之间的分布往往不一致,通过分析非独立同分布数据对联邦学习准确率的影响,来证明传统联邦学习方法得到的模型精度较低。因此,采用多样化抽样策略模拟数据倾斜度分布,提出了基于DBSCAN(Density-Based Spatial Clustering of Applications with Noise)聚类的集群联邦学习算法(DBSCAN Based Cluster Federated Learning,DCFL),解决了联邦学习中不同节点的数据非独立同分布降低了学习准确率的问题。在Mnist和Cifar-10标准数据集上进行了实验,相比传统的联邦学习算法,基于DBSCAN聚类的集群联邦学习算法对模型的准确率有较大的提升。
基金supported by the Guangdong Natural Science Foundation of China (Grant No. S2012040007829)the National High Technology Research and Development Program of China (Grant No. 2012AA101201)
文摘Genetic diversity of 299 inbred indica rice varieties, including 33 introduced varieties, applied in Guangdong Province of China were assessed using 20 ILP （intron length polymorphism） and 34 SSR （simple sequence repeat） markers. Totally, 154 loci were screened for the 299 varieties, with the average number of alleles （Na）, rare alleles （Nr）, and polymorphism information content （PIC） scored at 3.4, 0.7 and 0.32, respectively. The Nei＇s genetic distance （GD） was estimated ranging from 0 to 0.7529 with an average of 0.4797. There was no significant difference of Na, Nr, PIC or GDs between the introduced and local varieties. Neighbor-joining （N J） analysis showed that the 299 varieties failed into three main distinct groups, and the 33 introduced varieties were distributed over all the groups or subgroups. Model-based cluster analysis demonstrated that only 73 （24.4%） of the 299 varieties and 7 （21.2%） of the 33 introduced varieties could be distinctly classified into the three groups. Analysis of molecular variance showed that within the groups divided by NJ analysis, the genetic variations revealed by ILP, SSR and these two combined were 7.7%, 5.6% and 6.6%, and within the groups divided by region （Guangdong local and the introduced varieties）, the genetic variables were 2.1%, 4.6%, 5.4%, respectively. These results suggested that the genetic diversity of the 299 inbred rice varieties in Guangdong Province was low, simultaneously relationship among varieties was poor and close in all kind of groups. Hence, it is very necessary to extend the genetic diversity during the breeding and selection practical procedure.
