SBIR Phase I: Large Scale And Automated Unsupervised Machine Learning For Anomaly Detection

SBIR 第一阶段：用于异常检测的大规模自动化无监督机器学习

• 批准号: 1843988
• 负责人: Vinay Rao
• 金额: $ 22.5万
• 依托单位: RocketML Inc
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-01-01 至 2020-05-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1843988&HistoricalAwards=false
• 关键词: SBIR; Phase; Large; Scale; Automated

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will result from enabling businesses to process and extract insights from large unlabeled datasets, using machine learning with minimal human supervision, in application areas such as cyber security, precision medicine and predictive maintenance. Current deep learning approaches require large amounts of labeled data. Creation of labeled data is expensive, error prone and time consuming. The proposed software will provide fully automated capabilities for semi-supervised learning for anomaly detection in cyber security applications. All businesses ranging from large scale enterprises to boutique data science consulting firms will benefit from this project. The expected impact can range in the billions of dollars in the areas of cyber security and predictive maintenance, to name just two. More broadly the proposed technologies will enable both corporations and public institutions to harvest large datasets at minimal cost.This Small Business Innovation Research (SBIR) Phase I project will design, develop, and deploy high-performance computing (HPC) software for unsupervised learning and anomaly detection. In the last decades tremendous successes in machine learning have been achieved in the area of supervised learning that requires compilation of large datasets with labels (for example, grouping of pictures based on the individual depicted on the image). In contrast, unsupervised learning algorithms do not require labels and thus require minimal human participation. However, due to significant technical difficulties they have not been as successful as supervised learning algorithms. This software package circumvents these difficulties and opens the way to scaling unsupervised learning algorithms to large and complex datasets. The main research and development challenges that will be addressed in this project are the ability to integrate this new technology with real world complex datasets through the choice of the correct comparison function between the objects of the dataset and the fully automatic algorithm and algorithm parameter selection.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.

这个小企业创新研究（SBIR）第一阶段项目的更广泛的影响/商业潜力将来自于使企业能够在网络安全，精准医学和预测性维护等应用领域使用机器学习，在最少的人工监督下处理和提取大型未标记数据集的见解。 目前的深度学习方法需要大量的标记数据。 标记数据的创建是昂贵的、容易出错且耗时的。拟议的软件将为网络安全应用中的异常检测提供半监督学习的全自动化功能。从大型企业到精品数据科学咨询公司的所有企业都将从该项目中受益。在网络安全和预测性维护领域，预计影响可能达到数十亿美元，仅举两例。更广泛地说，拟议的技术将使企业和公共机构能够以最低的成本收集大型数据集。这个小企业创新研究（SBIR）第一阶段项目将设计，开发和部署高性能计算（HPC）软件，用于无监督学习和异常检测。在过去的几十年里，机器学习在监督学习领域取得了巨大的成功，这需要编译带有标签的大型数据集（例如，基于图像上描绘的个人对图片进行分组）。相比之下，无监督学习算法不需要标签，因此需要最少的人类参与。然而，由于存在重大技术困难，它们并不像监督学习算法那样成功。该软件包规避了这些困难，并开辟了将无监督学习算法扩展到大型复杂数据集的途径。 该项目将解决的主要研究和开发挑战是通过选择数据集对象之间的正确比较函数以及全自动算法和算法参数选择，将这项新技术与真实的世界复杂数据集集成的能力。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值进行评估，被认为值得支持和更广泛的影响审查标准。