SBIR Phase I: Solving Visual Data Problems for Large Scale Machine Learning Applications

这个小企业创新研究（SBIR）第一阶段项目的更广泛的影响/商业潜力将是提供一个统一的后端，为机器学习（ML）部署提供无缝和可扩展的视觉数据访问。这是通过在统一的管道中消除对众多独立产品的碎片化系统设计的需求来实现的。机器学习的改进使企业能够从信息丰富的视觉数据（如图像和视频）中提取见解。处理用于ML和查询目的的大可视化数据需要为可视化ML设计的存储和访问方法。由于目前现成的替代方案，ML工程师和数据科学家被迫合并未准备好的数据解决方案来解决可视化数据管理问题。企业支付技术债务的形式是：a)额外的数据平台资源；b)当机器学习工程师和数据科学家被迫设计基础设施时，人才不匹配；c)推迟产品发布。该项目创建了一个统一的系统，在机器学习的各个阶段都有一个解决方案，从数据收集、管理到训练、推理和业务查询。这个小企业创新研究（SBIR）第一阶段项目将导致一个为大规模视觉数据设计的新型数据管理平台，具有专门用于机器学习（ML）和专家见解查询的界面。该项目旨在构建基础设施，以支持数千个并发客户端、数万亿元数据实体和pb级的可视化数据，这将在越来越多地使用可视化数据的领域中变得常见。该平台是可伸缩的，不会影响性能，特别是对于ML部署的可视化数据管理的新兴领域。这项工作将包括从大量类似物联网的设备接收数据时的视觉数据存储，以及用于低延迟、高吞吐量访问大视觉数据的ml感知应用程序编程接口。可扩展的元数据数据库旨在利用新的内存技术和新的缓存和分层方法，通过新的格式使用基于内容的图像/视频数据知识。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。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will be to offer a unifying backend that will enable seamless and scalable visual data access for Machine Learning (ML) deployments. This is achieved by removing the need for a fragmented system design from numerous independent products in an otherwise unified pipeline. Improvements in ML have made it possible for businesses to extract insights from information-rich visual data such as images and video. Handling big-visual-data for ML and query purposes requires storage and access methods designed for visual ML. With the current off-the-shelf alternatives, ML engineers and data scientists are forced to merge unprepared data solutions to address visual data management. Businesses pay the technical debt in the form of a) extra data platform resources, b) talent mismatch when ML engineers and data scientists are forced to engineer infrastructure, and c) delayed product launches. This project creates a unified system with one solution across the various stages of ML starting from data collection, curation, to training, inference, and business queries.This Small Business Innovation Research (SBIR) Phase I project will lead to a novel data management platform designed for large-scale visual data, with an interface specialized for Machine Learning (ML) and Expert Insights queries. The project aims to build infrastructure to support thousands of concurrent clients, trillions of metadata entities, and petabytes of visual data, as will be common in the domains with increasing use of visual data. The platform is scalable without affecting performance, particularly for the emerging area of visual data management for ML deployments. The work will include visual data storage when receiving data from a large number of IoT-like devices and a ML-aware application programming interface for low-latency, high-throughput access of big-visual-data. The scalable metadata database is designed to exploit new memory technology and novel caching and tiering methods using content-based knowledge of image/video data via novel formats.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.