PPoSS: Planning: Eliminating the Bottlenecks to ML Usability and Scalability

机器学习（ML）有可能极大地造福于日常生活的许多方面。应用范围从消费者和商业技术（虚拟助理）到交通（驾驶辅助和自动驾驶汽车），再到医疗保健（医院观察和医生辅助）和科学。不幸的是，使用ML技术来解决具体的、现实世界的问题仍然很困难。构建机器学习应用程序通常需要一个多步骤、迭代的过程，包括原型设计、评估结果和修复故障的重复循环。这个项目旨在显著降低这个迭代过程的难度，使一个人更容易快速构建ML应用程序来解决现实世界的问题。该项目的技术目标是大幅提高整个端到端机器学习模型开发工作流程的生产力，以便单个主题专家拥有大型数据集并访问数据中心规模的加速计算能力，可以在数小时到数天内为新任务构建准确的模型，而不是数周或数月。为此，项目将创建自动化工具，直接授权主题专家在快速迭代循环中执行模型开发任务。这些智能工具必须以低延迟执行，并可扩展到庞大的数据集和潜在的大型复杂模型。为了满足这些需求，该项目将结合从弱源自动生成监督的新方法，以及为训练和验证管理关键数据的新方法，新的ML编程抽象，以及可重构的ML硬件加速器，以支持一组异构的未来用例和算法技术。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。

Machine learning (ML) has the potential to dramatically benefit many aspects of daily life. Applications range from consumer and business technologies (virtual assistants) to transportation (driver assistance and self-driving cars) to health care (hospital observation and physician assistance) and science. Unfortunately, using ML techniques to solve specific, real-world problems remains difficult. Building ML applications often requires a multi-step, iterative process that involves repeated cycles of prototyping, evaluating results, and fixing failures. This project aims to dramatically reduce the difficulty of this iterative process, making it easier for a single person to quickly build ML applications to solve real-world problems.The technical goal of this project is to dramatically scale the productivity of the entire end-to-end ML model-development workflow so that a single subject-matter expert, armed with a large dataset and access to datacenter-scale accelerated compute capability, can build accurate models for new tasks in hours to days instead of weeks or months. To this end, the project will create automated tools that directly empower subject-matter experts to perform model-development tasks, in rapid iterative loops. These intelligent tools must execute with low latency and scale to huge datasets and potentially large, complex models. To meet these requirements, this project will combine new methods for automatically generated supervision from weak sources and for curating critical data for training and validation, new ML programming abstractions, and reconfigurable ML hardware accelerators to support a heterogeneous set of future use cases and algorithmic techniques.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.