EAGER: Novel sampling algorithms for scaling up spectral methods for unsupervised learning

EAGER：用于扩大无监督学习光谱方法的新型采样算法

基本信息

批准号：
1650080
负责人：
Claire Monteleoni
金额：
$ 9万
依托单位：
George Washington University
依托单位国家：
美国
项目类别：
Standard Grant
财政年份：
2016
资助国家：
美国
起止时间：
2016-08-15 至 2018-07-31
项目状态：
已结题

来源：
https://www.nsf.gov/awardsearch/showAward?AWD_ID=1650080&HistoricalAwards=false
关键词：
EAGER Novel sampling algorithms scaling

项目摘要

In the era of big data, unsupervised learning has become increasingly important. At a high-level, unsupervised learning serves to reduce the data size, while capturing its important underlying structure. For a powerful and widely-used family of unsupervised learning techniques (those based on spectral methods), scaling up to large data sets poses significant computational challenges. This research project will develop extremely simple and lightweight sampling techniques for scaling up this family of unsupervised learning methods. Since big data is ubiquitous, these research advances are likely to be transformative to a range of fields. This project will benefit society through the research team's ongoing collaborations in climate science, agriculture, and finance. The team will also continue to engage the computer science community in this endeavor, by training students, developing tutorials, and broadening the participation of women and minorities in computing.This project will advance machine learning research by scaling up spectral methods for the analysis of large data sets. While spectral methods for the unsupervised learning tasks of clustering and embedding have found wide success in a variety of practical applications, scaling them up to large data sets poses significant computational challenges. In particular, the storage and computation needed to handle the affinity matrix (a matrix of pairwise similarities between data points) can be prohibitive. An approach that has found promise is to instead approximate this matrix in some sense. The goal of this project is to provide simple approximation techniques that manage the tradeoff between their space and time complexity vs. the quality of the approximation. The proposed approach involves sampling techniques that address this goal by exploiting latent structure in a data set, in order to minimize the amount of information that needs to be stored to (approximately) represent it. This leads to techniques that speed up the computation and reduce the memory requirements of spectral methods, while simultaneously providing better approximations. The project will also continue the team's momentum on leveraging advances in machine learning for data-driven discovery.

在大数据时代，无监督的学习变得越来越重要。在高级，无监督的学习中，学习可以减少数据大小，同时捕获其重要的基础结构。对于强大且普遍使用的无监督学习技术（基于光谱方法的家庭），扩展到大型数据集会带来重大的计算挑战。该研究项目将开发出极为简单且轻巧的抽样技术，以扩展这种无监督的学习方法。由于大数据无处不在，因此这些研究进步可能会变成一系列领域。该项目将通过研究团队在气候科学，农业和金融方面的持续合作来使社会受益。该团队还将通过培训学生，开发教程并扩大妇女和少数群体参与计算机的参与来继续与计算机科学界的参与。该项目将通过扩大频谱方法来扩展机器学习研究，以分析大型数据集。尽管无监督的学习和嵌入学习任务的光谱方法在各种实际应用中都取得了广泛的成功，但将它们扩展到大型数据集却带来了重大的计算挑战。特别是，处理亲和力矩阵所需的存储和计算（数据点之间成对相似性的矩阵）可能是令人难以置信的。一种发现承诺的方法是在某种意义上近似该矩阵。该项目的目的是提供简单的近似技术，以管理其空间和时间复杂性与近似质量之间的权衡。所提出的方法涉及采样技术，通过在数据集中利用潜在结构来解决该目标，以最大程度地减少需要存储（大约）代表它的信息的量。这导致技术可以加快计算并减少光谱方法的内存要求，同时提供更好的近似值。该项目还将继续在为数据驱动的发现方面利用机器学习方面的进步方面的势头。