基于训练字典的压缩感知光谱稀疏化方法

压缩感知理论利用目标的稀疏特性,能从极少的测量值中重构出目标图像,已成为突破奈奎斯特采样定理,实现超分辨成像的一个极具潜力的研究方向,其应用于对地观测遥感成像的一个核心问题在于面对复杂的地物场景,如何探求有效的稀疏化表达方法。对于具有超高数据量的高光谱成像而言,充分利用波段间丰富的冗余光谱信息,研究有效的光谱稀疏化表达方法更加具有实用价值。首先介绍了压缩感知光谱成像以及光谱稀疏化表达的基本原理,然后利用来自ASTER光谱库的多种类型地物光谱数据构建了一种基于K-SVD方法的训练字典,将其与DCT基、小波基分别作为稀疏基,对于几种典型地物目标进行仿真重构,结果表明:所构建的稀疏字典在采样数较少的情况下明显优于DCT基和小波基,在20%的低采样率时即可近乎完美地重构光谱曲线。     

Bipartite Matching in the Semi-streaming Model

We present the first deterministic 1+ε approximation algorithm for finding a large matching in a bipartite graph in the semi-streaming model which requires only O((1/ε)⁵) passes over the input stream. In this model, the input graph G=(V,E) is given as a stream of its edges in some arbitrary order, and storage of the algorithm is bounded by O(npolylog?n) bits, where $n = \lvert {V}\rvert $ . The only previously known arbitrarily good approximation for general graphs is achieved by the randomized algorithm of McGregor (Proceedings of the International Workshop on Approximation Algorithms for Combinatorial Optimization Problems and Randomization and Computation, Berkeley, CA, USA, pp. 170–181, 2005), which uses Ω((1/ε)^1/ε ) passes. We show that even for bipartite graphs, McGregor’s algorithm needs Ω(1/ε) ^Ω(1/ε) passes, thus it is necessarily exponential in the approximation parameter. The design as well as the analysis of our algorithm require the introduction of some new techniques. A novelty of our algorithm is a new deterministic assignment of matching edges to augmenting paths which is responsible for the complexity reduction, and gets rid of randomization. We repeatedly grow an initial matching using augmenting paths up to a length of 2k+1 for k=?2/ε?. We terminate when the number of augmenting paths found in one iteration falls below a certain threshold also depending on k, that guarantees a 1+ε approximation. The main challenge is to find those augmenting paths without requiring an excessive number of passes. In each iteration, using multiple passes, we grow a set of alternating paths in parallel, considering each edge as a possible extension as it comes along in the stream. Backtracking is used on paths that fail to grow any further. Crucial are the so-called position limits: when a matching edge is the ith matching edge in a path and it is then removed by backtracking, it will only be inserted into a path again at a position strictly lesser than i. This rule strikes a balance between terminating quickly on the one hand and giving the procedure enough freedom on the other hand.     

基于输入特征稀疏化的图神经网络训练加速

图神经网络(graph neural network, GNN)是处理图数据的重要方法. 由于计算复杂、图数据容量大, 在大规模图上训练图神经网络依赖于CPU-GPU协作和图采样训练方法, 其中图结构和特征数据存储在CPU内存中, 而采样得到的子图及其特征则传输至GPU进行训练. 然而, 这种方法面临着严重的图特征数据加载瓶颈, 显著降低了端到端训练性能, 且图特征占用过多内存, 严重限制了可训练的图规模. 为了解决这些问题, 我们提出了基于输入特征稀疏化的数据加载方法, 显著减少CPU内存占用和跨PCIe总线传输的数据量, 大幅缩短数据加载时间, 加速GNN的训练, 使其可以充分利用GPU计算资源. 针对图特征和GNN计算特性, 我们提出了适用于图特征数据的稀疏化方法, 在压缩比和模型准确度之间达到平衡. 我们在3个常见GNN模型和3个不同规模的数据集上进行了实验评估, 包括最大的公开数据集之一MAG240M. 结果表明, 此方法将特征尺寸减小了一个数量级以上, 并实现1.6–6.7倍的端到端训练加速, 而模型准确度的降低不超过1%. 此外, 在仅使用4个GPU的情况下, 仅需40 min就可以在MAG240M上完成GraphSAGE模型的训练并达到目标准确度.     

面向复杂网络的图稀疏算法综述

大规模数据下复杂网络的算法分析面临复杂度高的挑战,为此引入图稀疏的思想,在保持原始图性质的情况下以一定的精度在稀疏图上实现了高效的算法分析。图稀疏算法是一种保留顶点、对边稀疏采样的方法。按照相应算法分析所需要的原始图性质,提出图稀疏的边度量方式。文中系统回顾了4种边度量下的图稀疏采样方法:生成图稀疏、边连通图稀疏、聚类图稀疏、边传播性图稀疏,归纳了不同边度量方式下图稀疏的优缺点和适应性,并进一步讨论了动态图流稀疏化的最新研究进展。最后,总结了图稀疏领域有待解决的问题并展望了未来的研究方向。     

基于稀疏化的快速扩展信息滤波SLAM算法

随着SLAM技术的不断发展,计算效率已经成为制约SLAM发展的主要因素.所提出的算法从稀疏化的角度对扩展信息滤波SLAM算法进行改进.根据信息矩阵几乎稀疏的特点,该算法在合理稀疏化信息矩阵的同时利用环闭合检测技术,不仅大大提高了算法的计算效率,而且所得到的估计结果也很精确.通过仿真对信息矩阵稀疏化、算法效率、重定位以及误差和协方差四个关键问题进行了分析.分别就室内具有摄像头的两轮机器人和室外具有激光雷达的四轮机器人的情况进行了实验讨论.仿真与实验结果表明了所提算法的有效性.     

Sparsification from dilute connectivity in a neural network model of memory.

Several hypotheses concerning implementations of associative memory in the brain rely on analyses of the capabilities of simple network models. However, the low connectivity of cerebral networks imposes constraints which sometimes do not arise clearly from such analyses. We investigate an aspect of a simple, dilute network's operation that is sometimes overlooked, namely the setting of activation thresholds. An examination of several criteria for optimal threshold assignment affords several new insights. It becomes apparent that the network's capacity (which is simply derived) is insufficient to characterize the quality of its performance. We derive the degree of 'sparsification' or decrease in firing probability that arises from dilution, and also the consequent losses in representational ability, and propose that they should also be taken into account. To evaluate the model's performance and suitability, we argue that one should explicitly consider the trade-off that exists between storage of patterns and preservation of information, and its consequent constraints.     

一种基于边松弛的大规模WSN分簇定位算法

针对大规模WSN定位问题中,基于半定规划的分簇定位算法在分簇不均匀及节点密度较大时,部分簇会出现定位计算复杂度过高的问题,提出了一种新的基于边松弛的分簇定位算法-EES-Cluster.该算法通过对每一个网络簇子图进行边的松弛预处理,减少了边的数目;在网络分簇数目较少时,能有效降低定位过程的计算复杂度,同时较好地保持较高的定位精度,减小簇头节点信息融合的功耗.仿真实验及分析表明,EES-Cluster算法能有效降低分簇定位算法的计算复杂度,提高大规模WSN的定位效率.     

MR-GSpar:一种基于MapReduce的大图稀疏化算法

图的稀疏化是图聚类分析中数据预处理的关键操作,已得到广泛的关注.针对图数据日益普及、规模不断增大的现状,提出了一种基于MapReduce的面向大规模图的稀疏化算法,即MR-GSpar算法.该算法在MapReduce并行计算框架的基础上,通过对传统的最小哈希(Minhash)算法的并行化改造,使其可在分布式的集群环境中实现对大规模图数据的高效稀疏化处理.真实数据集上的实验表明了该算法的可行性与有效性.     

压缩感知中提升测量矩阵稀疏性的等效字典优化设计

在压缩感知中,可以通过减小等效字典（测量矩阵和稀疏字典的乘积）的互相干性值来提升稀疏重构算法的稳定性。已有的优化设计方法在减小等效字典互相干性值的同时没有考虑如何提高信号重构的计算效率,为了克服该问题,在稀疏字典固定的情形下,本文提出了一个关于测量矩阵的有约束光滑优化问题,其中第1个约束要求等效字典的Gram矩阵具有尽可能小的互相干性值;第2个则利用L₁范数来促进测量矩阵的稀疏性。然后,利用收敛的交替投影算法进行求解。数值实验表明：针对图像恢复问题,相对于采用已有优化设计方法得到的等效字典,本文提出的方法显著提高了测量矩阵中的零元素占比,同时使得压缩感知系统具有更高的信号重构精度。     

最小二乘支持向量机分类的稀疏化方法研究

最小二乘支持向量机在提高求解效率的同时,会丧失解的稀疏性,导致其在预测新样本时速度较慢。为此,提出一种稀疏化最小二乘支持向量机分类算法。在特征空间中寻找近似线性无关向量组,构造分类判别函数的稀疏表示,相应的最小二乘支持向量机优化问题可以通过线性方程组求解,从而得到最优判别函数。实验结果表明,该算法在不损失分类精度的前提下,能够获得比最小二乘支持向量机更快的预测速度。     

自动控制系统的参数整定及自整定

本文论述了参数整定在自动控制系统中的重要性，介绍了几种参数整定方法，并着重叙述了经验法的优点及实验方法，同时介绍了自整定的基本实施方法及其研究应用情况。     

我国典型地物标准波谱数据库数据管理子系统设计与开发

地物标准波谱数据库中的地面测量的波谱数据为高光谱数据,而一条波谱数据不仅包括各个波段和反射率的信息,还有数据说明及配套参数的信息。我国典型地物标准波谱数据库中存储了大量的地面测量数据,如何对这些海量数据进行有效的管理并呈现出来,是波谱库维护工作所必须解决的问题。该文主要介绍和探讨了如下几个问题:波谱数据的流程;采用回收站的机制在库中管理单条和成批的波谱数据;对于一条波谱数据的数据说明和配套参数信息一次不能彻底删除的问题采取定期清除库中的冗余数据的办法解决;统计库中数据及其使用情况;采用多种方式完整呈现一条波谱及配套参数信息。并编程在Oracel9i、Windows2000Server和IIS平台上用Asp实现。     

Setting the MIS Course

Designing reliable, flexible user interfaces can be more of a challenge than it first appears to be. To provide a system that can be used by many and still be responsive to individual needs is an elusive goal. This column discusses the real need that exists today for a truly user-friendly interface.     

基于地标表示的联合谱嵌入和谱旋转的谱聚类算法

经典的谱聚类算法包含两个步骤.(1)谱嵌入过程:求解Laplacian矩阵的特征值分解,得到分类指示矩阵的连续松弛解.(2)后处理过程:对谱嵌入连续松弛矩阵应用k-means或者谱旋转,得到最终的二值指示矩阵.由于有用信息的丢失,这种单独求解步骤不能保证最佳聚类结果.同时,谱聚类算法在处理大规模数据集时,存在聚类精度低...     

Spectral variations in the near-infrared ocean reflectance

The optical properties of natural waters beyond the visible range, in the near-infrared (NIR, 700-900 nm), have received little attention because they are often assumed to be mostly determined by the large absorption coefficient of pure water, and because of methodological difficulties. It is now growingly admitted that the NIR represents a potential optical source of unambiguous information about suspended sediments in turbid waters, thence the need for better understanding the NIR optical behaviour of such waters. It has recently been proposed (Ruddick et al., Limnology and Oceanography. 51, 1167-1179, 2006) that the variability in the shape of the surface ocean reflectance spectrum in the NIR is negligible in turbid waters. In the present study, we show, based on both in situ and remote sensing data, that the shape of the ocean reflectance spectrum in the NIR does vary in turbid to extremely turbid waters. Space-borne ocean reflectance data were collected using 3 different sensors (SeaWiFS, MODIS/Aqua and MERIS) over the Amazon, Mackenzie and Rio de la Plata turbid river plumes during extremely clear atmospheric conditions so that reliable removal of gas and aerosol effects on reflectance could be achieved. In situ NIR reflectance data were collected in different European estuaries where extremely turbid waters were found. In both data sets, a flattening of the NIR reflectance spectrum with increasing turbidity was observed. The ratio of reflectances at 765 nm and 865 nm, for instance, varied from ca. 2 down to 1 in our in situ data set, while a constant value of 1.61 had been proposed based on theory in a previous study. Radiative transfer calculations were performed using a range of realistic values for the seawater inherent optical properties, to determine the possible causes of variations in the shape of the NIR reflectance spectrum. Based on these simulations, we found that the most significant one was the gradual increase in the contribution of suspended sediments to the color of surface waters, which often leads to the flattening of the reflectance spectrum. Changes in the scattering and absorption properties of particles also contributed to variations in the shape of the NIR surface ocean reflectance spectrum. The impact of such variations on the interpretation of ocean color data is discussed.