BIGDATA: IA: Collaborative Research: Asynchronous Distributed Machine Learning Framework for Multi-Site Collaborative Brain Big Data Mining

BIGDATA：IA：协作研究：用于多站点协作大脑大数据挖掘的异步分布式机器学习框架

• 批准号: 2348159
• 负责人: Heng Huang
• 金额: $ 78万
• 依托单位: University of Maryland, College Park
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-09-01 至 2024-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2348159&HistoricalAwards=false
• 关键词: BIGDATA; IA; Collaborative; Research; Asynchronous

Recent advances in multimodal brain imaging and high throughput genotyping and sequencing techniques provide exciting new opportunities to ultimately improve our understanding of brain structure and neural dynamics, their genetic architecture, and their influences on cognition and behavior. However, data privacy and security issues have inhibited data sharing across institutes. Emerging multi-site collaborative data analysis can address these issues and facilitate data and computing resource sharing. In collaborative data analysis, the participating institutes keep their own data, which are analyzed and computed locally, and only share the computed results by communicating with a server. The server communicates with all institutes and updates the local models such that the trained machine learning models indirectly use all data and are shared with all institutes. Although some distributed/parallel computation techniques were recently proposed to address big data mining problems, most of them are synchronous models. Asynchronous distributed learning methods are much more efficient, because they allow the server to update the model with information from only one worker node without waiting for slow worker nodes in each round. However, the convergence analysis for the asynchronous distributed algorithms is much more difficult due to the inconsistent variables update across nodes. Thus, it is challenging to design efficient distributed machine learning algorithms for collaborative big data analysis. The research objective of this project is to address the computational challenges in the emerging multi-site collaborative data mining for brain big data. This project seeks to harness the opportunities of designing new efficient asynchronous distributed machine learning algorithms with rigorous theoretical foundations for multi-site collaborative brain big data mining, creating large-scale computational strategies and effective software tools to reveal sophisticated relationships among heterogeneous brain data. This project designs the asynchronous distributed machine learning and principled big data mining models to conduct the comprehensive study of brain imaging genomics and connectomics. Specifically, the principal investigators investigate: 1) collaborative genotype and phenotype association study using new asynchronous doubly stochastic proximal gradient algorithms; 2) communication-efficient multi-site collaborative data integration models to integrate imaging genomics data for predicting outcomes of interest; 3) collaborative deep learning algorithm speedup by the asynchronous distributed algorithms with applications in temporal cognitive change prediction; and 4) new graph convolutional deep learning models for brain network mining. It is innovative to integrate new distributed machine learning and data-intensive computing with brain imaging genomics and connectomics that hold great promise for a systems biology of the brain. The developed methods and tools impact other neuroimaging, genomics, and neuroscience research, and enable investigators working on brain science to effectively test their scientific hypotheses. This project will also facilitate the development of novel educational tools.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.

多模式脑成像和高通量基因分型和测序技术的最新进展提供了令人兴奋的新机会，最终提高我们对大脑结构和神经动力学，其遗传结构及其对认知和行为的影响的理解。然而，数据隐私和安全问题阻碍了各机构之间的数据共享。新兴的多站点协作数据分析可以解决这些问题，并促进数据和计算资源共享。在协同数据分析中，参与机构保持自己的数据，这些数据在本地进行分析和计算，并且仅通过与服务器通信来共享计算结果。服务器与所有机构进行通信并更新本地模型，以便经过训练的机器学习模型间接使用所有数据并与所有机构共享。虽然最近提出了一些分布式/并行计算技术来解决大数据挖掘问题，但大多数都是同步模型。异步分布式学习方法效率更高，因为它们允许服务器仅使用来自一个工作节点的信息更新模型，而无需在每一轮中等待缓慢的工作节点。然而，异步分布式算法的收敛性分析是非常困难的，由于不一致的变量更新跨节点。因此，设计高效的分布式机器学习算法用于协作大数据分析是具有挑战性的。该项目的研究目标是解决新兴的大脑大数据多站点协作数据挖掘中的计算挑战。该项目旨在利用设计新的高效异步分布式机器学习算法的机会，为多站点协作大脑大数据挖掘提供严格的理论基础，创建大规模计算策略和有效的软件工具，以揭示异构大脑数据之间的复杂关系。本项目设计了异步分布式机器学习和原则性大数据挖掘模型，开展脑成像基因组学和连接组学的综合研究。具体而言，主要研究人员研究：1）使用新的异步双随机近端梯度算法进行协作基因型和表型关联研究; 2）通信高效的多站点协作数据集成模型，以集成成像基因组学数据来预测感兴趣的结果; 3）异步分布式算法加速协作深度学习算法，并应用于时间认知变化预测;以及4）用于大脑网络挖掘的新的图卷积深度学习模型。将新的分布式机器学习和数据密集型计算与大脑成像基因组学和连接组学相结合是创新的，这对大脑的系统生物学有很大的希望。开发的方法和工具影响了其他神经成像，基因组学和神经科学研究，并使脑科学研究人员能够有效地测试他们的科学假设。该奖项反映了NSF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。