SHF: Small: Program Analysis for Dependable Clustering

聚类分析，又称聚类，是一种机器学习技术，用于将相关或具有相似特征的实体分组在一起。聚类应用于医学、生物学、社会科学、机器人和地球科学，包括高风险领域，如医学图像处理或医学诊断、预测疾病相关基因或急性疾病的资源分配。然而，目前，包含集群的应用程序的用户或开发人员不能保证应用程序是可靠的；这引起了对使用聚类获得的结果或诊断的质疑，并阻碍了研究人员或从业者使用聚类应用程序。这个项目将使集群实现更可靠、更容易开发和更容易修复。软件工程师将受益于更容易编程/修复/检查的实现。反过来，最终的软件将更加可靠，使最终用户受益。该项目将向学生和IT专业人士介绍可靠机器学习的挑战和方法；这将使学生和专业人士更好地应对新兴的软件研究和开发挑战。将继续开展对少数民族和代表性不足群体的外联和支助工作。随着人们对机器学习越来越感兴趣，以及在软件驱动的产品中越来越多地采用机器学习实现（硬件或软件），集群的使用将会扩展。因此，集群实现必须是可靠的。这个领域的研究人员缺乏基本聚类正确性的定义，以及验证聚类实现属性的有效/高效分析。这个项目将通过从第一原则开始定义集群正确性来解决上述问题，例如，程序决定论；并构造了通过程序分析验证正确性的方法。方法将包括差分执行、白盒和黑盒技术、动态切片和符号执行。这项工作的范围包括“纯软件”集群实现以及使用硬件加速的实现。使用这些工具，开发人员和研究人员将能够有效地了解集群实现的行为和可靠性。研究人员将能够使用在这项工作中开发的一般原理来构建其他领域的程序分析，例如，科学计算，数值计算，高性能计算，并使用锁步执行，切片或符号执行其他类型的数据密集型应用程序的工具/方法。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Cluster analysis, a.k.a. Clustering, is a machine-learning technique used to group together entities that are related or share similar characteristics. Clustering has applications in medicine, biology, social sciences, robotics, and earth sciences, including high-stakes domains such as medical image processing or medical diagnosis, predicting disease-related genes, or resource allocation in acute disease. However, currently, users or developers of applications that incorporate clustering have no assurances that the applications are reliable; this calls into question results or diagnoses obtained with the use of clustering, and discourages researchers or practitioners from using clustering applications. This project will make clustering implementations more reliable, easier to develop, and easier to fix. Software engineers will benefit from implementations that are easier to program/fix/check. In turn, the resulting software will be more reliable, benefiting end-users. The project will introduce students and IT professionals to challenges in, as well as approaches for, dependable machine learning; this will make students and professionals better equipped for tackling emerging software research and development challenges. Ongoing outreach and support efforts, to minorities and underrepresented groups, will continue. Clustering use will expand with the increasing interest in machine learning in general, and increased adoption of machine-learning implementations (hardware or software) in software-driven products. Hence it is imperative that clustering implementations be dependable. Researchers in this space lack definitions of basic clustering-correctness properties and effective/efficient analyses for verifying clustering implementations' properties. This project will address the aforementioned issues by defining clustering correctness starting from first principles, e.g., program determinism; and constructing approaches for verifying correctness via program analysis. Approaches will include differential execution, white-box as well as black-box techniques, dynamic slicing, and symbolic execution. The scope of this work includes "purely software" clustering implementations as well as implementations that use hardware acceleration. Using these tools, developers and researchers will be able to gain effective insights into clustering-implementation behavior and dependability. Researchers will be able to use the general principles introduce developed in this work to construct program analyses in other domains, e.g., scientific computing, numerical computing, high-performance computing, and use the tools/approaches for lockstep-executing, slicing, or symbolically executing other categories of data-intensive applications.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.