Demonstrator of Flower Federated Learning Tool: DEFT

Flower联邦学习工具演示者：DEFT

• 批准号: EP/Y024370/1
• 负责人: Nicholas Lane
• 金额: $ 16.19万
• 依托单位: University of Cambridge
• 依托单位国家: 英国
• 项目类别: Research Grant
• 财政年份: 2023
• 资助国家: 英国
• 起止时间: 2023 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=EP%2FY024370%2F1
• 关键词: Demonstrator; Flower; Federated; Learning; Tool

Building upon recent results from the ERC StG REDIAL project, we seek to develop a commercial proof-of-concept of federated learning technology. Core innovation in the REDIAL project has been been made in the direction of scalable simulation and efficient training of federated systems -- along with enabling methods for unsupervised video understanding and speech. The technical breakthrough is the enabling of machine learning systems (e.g., detection of cancer, performing drug discovery, detecting financial fraud) that can be trained even if the data remains at the originating institution -- and yet a model can be built that benefits from the complete set of data that spans all of these locations. Often due to legal, ethical, intellectual property, difficulty of system compatibility -- models are trained on data sets far smaller than would be expected because of the difficulties of bring data into one single data center. A secondary motivation behind interest in the ability to build privacy preserving user systems, such as federated forms of digital assistants, IoT devices and smartphone services. The commercial opportunity extends from the fact REDIAL technology has been integrated within a highly popular completely open-source federated learning toolkit called FLOWER -- which has also developed in the lab of the PI of REDIAL. More than 420,000 downloads of FLOWER with REDIAL based technology have occurred. There is a growing user base and community around FLOWER. But developing federated solutions remains too hard for mainstream use -- and companies require more support, and easy to use specific tools. This proof-of-concept grant would be to facilitate the development of a hardened simple set of tools (we call it LEAP) and support for companies build federated learning applications. We believe they will come in the areas such as healthcare, preventative maintenance, financial analysis, and drug discovery.

基于ERC StG REDIAL项目的最新成果，我们寻求开发联邦学习技术的商业概念验证。REDIAL项目的核心创新是在联邦系统的可扩展模拟和有效训练方面--沿着无监督视频理解和语音的启用方法。技术突破是机器学习系统的实现（例如，检测癌症、进行药物发现、检测金融欺诈），即使数据保留在原始机构，也可以对其进行训练--而且可以建立一个模型，该模型受益于跨越所有这些位置的完整数据集。通常，由于法律的、道德的、知识产权的、系统兼容性的困难--模型是在比预期小得多的数据集上训练的，因为将数据带入单个数据中心很困难。第二个动机是对构建隐私保护用户系统的能力感兴趣，例如联合形式的数字助理，物联网设备和智能手机服务。商业机会来自于REDIAL技术已被集成到一个非常流行的完全开源的联邦学习工具包FLOWER中，该工具包也是在REDIAL的PI实验室中开发的。超过420，000次下载基于REDIAL技术的FLOWER。有一个不断增长的用户群和社区周围的花。但是开发联邦解决方案对于主流应用来说仍然太难了--公司需要更多的支持和易于使用的特定工具。这一概念验证资助将促进一组强化的简单工具（我们称之为LEAP）的开发，并支持公司构建联邦学习应用程序。我们相信它们将出现在医疗保健、预防性维护、财务分析和药物发现等领域。