G&V: Medium: Collaborative Research: Large Data Visualization Using An Interactive Machine Learning Framework

摘要-Mchiraju、Rangarajan和Thompson随着计算机能力的不断增强，模拟的复杂性也随之增加，从而产生了前所未有的数据集。如果没有有效的分析工具，这些大规模模拟的结果就不能得到最充分的利用。这项研究通过使用交互式多尺度机器学习来解决大数据可视化和探索的问题，它利用了一种高效的基于特征的多分辨率数据表示。研究人员正在利用机器学习领域的方法来执行两项不同的任务：识别感兴趣的区域和增强特征检测算法的稳健性。这一努力的主要成果是实现了探索大型数据集的框架。此外，这项工作正在将机器学习中的大量工作引入可视化领域。这项研究的成功完成将有助于克服现有可视化方法的脆弱性，并在许多科学和工程领域促进权宜之计的发现。这里开发的多分辨率技术将采用双重策略。首先，基于与领域专家的训练的半监督学习被用于制定选择性地对数据进行空间和时间细化的策略。构造分类器以将粗分辨率特征检测(即区域)的输出标记为感兴趣或不感兴趣。然后，在最精细的尺度上，识别包含感兴趣的特征的有趣的本地数据块以供进一步分析。其次，使用自适应增强(AdaBoost)将几个局部特征检测算法或弱分类器组合成一个更健壮的复合分类器，以及一个称为CAVIAR的数据自适应变体，该变量有助于验证特征检测。理想情况下，复合分类器结合了所有弱分类器中最好的，因为它们响应潜在的物理信号。这项研究通过应用现有的局部检测算法来显示湍流流场中的涡旋，从而证明了这些方法的有效性。

Abstract - Machiraju, Rangarajan, and ThompsonAs computer power continues to increase, the complexity of simulations also increases thereby producing datasets of unprecedented size. Without effective analysis tools, results from these large-scale simulations cannot be utilized to their fullest extent. This research addresses the problem of large-data visualization and exploration by employing interactive multi-scale machine learning, which exploits an efficient feature-based, multi-resolution representation of the data. The investigators are leveraging methods from the field of machine learning to perform two distinct tasks: identify regions of interest and enhance robustness of feature detection algorithms. The primary outcome of this effort is the realization of a framework for exploring large datasets. Further, this work is introducing a large body of work in machine learning to the field of visualization. Successful completion of this research will help overcome the brittleness of existing visualization methods and foster expedient discovery in many areas of science and engineering.The multi-resolution techniques developed here will employ a two-fold strategy. First, semi-supervised learning based on training with the domain expert is used to develop strategies for selective spatial and temporal refinement of the data. A classifier is constructed to tag the output of the coarse resolution feature detection (i.e. regions) as either interesting or not interesting. Then at the finest scale, interesting local data chunks containing features of interest are identified for further analysis. Second, several local feature detection algorithms, or weak classifiers, are combined into a single, more robust compound classifier using adaptive boosting, or AdaBoost, and a data adaptive variant called CAVIAR that facilitates validated feature detection. Ideally, the compound classifier combines the best of all weak classifiers as they respond to the underlying physical signal. This research is demonstrating the effectiveness of these methods by applying existing local detection algorithms for visualizing vortices in turbulent flow fields.