SHF: Small: Learning Circuit Networks from Measurements

SHF：小型：从测量中学习电路网络

• 批准号: 2205572
• 负责人: Zhuo Feng
• 金额: $ 55万
• 依托单位: Stevens Institute of Technology
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-10-01 至 2025-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2205572&HistoricalAwards=false
• 关键词: SHF; Small; Learning; Circuit; Networks

Recent graph-learning techniques exploit both graph topological properties and node-feature (attribute) information to achieve promising results for various important applications such as vertex (data) classification, link prediction (recommendation systems), community detection, drug discovery, partial differential equation (PDE) solvers, and electronic design automation (EDA). On the other hand, graph learning can still be extremely challenging when there exists only partial or no knowledge about the underlying graph topologies. Fortunately, recent research shows that it is possible to learn graph topologies from node-feature (attribute) data so that existing graph-learning algorithms can be applied subsequently. However, even the state-of-the-art graph-topology-learning algorithms do not scale to large data sets due to their high computational complexity, which may prohibit their applications in real-world large-scale graph-learning tasks, such as those adopted for integrated-circuit networks involving billions of components. This project is also likely to spark new research in many other related fields, such as complex system/network modeling, model order reduction, computational biology, precision medicine, and transportation networks. The source code of the developed algorithms will be released on a public website managed by the PI to facilitate technology transfers to the industry, especially to the leading EDA companies. This research project will investigate highly scalable yet sample-efficient spectral methods for learning graph topologies from potentially high-dimensional data samples, such as voltage and current measurements in circuit networks. The proposed approach is based on a novel spectral-graph densification framework to allow for more efficient estimations of attractive Gaussian Markov Random Fields (GMRFs). A unique property of the learned graphs is that the effective-resistance distances on the learned graph will encode the similarities between the original data samples. The success of this research plan will immediately lead to the development of more scalable data-driven, physics-informed EDA algorithms for modeling, simulation, optimization, and verification of integrated circuits (ICs). The accomplished theoretical results will likely advance state of the art in spectral graph theory, dimensionality reduction, scientific computation, data visualization, and machine learning (ML).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

最近的图学习技术利用图的拓扑属性和节点特征(属性)信息来获得各种重要的应用，如顶点(数据)分类、链接预测(推荐系统)、社区检测、药物发现、偏微分方程(PDE)解算器和电子设计自动化(EDA)。另一方面，当只有部分或根本没有关于底层图拓扑的知识时，图学习仍然是极具挑战性的。幸运的是，最近的研究表明，从节点特征(属性)数据中学习图拓扑是可能的，以便随后可以应用现有的图学习算法。然而，即使是最先进的图拓扑学习算法，由于其高度的计算复杂性，也不能扩展到大数据集，这可能会阻碍它们在现实世界的大规模图学习任务中的应用，例如涉及数十亿个元件的集成电路网络所采用的任务。该项目还可能在许多其他相关领域引发新的研究，如复杂系统/网络建模、模型降阶、计算生物学、精确医学和交通网络。开发的算法的源代码将在PI管理的公共网站上发布，以促进向行业，特别是领先的EDA公司转让技术。这项研究项目将研究从潜在的高维数据样本中学习图拓扑的高可伸缩性和样本效率的谱方法，例如电路网络中的电压和电流测量。该方法基于一种新的谱图增密框架，以允许更有效地估计有吸引力的高斯马尔可夫随机场(GMRF)。学习图的一个独特性质是，学习图上的有效电阻距离将编码原始数据样本之间的相似性。这一研究计划的成功将立即导致更具可扩展性的数据驱动、物理信息的EDA算法的开发，用于集成电路(IC)的建模、仿真、优化和验证。已完成的理论成果可能会促进谱图理论、降维、科学计算、数据可视化和机器学习(ML)的最新水平。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

SF-SGL: Solver-Free Spectral Graph Learning From Linear Measurements

• DOI: 10.1109/tcad.2022.3198513
• 发表时间: 2023-02
• 期刊: IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
• 影响因子: 2.9
• 作者: Ying Zhang;Zhiqiang Zhao;Zhuo Feng

SF-GRASS: Solver-Free Graph Spectral Sparsification

• DOI: 10.1145/3400302.3415629
• 发表时间: 2020-08
• 期刊: 2020 IEEE/ACM International Conference On Computer Aided Design (ICCAD)
• 作者: Ying Zhang;Zhiqiang Zhao;Zhuo Feng

GARNET: Reduced-Rank Topology Learning for Robust and Scalable Graph Neural Networks

• 发表时间: 2022-01
• 期刊: ArXiv
• 作者: Chenhui Deng;Xiuyu Li;Zhuobo Feng;Zhiru Zhang