III: Small: Declarative Recursive Computation on a Database System

机器学习（ML）将在未来几十年内产生巨大的经济和科学影响。但是，尽管机器学习的重要性越来越大，但针对支持机器学习的系统应该是什么样子的研究相对较少。结果是很难将ML应用于非标准情况：大数据，需要很长时间训练或需要比图形处理单元（GPU）更多RAM的大型或复杂模型，或具有严格训练时间限制的学习问题，仅举几例。该项目旨在通过将关系数据库系统的思想应用于ML系统的设计和实现来解决这些缺陷。 在关系式引擎之上构建ML系统将使ML系统的设计能够自动生成特定ML任务的计算计划，而只需很少的程序员工作。 这些计划将被优化和执行，以匹配数据大小，布局和计算硬件。无论计算是在本地机器上还是在分布式环境中运行，实现ML算法的代码都是相同的。 如果成功，该项目将从根本上扩展ML的易用性和适用性。在关系系统上运行ML计算需要回答许多技术问题，回答这些问题将是该项目的核心。 例如：ML基元（卷积、循环模块等）如何映射到关系原语上 如何将大型对象（矩阵/张量）放入记录中，以便关系实现高效运行？ 由于ML算法的开发人员不太可能接受SQL作为编程语言：如何将类似Karas的Python程序转换为关系代数？该奖项反映了NSF的法定使命，并被认为是值得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的支持。

Machine learning (ML) will have a huge economic and scientific impact over the upcoming decades. But despite the growing importance of ML, there has been relatively little work focused on asking what systems to support machine learning should look like. The result is that it is difficult to apply ML to non-standard cases: Big Data, large or complex models that take a long time to train or require more RAM than is available on a graphics processing unit (GPU), or learning problems with hard training time constraints, to name a few. The proposed project aims to address such deficiencies by applying ideas from relational database systems to the design and implementation of systems for ML. Building an ML system on top of a relational-style engine will enable the design of ML systems that are able to automatically generate compute plans for specific ML tasks with little programmer effort. Those plans will be optimized and executed to match the data size, layout, and the compute hardware. The code to implement an ML algorithm will be the same no matter whether the computation is run on a local machine, or in a distributed environment. If successful, the project will radically expand the ease-of-use and applicability of ML.There are a number of technical questions that need to be answered for ML computations to be run on top of a relational system, and answering such questions will be at the heart of the project. For example: How can ML primitives (convolutions, recurrent modules, etc.) be mapped onto relational primitives? How can large objects (matrices/tensors) be chucked into records so relational implementations run efficiently? And since a developer of ML algorithms is unlikely to accept SQL as a programming language: How to translate Karas-like Python programs into relational algebra?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.