Cloudless-Training:基于serverless的高效跨地域分布式ML训练框架

跨地域分布式机器学习(ML)训练能够联合多区域的云资源协作训练,可满足许多新兴ML场景(比如大型模型训练、联邦学习)的训练需求。但其训练效率仍受2方面挑战的制约。首先,多区域云资源缺乏有效的弹性调度,这会影响训练的资源利用率和性能;其次,模型跨地域同步需要在广域网(WAN)上高频通信,受WAN的低带宽和高波动的影响,会产生巨大通信开销。本文提出Cloudless-Training,从3个方面实现高效的跨地域分布式ML训练。首先,它基于serverless计算模式实现,使用控制层和训练执行层的2层架构,支持多云区域的弹性调度和通信。其次,它提供一种弹性调度策略,根据可用云资源的异构性和训练数据集的分布自适应地部署训练工作流。最后,它提供了2种高效的跨云同步策略,包括基于梯度累积的异步随机梯度下降(ASGD-GA)和跨云参数服务器(PS)间的模型平均(MA)。Cloudless-Training是基于OpenFaaS实现的,并被部署在腾讯云上评估,实验结果表明Cloudless-Training可显著地提高跨地域分布式ML训练的资源利用率(训练成本降低了9.2%~24.0%)和同步效率(训练速度最多比基线快1.7倍),并能保证模型的收敛精度。

体积>80mL的良性前列腺增生患者经尿道等离子前列腺电切术中应用尖部收切法的可行性与安全性分析

目的:分析体积>80mL的良性前列腺增生(BPH)患者经尿道等离子前列腺电切术(TUP-KP)术中应用尖部收切法的可行性与安全性。方法:选取2019年10月至2022年10月196例我院体积>80mL、择期手术的BPH患者,随机分为两组,研究组(n=98,采用TUPKP术中应用尖部收切法),常规组(n=98,采用常规TUPKP术),比较两组患者手术相关指标、逼尿肌稳定性相关指标、前列腺症状、生活质量及并发症。结果:两组手术相关指标差异无统计学意义(P>0.05);术后3个月,两组患者逼尿肌压力、初始尿意容量、排尿后残尿量以及最大尿意容量各自较术前相比皆有所改善(P<0.05),但两组患者组间上述指标差值均无统计学意义(P>0.05);术后3个月,研究组国际前列腺症状评分等级优于常规组(P<0.05),且时间与前列腺症状分级的交互项具有显著性(P<0.05);术后3个月,两组患者生活质量评分量表各维度较术前均升高(P<0.05),研究组生活质量评分量表各维度差值均高于常规组(P<0.05);术后3个月,研究组总并发症发生率为2.04%,低于常规组的9.18%(P<0.05)。结论:TUPKP术中应用尖部收切法治疗体积>80mL的BPH患者可缓解症状,改善生活质量,且并发症少,安全可靠。

Empirical likelihood for spatial cross-sectional data models with matrix exponential spatial specification

In this paper,we study spatial cross-sectional data models in the form of matrix exponential spatial specification(MESS),where MESS appears in both dependent and error terms.The empirical likelihood(EL)ratio statistics are established for the parameters of the MESS model.It is shown that the limiting distributions of EL ratio statistics follow chi-square distributions,which are used to construct the confidence regions of model parameters.Simulation experiments are conducted to compare the performances of confidence regions based on EL method and normal approximation method.

基于PSO和MLEM混合算法的NDP测量反演算法研究

中子深度剖面(NDP)分析技术是一种无损检测方法,能够同时测量样品中目标核素的浓度与空间信息,已被广泛应用于锂电池、半导体等产业。在NDP分析过程中,由测量能谱反演出目标核素浓度的分布信息是关键步骤。目前NDP测量反演中常用的算法为最大似然期望最大化(MLEM)算法。针对MLEM算法计算结果易陷入局部最优解的情况,本文提出了粒子群(PSO)与MLEM混合(PSO-MLEM)算法,并通过动态加速因子提高了算法的收敛速度与计算精度。应用PSO-MLEM算法、PSO算法、MLEM算法、奇异值分解求解最小二乘(SVDLS)算法对锂电池中^(6)Li的NDP模拟能谱进行反演,并对反演计算结果进行了评价。结果表明:对比PSO算法,PSO-MLEM算法的收敛效率与计算精度明显提升;对比MLEM算法,PSO-MLEM算法的全局寻优能力有效提升了反演精度,避免了局部最优解的影响;对比SVDLS算法,PSO-MLEM算法的反演精度明显提升。

Beamspace maximum likelihood algorithm based on sum and difference beams for elevation estimation

Beamspace super-resolution methods for elevation estimation in multipath environment has attracted significant attention, especially the beamspace maximum likelihood(BML)algorithm. However, the difference beam is rarely used in superresolution methods, especially in low elevation estimation. The target airspace information in the difference beam is different from the target airspace information in the sum beam. And the use of difference beams does not significantly increase the complexity of the system and algorithms. Thus, this paper applies the difference beam to the beamformer to improve the elevation estimation performance of BML algorithm. And the direction and number of beams can be adjusted according to the actual needs. The theoretical target elevation angle root means square error(RMSE) and the computational complexity of the proposed algorithms are analyzed. Finally, computer simulations and real data processing results demonstrate the effectiveness of the proposed algorithms.

心理疏导及认知干预对提升400 mL献血率的影响

目的 探讨心理疏导和认知干预对提升400 mL献血率的影响。方法 选取2021年1月—2023年8月青岛市中心血站西海岸第一献血服务部的24 700名志愿者参与研究,按献血时间先后顺序进行分组,其中2021年1月—2022年4月接受常规健康教育的12 350名志愿者为对照组,2022年5月—2023年8月接受心理疏导和认知干预的12 350名志愿者为试验组。比较2组志愿者的献血知识知晓率、400 mL献血率、焦虑和抗拒情绪评分、献血满意度。结果 试验组献血知识知晓率为99.24%,400 mL献血率为85.11%,均高于对照组的98.61%、81.85%(P <0.05);试验组干预后的焦虑和抗拒情绪评分均低于对照组(P <0.05);试验组献血总满意度高于对照组,差异有统计学意义(P <0.05)。结论 心理疏导和认知干预在提升400 mL献血率、减轻焦虑和抗拒情绪方面具有显著效果;这些干预方法不仅在促进献血行为方面有潜在应用,还能提升献血体验的满意度。

顾及RMLE的GNSS时序噪声特性及环境负载修正

针对环境负载易对全球卫星导航系统(GNSS)站坐标时序修正前后噪声模型特性及其站速度造成影响的问题,提出一种顾及约束最大似然估计(RMLE)的GNSS时序噪声特性及环境负载修正方法:选用206个连续13a的GNSS站时间序列,采用改进的贝叶斯信息模型估计准则,利用约束最大似然估计(RMLE)方法探讨和分析大气压负载、非海洋潮汐负荷、积雪负载及土壤水负载等环境负载对GNSS时序修正前后噪声模型特性及其站速度的影响。结果表明:经环境负载修正后,GNSS时序中的噪声主要表现为闪烁噪声+白噪声(FNWN)与幂律噪声+白噪声(PLWN),北(N)、东(E)、天(U)3个分量上噪声模型修正前后分别有约19.9%、36.5%、40.8%的噪声模型发生变化,N、E、U分量上分别有96.6%、84.5%、88.3%的测站速度不确定度减小,对垂直速度估计的最大影响可达0.5mm/a;证明环境负载修正能够提高速度估值的精度。

Vulnerable brain regions in adolescent major depressive disorder:A resting-state functional magnetic resonance imaging activation likelihood estimation meta-analysis

BACKGROUND Adolescent major depressive disorder(MDD)is a significant mental health concern that often leads to recurrent depression in adulthood.Resting-state functional magnetic resonance imaging(rs-fMRI)offers unique insights into the neural mechanisms underlying this condition.However,despite previous research,the specific vulnerable brain regions affected in adolescent MDD patients have not been fully elucidated.AIM To identify consistent vulnerable brain regions in adolescent MDD patients using rs-fMRI and activation likelihood estimation(ALE)meta-analysis.METHODS We performed a comprehensive literature search through July 12,2023,for studies investigating brain functional changes in adolescent MDD patients.We utilized regional homogeneity(ReHo),amplitude of low-frequency fluctuations(ALFF)and fractional ALFF(fALFF)analyses.We compared the regions of aberrant spontaneous neural activity in adolescents with MDD vs healthy controls(HCs)using ALE.RESULTS Ten studies(369 adolescent MDD patients and 313 HCs)were included.Combining the ReHo and ALFF/fALFF data,the results revealed that the activity in the right cuneus and left precuneus was lower in the adolescent MDD patients than in the HCs(voxel size:648 mm3,P<0.05),and no brain region exhibited increased activity.Based on the ALFF data,we found decreased activity in the right cuneus and left precuneus in adolescent MDD patients(voxel size:736 mm3,P<0.05),with no regions exhibiting increased activity.CONCLUSION Through ALE meta-analysis,we consistently identified the right cuneus and left precuneus as vulnerable brain regions in adolescent MDD patients,increasing our understanding of the neuropathology of affected adolescents.

Multimodal abnormalities of brain structures in adolescents and young adults with major depressive disorder:An activation likelihood estimation meta-analysis

BACKGROUND Major depressive disorder(MDD)in adolescents and young adults contributes significantly to global morbidity,with inconsistent findings on brain structural changes from structural magnetic resonance imaging studies.Activation likeli-hood estimation(ALE)offers a method to synthesize these diverse findings and identify consistent brain anomalies.METHODS We performed a comprehensive literature search in PubMed,Web of Science,Embase,and Chinese National Knowledge Infrastructure databases for neuroi-maging studies on MDD among adolescents and young adults published up to November 19,2023.Two independent researchers performed the study selection,quality assessment,and data extraction.The ALE technique was employed to synthesize findings on localized brain function anomalies in MDD patients,which was supplemented by sensitivity analyses.RESULTS Twenty-two studies comprising fourteen diffusion tensor imaging(DTI)studies and eight voxel-based morphome-try(VBM)studies,and involving 451 MDD patients and 465 healthy controls(HCs)for DTI and 664 MDD patients and 946 HCs for VBM,were included.DTI-based ALE demonstrated significant reductions in fractional anisotropy(FA)values in the right caudate head,right insula,and right lentiform nucleus putamen in adolescents and young adults with MDD compared to HCs,with no regions exhibiting increased FA values.VBM-based ALE did not demonstrate significant alterations in gray matter volume.Sensitivity analyses highlighted consistent findings in the right caudate head(11 of 14 analyses),right insula(10 of 14 analyses),and right lentiform nucleus putamen(11 of 14 analyses).CONCLUSION Structural alterations in the right caudate head,right insula,and right lentiform nucleus putamen in young MDD patients may contribute to its recurrent nature,offering insights for targeted therapies.

Optimization of Generator Based on Gaussian Process Regression Model with Conditional Likelihood Lower Bound Search

The noise that comes from finite element simulation often causes the model to fall into the local optimal solution and over fitting during optimization of generator.Thus,this paper proposes a Gaussian Process Regression(GPR)model based on Conditional Likelihood Lower Bound Search(CLLBS)to optimize the design of the generator,which can filter the noise in the data and search for global optimization by combining the Conditional Likelihood Lower Bound Search method.Taking the efficiency optimization of 15 kW Permanent Magnet Synchronous Motor as an example.Firstly,this method uses the elementary effect analysis to choose the sensitive variables,combining the evolutionary algorithm to design the super Latin cube sampling plan;Then the generator-converter system is simulated by establishing a co-simulation platform to obtain data.A Gaussian process regression model combing the method of the conditional likelihood lower bound search is established,which combined the chi-square test to optimize the accuracy of the model globally.Secondly,after the model reaches the accuracy,the Pareto frontier is obtained through the NSGA-II algorithm by considering the maximum output torque as a constraint.Last,the constrained optimization is transformed into an unconstrained optimizing problem by introducing maximum constrained improvement expectation(CEI)optimization method based on the re-interpolation model,which cross-validated the optimization results of the Gaussian process regression model.The above method increase the efficiency of generator by 0.76%and 0.5%respectively;And this method can be used for rapid modeling and multi-objective optimization of generator systems.

ML20MnTiB线材表面翘皮缺陷成因分析

借助光学显微镜及扫描电镜SEM对缺陷横截面的形貌、组分等进行检测,通过氮氧分析仪对成品线材的气体含量进行检验,认为因钢中氮含量高生成的大量TiN夹杂物是导致线材表面缺陷形成的根本原因。通过对ML20MnTiB生产过程加强控制,控制钢中氮含量≤0.0060%,可有效控制TiN夹杂物的生成,进而避免线材表面翘皮缺陷的产生。

超声引导下22G八光针与5mL注射器针头细针穿刺活检在甲状腺TI-RADS 4类以上结节中的诊断价值

目的:比较可疑恶性的甲状腺结节经超声引导下22 G八光针与5 mL注射器针头细针穿刺活检(FNA)检查的诊断效能。方法:选取2021年1月—2022年9月佛山市南海区人民医院收治的甲状腺结节TI-RADS分类4类以上行FNA检查的患者291例,分为22 G八光针组(71例)和5 mL注射器针头组(220例),以手术病理结果为金标准,对比采用两种针具行US-FNA检查在取材满意程度、并发症发生情况以及对甲状腺良恶性病变的诊断效能差异。结果:对最大径<5 mm、5~10 mm、>10 mm的结节,有微小钙化灶、无钙化灶的结节,表现为低回声、等-高回声的结节以及不同血供类型的结节,5 mL注射器针头的取材标本满意度均高于22 G八光针组(P<0.05);两组并发症发生例数均较少(1/220、1/71),差异无统计学意义(P>0.05);5 mL注射器针头组对恶性病变的诊断灵敏度、准确率高于22 G八光针组,特异度低于22 G八光针组,差异有统计学意义(P<0.05)。结论:与八光针比较,超声引导下应用5 mL注射器针头进行FNA取材较满意、并发症发生率低、对甲状腺恶性病变有较好的灵敏度和准确率,同时5 mL注射器针头价格低廉,在患者群体较大的基层可作为FNA常用针具之一。

基于超效率SBM和ML的航空公司碳排放效率研究

为探究我国航空公司碳排放效率的特征和差异,探寻民航运输企业碳减排的多元路径,论文测算了我国九家航空公司2011-2019年碳排放量,运用超效率SBM模型和Malmquist-Luenberger指数法对航空公司碳排放效率进行了针对性研究,结果表明:我国航空公司碳排放效率整体处于中等偏上水平,呈波动增长趋势,技术效率是主要提升力;各类航空公司碳排放效率差异显著,呈现出民营>国有,小型>中型>大型的格局,但近几年这种差异有所减小;航空公司碳排放全要素生产率指数整体为下降趋势,但发展潜力较好,主要受技术进步的约束。因此,航空公司需要不断推进低碳技术的应用,合理规划自身运营管理的规模,积极构建绿色发展的治理体系。

区域科技人才开发效率测度与影响因素研究——基于EBM-ML-Tobit模型

科技是第一生产力,人才是第一资源,科技人才是国家和区域发展的重要保证,为测度我国典型区域科技人才的发展,以京津冀地区、长江经济带、黄河经济带三大经济带的区域科技人才开发效率为研究对象,运用EBM-ML模型对2007—2020年三个经济带的科技人才开发效率进行测算和动态分析,采用Tobit模型进行影响因素实证分析。结果表明:三个区域人才开发效率均处于较高水平,其中京津冀地区>长江经济带>黄河经济带,区域科技人才开发动态效率的提升与综合技术效率、技术进步的变化密切相关;人才开发效率较低的地区提升速度较快,存在落后者对先进者的“追赶效应”,各地区发展趋势趋同;Tobit回归分析结果表明,三大经济带的关键性驱动因素为政府资助强度,除此之外,教育水平、产业结构对于区域科技人才的开发效率提高也有重要的作用,因此政府应当加大资助强度,提升区域教育水平,促进产业结构升级,进而提高区域人才开发效率。

基于MLS形函数处理边界的电阻率法有限单元正演模拟

电阻率法有限单元正演模拟中,采用第三类边界条件时为保证精度仍要求较大范围的计算域.无单元法为地球物理领域的新兴正演模拟方法,其计算效率低,但其中采用的移动最小二乘(MLS)形函数相比于有限单元法形函数具有良好的连续性,模拟精度高.本文将MLS形函数应用于电阻率法有限单元2.5维正演的第三类边界条件处理,提出电阻率法有限单元-移动最小二乘(FEM-MLS)耦合正演方法.通过不同正演方法的模型算例模拟结果对比,验证了本文算法的有效性,并讨论了各个参数选择对模拟结果的影响.本文数值模拟结果表明采用第三类边界条件时,在同等计算精度前提下,FEM-MLS耦合法相比于有限单元法可进一步缩小计算域并提高了计算效率,相比于采用较大计算域满足边界条件的有限单元法计算效率提高了约一倍,相比于采用相同小范围计算域的有限单元法平均精度提高了约一倍.

WOMBAT—A tool for mixed model analyses in quantitative genetics by restricted maximum likelihood (REML)

WOMBAT is a software package for quantitative genetic analyses of continuous traits, fitting a linear, mixed model; estimates of covariance components and the resulting genetic parameters are obtained by restricted maximum likelihood. A wide range of models, comprising numerous traits, multiple fixed and random effects, selected genetic covariance structures, random regression models and reduced rank estimation are accommodated. WOMBAT employs up-to-date numerical and computational methods. Together with the use of efficient compilers, this generates fast executable programs, suitable for large scale analyses. Use of WOMBAT is illustrated for a bivariate analysis. The package consists of the executable program, available for LINUX and WINDOWS environments, manual and a set of worked example, and can be downloaded free of charge from http://agbu. une.edu.au/~kmeyer/wombat.

Studies on unfolding energy spectra of neutrons using maximumlikelihood expectation–maximization method

Energy spectra of neutrons are important for identification of unknown neutron sources and for determination of the equivalent dose. Although standard energy spectra of neutrons are available in some situations, e.g., for some radiotherapy treatment machines, they are unknown in other cases, e.g., for photoneutrons created in radiotherapy rooms and neutrons generated in nuclear reactors. In situations where neutron energy spectra need to be determined, unfolding the required neutron energy spectra using the Bonner sphere spectrometer (BSS) and nested neutron spectrometer (NNS) has been found promising. However, without any prior knowledge on the spectra, the unfolding process has remained a tedious task. In this work, a standalone numerical tool named ‘‘NRUunfold’’ was developed which could satisfactorily unfold neutron spectra for BSS or NNS, or any other systems using similar detection methodology. A generic and versatile algorithm based on maximum-likelihood expectation– maximization method was developed and benchmarked against the widely used STAY’SL algorithm which was based on the least squares method. The present method could output decent results in the absence of precisely calculated initial guess, although it was also remarked that employment of exceptionally bizarre initial spectra could lead to some unreasonable output spectra. The neutron count rates computed using the manufacturer’s response functions were used for sensitivity studies. The present NRUunfold code could be useful for neutron energy spectrum unfolding for BSS or NNS applications in the absence of a precisely calculated initial guess.

Joint polarization and DOA estimation based on improved maximum likelihood estimator and performance analysis for conformal array

The conformal array can make full use of the aperture,save space,meet the requirements of aerodynamics,and is sensitive to polarization information.It has broad application prospects in military,aerospace,and communication fields.The joint polarization and direction-of-arrival(DOA)estimation based on the conformal array and the theoretical analysis of its parameter estimation performance are the key factors to promote the engineering application of the conformal array.To solve these problems,this paper establishes the wave field signal model of the conformal array.Then,for the case of a single target,the cost function of the maximum likelihood(ML)estimator is rewritten with Rayleigh quotient from a problem of maximizing the ratio of quadratic forms into those of minimizing quadratic forms.On this basis,rapid parameter estimation is achieved with the idea of manifold separation technology(MST).Compared with the modified variable projection(MVP)algorithm,it reduces the computational complexity and improves the parameter estimation performance.Meanwhile,the MST is used to solve the partial derivative of the steering vector.Then,the theoretical performance of ML,the multiple signal classification(MUSIC)estimator and Cramer-Rao bound(CRB)based on the conformal array are derived respectively,which provides theoretical foundation for the engineering application of the conformal array.Finally,the simulation experiment verifies the effectiveness of the proposed method.

基于AutoML的低频波浪载荷智能预报方法

为了更加高效和准确地计算船舶在不同航速、海浪周期、波高及浪向角下任意剖面位置的低频波浪载荷响应,将贝叶斯优化算法和深度神经网络结合建立基于自动化机器学习(Auto ML)的波浪载荷智能预报方法。将拉丁超立方抽样(LHS)、数据处理与分析、切片理论和深度神经网络结合建立低频波浪载荷智能预报(LFWLIP)方法,探讨LHS方法和数据处理方法对LFWLIP模型计算精度的影响。利用基于Auto ML的低频波浪载荷智能预报方法对某船进行应用,通过优化得到自适应的网络模型超参数。结果表明:贝叶斯算法能够自动优化LFWLIP模型的损失函数,机器学习无需人工调参即可得到最优的批量大小、学习率、激活函数和隐含层结构。经验证,基于Auto ML的低频波浪载荷智能预报方法能够高精度地评估全域空间中的船舶低频波浪载荷。

基于MLS的三维扩展目标PMBM跟踪算法

针对传统三维扩展目标跟踪算法形状估计精度低的问题,提出了一种基于移动最小二乘的泊松多伯努利混合(Poisson multi-Bernoulli mixture based on the moving least square,MLS-PMBM)滤波跟踪算法。该算法基于MLS模型构建三维扩展目标的形状矩阵,通过PMBM滤波器预测和更新目标的运动状态,利用移动最小二乘算法更新形状矩阵,结合目标质心状态与形状估计完成对三维扩展目标的跟踪。仿真实验与实际点云数据的验证表明,与现有算法相比,本文所提算法在多扩展目标的形状估计方面具有更优的性能,具有较高的泛用性。