耗散粒子动力学GPU并行计算研究

研究了耗散粒子动力学基于计算统一设备架构的图形处理器(GPU)并行计算的实施.对其中涉及的算法映射模型、Cell-List法数组的并行化更新、随机数生成、存储器访问优化、负载平衡等进行了详细的讨论.进一步模拟了Poiseuille流动和突扩突缩流动,从而验证了GPU计算结果的正确性.计算结果表明,相对于基于中央处理器的串行计算,在耗散粒子动力学中实施GPU并行计算可以获得约20倍的加速比.     

A proteome-wide analysis of domain architectures of prokaryotic single-spanning transmembrane proteins

We performed a proteome-wide survey of the domain architectures in single-spanning transmembrane (TM) proteins (single-spannings) from 87 sequenced prokaryotic (Bacterial and Archaean) genomes by assigning Pfam domains to their N-tail and C-tail loops. Out of 14,625 single-spannings, 3,516 sequences have at least one domain assigned, and no domains were assigned to 7,850, with the remaining 3,259 with less reliable assignment. In the domain-assigned sequences, 3116 sequences are with at most two domains, and the other 400 sequences with more than two. The assigned domains distribute over 651 Pfam families, which account for 11.4% of the total Pfam-A families. Among the 651 families are mostly soluble-protein-originated ones, but only 21 families are unique to TM proteins. The occurrence frequency of the individual domain families follows a power-law, that is, 264 families occur only once, 106 just twice, and the families appeared more than 30 times are counted by only 39. It is found that the great majority of the sequences having one or two domains are of the type II topology with the C-tail loop containing domains on it. On the contrary, the N-tail loop of the same type topology seldom carries domains. Importantly, the assigned domains are always found on the tail loops longer than 60 residues, even for the small domains with less than 30 residues. There are still as many as 5,800 sequences without assigned domains in spite of having at least one long tail, on which no less than 1,000 novel domain families are expected most likely to lie concealed unknown yet. We also investigated the domain arrangement preference and the domain family combination patterns in 'singlets' (single-spannings with one assigned domain) and 'doublets' (with two domains).     

Closed-form analysis for indirect learning digital pre-distorter for wideband wireless systems

Indirect learning architecture (ILA) for digital pre-distortion (DPD) is commonly used to linearize power amplifiers (PA). To the author’s best knowledge, most of the DPD results in the literature obtain the matrix form of the least-square solution in order to get the DPD coefficients numerically. There exists no explicit closed-form for these coefficients that can be used as plug-and-play in simulations, or used for further closed-form analysis of important measures such as signal-to-noise ratio (SNR) and mean square error (MSE), bit-error rate (BER), …etc. In this paper, we analyze the ILA-DPD system for general memory-polynomial PA models. We provide a closed-form solution for the DPD coefficients. We first present the analytical methodology for deriving the mathematical expressions for each DPD coefficient and then introduce an open-access code that generates the DPD coefficients in symbolic form that is used to mathematically model the DPD. We consider case studies for PA and show that the analytical DPD solution matches the Monte Carlo simulations. Moreover, we also provide a closed-form solution for the iterative adaptive ILA-DPD. Our analysis shows that in the case of a large training block length the non-iterative DPD achieves approximately the same performance as an iterative DPD with a shorter training block length. System impairments are also considered, e.g. the thermal noise and the quantization noise in analog–digital conversion (ADC). We derive the normalized mean square error (NMSE) for the transmit chain in the presence of these impairments. The NMSE expression is verified through numerical simulations.     

An Algorithm Based on Simple Rules for the Determination of the Architecture of Branched Polymer Molecules from the Number of Branching Points per Molecule

A general simple algorithm is proposed to determine the average architecture of an acyclic branched polymer macromolecule from its number of branching points, whatever their functionality. The number of branching points can be derived from SEC measurements using a coupling of viscosimetric and refractometric detectors through the use of the Zimm‐Stockmayer equation. Since various isomers can be built for a given number of branching points, some basic rules are proposed in order to select the structures that are physically meaningful; the possibility of structural isomers has been also investigated and a probability of presence is assigned to each species starting from the compactness of the molecule. For each species corresponding to a given molecular weight in the sample, the average structure is then obtained from the number of branching points and the universal probability distribution of the isomers.

     

中药复杂组效关系的变结构神经网络辨识方法

针对中药复杂组效关系的辨识问题,研究了变结构多层前馈神经网络,推导出一种新型的变结构网络学习算法,成功地应用于中药川芎药效活性预测计算.该方法从一个规模较小的网络出发,当网络无法达到预定的学习精度时,自动增加隐含层神经元个数,并在原有学习结果的基础上确定新的网络参数,自适应地确定前馈神经网络结构,可用于处理复杂化学模式信息.计算机仿真实验结果表明,该方法能有效地确定多层前馈神经网络的最佳结构,提高网络学习效率和函数逼近精度,解决复杂非线性函数映射关系准确建模问题.     

基于快速神经网络架构搜索的鲁棒图像水印网络算法

为解决深度学习在图像水印算法中计算量大且模型冗余的问题,提高图像水印算法在抵抗噪声、旋转和剪裁等攻击时的鲁棒性,提出基于快速神经网络架构搜索（neural architecture search,NAS）的鲁棒图像水印网络算法。通过多项式分布学习快速神经网络架构搜索算法,在预设的搜索空间中搜索最优网络结构,进行图像水印的高效嵌入与鲁棒提取。首先,将子网络中线性连接的全卷积层设置为独立的神经单元结构,并参数化表示结构单元内节点的连接,预先设定结构单元内每个神经元操作的搜索空间;其次,在完成一个批次的数据集训练后,依据神经元操作中的被采样次数和平均损失函数值动态更新概率;最后,重新训练搜索完成的网络。水印网络模型的参数量较原始网络模型缩减了92%以上,大大缩短了模型训练时间。由于搜索得到的网络结构更为紧凑,本文算法具有较高的时间性能和较好的实验效果,在隐藏图像时,对空域信息的依赖比原始网络更少。对改进前后的2个网络进行了大量鲁棒性实验,对比发现,本文算法在CIFAR-10数据集上对抵抗椒盐噪声和旋转、移除像素行（列）等攻击优势显著;在ImageNet数据集上对抵抗椒盐高斯噪声、旋转、中值滤波、高斯滤波、JPEG压缩、裁剪等攻击优势显著,特别是对随机移除行（列）和椒盐噪声有较强的鲁棒性。     

GPON体系结构研究

千兆比特无源光网络(GPON)是一种面向下一代网络(NGN)的接入技术,它具有速率高、传输距离长、效率高和可伸缩性强等关键优势.同时,GPON支持引入更高带宽的以太网新业务,并能低成本提供传统的语音服务.介绍了GPON的标准化进展,分析了GPON的系统结构和协议参考模型,并给出了传输汇聚层的分层结构.     

SpinStudioJ: A cross-platform NMR data acquisition and processing workbench based on a plug-in architecture

Flexibility and extensibility are important issues in the design of nuclear magnetic resonance (NMR) software, as these determine the ability to integrate a variety of continuously evolving data acquisition and processing methods. Here, SpinStudioJ is introduced. It is an NMR data acquisition and processing workbench with a plug-in-based architecture. The workbench is based on Eclipse Rich Client Platform, which provides a plug-and-play runtime mechanism and rich graphical user interface functionality. New data acquisition methods and processing algorithms can be easily integrated into the SpinStudioJ workbench by defining extension points, without the need to redistribute existing modules. The software is independent of operating systems, as it leverages the cross-platform feature of the Java virtual machine.     

云端结合的书法大数据平台

基于文化强国的国家重大战略，为满足传承和创新我国传统文化这一重要需求，实现文化休闲与共享、书法普及与教育的具体目标，研发了基于大数据和人工智能等核心技术建构的云端结合的书法大数据平台。平台连通了线下和线上的信息通道，实现古代经典与当下书写的智能对接、线上与线下作品的共创共享以及个体创作与群体创作的有机结合。在此基础上，实现了数字书法创作与交流、作品集的生成与推广、线上线下协同创作、书法作品的智能竞技、书法创作数据分析等形式的智能书法服务。