面向隐私保护联邦学习的医学影像目标检测算法

联邦学习技术是一种新型多机构协同训练模型范式,广泛应用于多领域,其中模型参数隐私保护是一个关键问题.针对CT影像综合性病灶检测任务,提出隐私保护的联邦学习算法.首先部署松散耦合的客户端-服务器架构;其次在各客户端使用改进的RetinaNet检测器,引入上下文卷积和后向注意力机制;最后完成联邦训练.各客户端使用局部更新策略,采用自适应训练周期,局部目标函数中加入了限制项;服务器使用自适应梯度裁剪策略和高斯噪声差分隐私算法更新全局模型参数.在DeepLesion数据集上的消融分析说明了算法各部分的重要性.实验结果表明,改进的RetinaNet检测器有效地提升了多尺度病灶的检测精度.与集中数据训练模型范式相比,联邦学习所得模型性能略低(mAP分别为75.33％和72.80％),但训练用时缩短近38％,有效地实现了隐私保护、通信效率和模型性能的良好权衡.     

RLINK: Deep reinforcement learning for user identity linkage

World Wide Web - User identity linkage is a task of recognizing the identities of the same user across different social networks (SN). Previous works tackle this problem via estimating the pairwise...     

基于稳健设计的SGNS算法的超参数调优方法

通常给定超参数的若干取值选取性能最大的为最优组合(称为直接选优法),但是此方法的稳健性差。为此,提出了一种基于稳健设计的超参数调优方法(称为稳健调优法)。具体地,以SGNS算法中的超参数调优为例,在词语推断任务上实验并得出:经方差分析得到SGNS算法中的七个超参数中的五个对算法预测性能有显著影响,确定为主控因子,其余两个确定为噪声因子,且主控因子中有三个对性能估计的方差有显著影响,因此,调优中仅从期望最大来直接选优是不合理的;稳健调优法与直接选优法两者在预测性能上没有显著差异,但稳健调优法对噪声因子具有较好的稳健性。稳健调优法对一般的深度神经网络的调参有实际的借鉴意义。     

Novel Apriori-Based Multi-Label Learning Algorithm by Exploiting Coupled Label Relationship

It is a key challenge to exploit the label coupling relationship in multi-label classification (MLC) problems. Most previous work focused on label pairwise relations, in which generally only global statistical information is used to analyze the coupled label relationship. In this work, firstly Bayesian and hypothesis testing methods are applied to predict the label set size of testing samples within their k nearest neighbor samples, which combines global and local statistical information, and then apriori algorithm is used to mine the label coupling relationship among multiple labels rather than pairwise labels, which can exploit the label coupling relations more accurately and comprehensively. The experimental results on text, biology and audio datasets shown that, compared with the state-of-the-art algorithm, the proposed algorithm can obtain better performance on 5 common criteria.     

Effect of DAPBO segment on the structure and performance enhancement of powder foamed BTDA-ODA polyimide

Achieving excellent thermal stability, good mechanical strength, and low-density in one product, is a big challenge in the design and synthesis of polyimide (PI) foams. We selected a rigid diamine monomer, 2-(4-aminophenyl)-5-aminobenzoxazole (DAPBO) to partially replace 4,4′-oxydianiline (ODA) in BTDA-ODA system, and studied the effect on properties of foamed PI products. With various molar ratio between DABPO and ODA (0:5, 1:4, 2:3, 3:2, 4:1, 5:0), the diamine(s) polymerized with 3,3′,4,4′-benzophenonetetracarboxylic dianhydride (BTDA) and produced six PI precursors (PAE-0 ~ PAE-5). After foaming and thermal imidization, corresponding PI foams (PI-0 ~ PI-5) prepared. Among them, PI-3 has a glass transition temperature of 386°C, which is 95°C higher than PI-0, and its compressive strength reaches 1.32 MPa. Moreover, PI-3 has a uniform cell structure with an average pore size of ~200 μm. The analysis results show that PI-3 is in a transition state from amorphous to partial crystalline, maintaining a balance between flexibility and rigidity. The addition of DABPO not only increases the rigidity of polymer molecular segments, but also introduces N and O atoms that can form intermolecular hydrogen bonds between molecules, which increases the packing density and arrangement regularity of polymer molecules.     

Effect of special structure of clam shell powder on structure and properties of castor oil-based composites

Castor oil (CO) is an environmentally friendly renewable green resource and ideal alternative to petroleum resources. The preparation of high strength and high toughness castor oil-based polyurethane prepolymer (COPU) composites has significant applications such as supporting material and engineering plastic sheet. In this study, unmodified clam shell powder (CSP) with a unique CaCO₃-proteoglycan structure was used as a filler to prepare compatible reinforced COPU composite materials. Investigation of the mechanical properties revealed that the elastic modulus of the composite COPU reinforced with 50 wt% of CSP had increased to 5859.0 ± 8.4 MPa representing 187.77% to obtain stiffer and stronger material over pure COPU (2036.6 ± 196.9 MPa). Moreover, the scanning electron microscopy, thermogravimetric analysis and contact angle results demonstrated that the reinforced COPU composites have better compatibility, thermal stability, and water resistance than pure COPU. This work will promote the application prospects of CO-based polyurethane.     

高压凝固对Al-Mg合金枝晶形态的影响

利用OLYMPUS光学显微镜研究了高压凝固对Al-9．6Mg合金枝晶形态的影响。结果表明，与常压下典型的破碎的粗大枝晶组织相比，高压凝固条件下，枝晶臂较完整，二次枝晶间距减小，并且大部分一次枝晶臂增长。从高压对成分过冷和热过冷的影响及溶质扩散角度探讨了枝晶形态的变化机理。高压这种极端凝固条件为改变合金的组织形态提供了一种新途径。     

不锈钢过钝化－二次钝化的研究

通过对３０４不锈钢在较高电位下过钝化和二次钝化的稳态过程和电化学交流阻抗（ＥＩＳ）的分析，认为不锈钢过钝化和二次钝化现象与膜／溶液（ｆ／ｓ）界面上的电化学反应有密切关系，并通过理论模型分析，解释了过钝化和二次过钝化膜在高电位下溶解速度高，而在开路自腐蚀电位下反而有较好的稳定性这一现象．     

奔驰重卡转向节挤压锻造复合工艺的有限元分析

奔驰重型卡车的转向节挤压锻造工艺，分为镦粗、挤压、预锻和终锻4个工步。针对汽车转向节的挤压锻造进行的有限元分析。模拟软件采用Deform 3D，锻件采用刚塑性材料模型，成形过程看作为等温成形。报告给出了毛坯形状和尺寸、锻件形状变化，成形等效应变分布以及载荷——行程曲线等。实验表明，模拟结果和实际成形过程吻合良好。     

体积成形数值模拟中利用螺旋压力机成形的能量法与速度法分析

体积成形模拟中的运动通常是通过对模具施加一定的速度来实现的，这种输入方式简单、直接，通常适用于普通压力机。而利用螺旋压力机成形的工艺有两种输入方式可供选择：速度法和能量法。本文针对体积成形数值模拟中螺旋压力机运动参数的输入方法进行研究，将与螺旋压力机成形特点结合更紧密的能量法引入模拟中，其目的是就螺旋压力机的不同运动输入形式对模拟结果的影响进行系统地分析、比较，概括出利用高能螺旋压力机成形时，利用两种不同输入法辅助工程设计的适用范围。