BIGDATA: Collaborative Research: F: From Data Geometries to Information Networks

BIGDATA：协作研究：F：从数据几何到信息网络

• 批准号: 1546206
• 负责人: Leonidas Guibas
• 金额: $ 50万
• 依托单位: Stanford 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: 2016-01-01 至 2019-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1546206&HistoricalAwards=false
• 关键词: BIGDATA; Collaborative; Research; Data; Geometries

Big Data often results from multiple sources, giving collections that contain multiple, often partial, "views" of the same object, space, or phenomenon from various observers. Extracting information robustly from such data sets calls for a joint analysis of a large collection of data sets. The project is developing a novel geometric framework for modeling, structure detection, and information extraction from a collection of large related data sets, with an emphasis on the relationships between data. While this approach clearly applies to data with a clear geometric character (e.g., objects in images), the work is also applied to datasets as diverse as computer networks (identifying common structure in subnets) and Massive Open Online Course homework data (automatically carrying grader annotations to similar problems in other students' homeworks).The novel framework is based on the construction of maps between the objects under considerations (point clouds, graphs, images, etc...), and on the analysis of the networks of maps that result as a way of extracting information, generating latent models for the data, and transporting or inferring functional / semantic information. These tasks define a new field of map processing between data sets and require tool sets with new ideas from functional analysis, non-convex optimization, and homological algebra in mathematics, and geometric algorithms, machine learning, optimization, and approximation algorithms in computer science. Sophisticated algorithmic techniques for attacking the large-scale non-linear optimization problems that emerge within the framework will also be investigated.

大数据通常来自多个来源，提供包含来自不同观察者的同一对象、空间或现象的多个（通常是部分的）“视图”的集合。从这些数据集中健壮地提取信息需要对大量数据集进行联合分析。该项目正在开发一种新的几何框架，用于建模、结构检测和从大量相关数据集中提取信息，重点是数据之间的关系。虽然这种方法显然适用于具有明确几何特征的数据（例如，图像中的对象），但这项工作也适用于各种数据集，如计算机网络（识别子网中的共同结构）和大规模开放在线课程作业数据（自动携带评分者对其他学生作业中类似问题的注释）。这个新框架是基于在考虑的对象（点云、图形、图像等）之间构建地图，以及对地图网络的分析，这些网络作为提取信息、为数据生成潜在模型以及传输或推断功能/语义信息的一种方式。这些任务定义了数据集之间地图处理的新领域，并且需要具有新思想的工具集，这些思想来自数学中的泛函分析、非凸优化和同调代数，以及计算机科学中的几何算法、机器学习、优化和近似算法。复杂的算法技术，攻击大规模非线性优化问题，出现在框架内也将进行研究。

项目成果

Learning Transformation Synchronization

• DOI: 10.1109/cvpr.2019.00827
• 发表时间: 2019-01
• 期刊: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
• 作者: Xiangru Huang;Zhenxiao Liang;Xiaowei Zhou;Yao Xie;L. Guibas;Qi-Xing Huang

PeerNets: Exploiting Peer Wisdom Against Adversarial Attacks

• 发表时间: 2018-05
• 期刊: ArXiv
• 作者: Jan Svoboda;Jonathan Masci;Federico Monti;M. Bronstein;L. Guibas

Parsing Geometry Using Structure-Aware Shape Templates

使用结构感知形状模板解析几何

• DOI: 10.1109/3dv.2018.00082
• 发表时间: 2018
• 期刊: 2018 International Conference on 3D Vision (3DV
• 作者: Ganapathi-Subramanian, Vignesh;Diamanti, Olga;Pirk, Soeren;Tang, Chengcheng;Niessner, Matthias;Guibas, Leonidas
• 通讯作者: Guibas, Leonidas