基于伪数据的机器翻译质量估计模型的训练

为向基于深度学习的机器翻译质量估计模型提供高效的训练数据, 提出面向目标数据集的伪数据构造方法, 采用基于伪数据预训练与模型精调相结合的两阶段模型训练方法对模型进行训练, 并针对不同伪数据规模设计实验。结果表明, 在构造得到的伪数据下, 利用两阶段训练方法训练得到的机器翻译质量估计模型给出的得分与人工评分的相关性有显著的提升。     

基于卷积核分解的深度CNN模型结构优化及其在小图像识别中的应用

小图像由于像素少、分辨率低、整幅图像包含信息较少,识别较为困难。目前优秀的深度卷积神经网络模型多为大图像而设计,而用于小图像的模型则存在着层次不够深、难以对特征进行充分抽象的不足。本文基于VGG19模型,依据卷积核分解的原理,设计了一种KDS-DCNN模型,模型深度达到31层,解决了目前超深度模型不能直接用于小图像识别的问题,实验表明该方法不但提升了识别性能,而且还降低了模型的时间复杂度。在CIFAR-10、CIFAR-100和SVHN三个数据集上的验证结果显示,KDS-DCNN模型性能优越,其识别错误率分别降低到29.46%、6.02%和2.17%。     

中心损失与Softmax损失联合监督下的人脸识别

深度学习在人脸识别领域已经取得了巨大的成就,针对当前大多数卷积神经网络采用Softmax损失函数进行特征分类,增加新的类别样本会减小类间距离的增长趋势,影响网络对特征判别的问题,采用了一种基于中心损失与Softmax损失联合监督的人脸识别算法,来提高网络对特征的识别能力。在Softmax基础上,首先,分别对训练集每个类别在特征空间维护一个类中心,训练过程新增加样本时,网络会约束样本的分类中心距离,从而兼顾了类内聚合与类间分离。其次,引入动量概念,在分类中心更新的时候,通过保留之前的更新方向,同时利用当前批次的梯度微调最终的更新方向,该方法可以在一定程度上增加稳定性,提高网络的学习效率。最后,在人脸识别基准库LFW上的测试实验证明:所提的联合监督算法,在较小的网络训练集上,获得了99.31%的人脸识别精度。     

堪为物理学第三分支的计算物理

本文是作者1982年在中国访问时所作的讲演,论述物理学各领域的前沿工作,用许多事例说明计算物理对推动物理理论的重大意义。文中还指出计算物理在其性质、方法及需要等方面不同于和独立于解析的理论物理和实验物理,而成为物理学的第三分支。     

Determination of hydroxysafflor yellow A in biological fluids of patients with traumatic brain injury by UPLC‐ESI‐MS/MS after injection of Xuebijing

A simple, novel, specific, rapid and reproducible ultra‐performance liquid chromatography electrospray ionization tandem mass spectrometry method has been developed and validated for the determination of hydroxysafflor yellow A (HSYA) in biological fluids (plasma, urine and cerebrospinal fluid) of patients with traumatic brain injury after intravenous injection of Xuebijing (XBJ). Liquid–liquid extraction was performed, and separation was carried out on an Acquity UPLC? BEH C₁₈ column, with gradient elution using a mobile phase composed of methanol and 0.1% formic acid at a flow rate of 0.3 mL/min. A triple quadrupole tandem mass spectrometer with electrospray ionization was used for the detection of HSYA. The mass transition followed was m/z 611.0 → 491. The retention time was less than 3.0 min. The calibration curve was linear in the concentration range from 2 to 6125 ng/mL for cerebrospinal fluid, plasma and urine. The intra‐ and inter‐day precisions were <10%, and the relative standard deviation of recovery was <15% for HSYA in biological matrices. The method was successfully applied for the first time to quantify HSYA in the biological fluids (especially in cerebrospinal fluid) of patients with traumatic brain injury following intravenous administration of XBJ. Copyright © 2014 John Wiley & Sons, Ltd.     

Validated UPLC‐MS/MS method for determination of moclobemide in human brain cell supernatant and its application to bidirectional transport study

A simple and sensitive analytical method based on ultraperformance liquid chromatography–tandem mass spectrometry (UPLC‐MS/MS) has been developed for determination of moclobemide in human brain cell monolayer as an in vitro model of blood–brain barrier. Brucine was employed as the internal standard. Moclobemide and internal standard were extracted from cell supernatant by ethyl acetate after alkalinizing with sodium hydroxide. The UPLC separation was performed on an Acquity UPLC^TM BEH C₁₈ column (50 × 2.1 mm, 1.7 µm, Waters, USA) with a mobile phase consisting of methanol–water (29.5:70.5, v/v); the water in the mobile phase contained 0.05% ammonium acetate and 0.1% formic acid. Detection of the analytes was achieved using positive ion electrospray via multiple reaction monitoring mode. The mass transitions were m/z 269.16 → 182.01 for moclobemide and m/z 395.24 → 324.15 for brucine. The extraction recovery was 83.0–83.4% and the lower limit of quantitation (LLOQ) was 1.0 ng/mL for moclobemide. The method was validated from LLOQ to 1980 ng/mL with a coefficient of determination greater than 0.999. Intra‐ and inter‐day accuracies of the method at three concentrations ranged from 89.1 to 100.9% for moclobemide with precision of 1.1–9.6%. This validated method was successfully applied to bidirectional transport study of moclobemide blood–brain barrier permeability. Copyright © 2013 John Wiley & Sons, Ltd.     

LC/MS/MS analysis of the endogenous dimethyltryptamine hallucinogens,their precursors,and major metabolites in rat pineal gland microdialysate

We report a qualitative liquid chromatography–tandem mass spectrometry (LC/MS/MS) method for the simultaneous analysis of the three known N,N‐dimethyltryptamine endogenous hallucinogens, their precursors and metabolites, as well as melatonin and its metabolic precursors. The method was characterized using artificial cerebrospinal fluid (aCSF) as the matrix and was subsequently applied to the analysis of rat brain pineal gland‐aCSF microdialysate. The method describes the simultaneous analysis of 23 chemically diverse compounds plus a deuterated internal standard by direct injection, requiring no dilution or extraction of the samples. The results demonstrate that this is a simple, sensitive, specific and direct approach to the qualitative analysis of these compounds in this matrix. The protocol also employs stringent MS confirmatory criteria for the detection and confirmation of the compounds examined, including exact mass measurements. The excellent limits of detection and broad scope make it a valuable research tool for examining the endogenous hallucinogen pathways in the central nervous system. We report here, for the first time, the presence of N,N‐dimethyltryptamine in pineal gland microdialysate obtained from the rat. Copyright © 2013 John Wiley & Sons, Ltd.     

Microwave Assisted Rupture of Furoxans to Nitrile Oxides and Intermolecular Capture by Dipolarophiles

3,4-Diaroylfuroxans 1 react with various dipolarophiles 3 under microwave activation to afford the cycloadduct 4 instead of the expected isoxazole 5 in good yields in the absence of solvent.     

Deep cleaning of a metallurgical zinc leaching residue and recovery of valuable metals

Huge quantities of zinc leaching residues (ZLRs) generated from zinc production are dumped continuously around the world and pose a potential environmental threat because of their considerable amounts of entrained heavy metals (mainly lead). Most ZLRs have not been properly treated and the valuable metals in them have not yet been effectively recovered. Herein, the deep cleaning of a ZLR and recovery of valuable metals via a hydrometallurgical route were investigated. The cleaning process consists of two essential stages:acid leaching followed by calcium chloride leaching. The optimum conditions for extracting zinc, copper, and indium by acid leaching were a sulfuric acid concentration of 200 g·L-1, a liquid/solid ratio of 4:1 (mL/g), a leaching time of 2 h, and a temperature of 90℃. For lead and silver extractions, the optimum conditions were a calcium chloride concentration of 400 g·L-1, a pH value of 1.0, a leaching time of 1 h, and a temperature of 30℃. After calcium chloride leaching, silver and lead were extracted out and the lead was finally recovered as electrolytic lead by electrowinning. The anglesite phase, which poses the greatest potential environmental hazard, was removed from the ZLR after deep cleaning, thus reducing the cost of environmental management of ZLRs. The treatment of chlorine and spent electrolyte generated in the process was discussed.