Collaborative Research: Frameworks: Machine learning and FPGA computing for real-time applications in big-data physics experiments

近年来，引力波天体物理、高能物理和大规模电磁调查所需的网络基础设施迅速发展。用于在这些不同的研究领域进行科学发现的设施的建设和升级导致了两个共同的计算重大挑战：(1)日益复杂和数量不断增加的数据集；(2)必须在超额订阅的计算资源下实时执行的数据挖掘分析。此外，引力波天体物理学与电磁和天体粒子测量的融合，就是多信使天体物理学的诞生，已经提供了一个在未来几年它将使之能够实现的变革性发现的一瞥。鉴于大型强子对撞机(LHC)以及激光干涉仪引力波天文台(LIGO)和用于多信使天体物理的大型天气观测望远镜(LSST)的组合在科学发现方面的独特潜力，社区需要加快开发和利用将超过现有方法的深度学习算法。正如美国国家科学基金会的使命所说，这个项目通过促进科学进步来服务于国家利益。它将在规模上推动深度学习的前沿，展示这些方法的多功能性和可扩展性，以加速和实现大数据时代的新物理。由于这些方法也适用于我们国家和全球经济社会的许多其他部分，这项工作将对许多领域产生积极影响。在这项研究中接受指导和培训的学生和初级科学家将与我们的行业合作伙伴密切互动，创造新的职业机会，并加强学术界和产业界之间的协同效应。该团队将通过开源软件库与社区共享算法，并通过我们的教程和研讨会培训社区关于软件信用和软件引用。在这个项目中，PI将基于我们最近的工作开发高质量的深度学习算法，用于作为开源软件的时间序列和图像数据集的实时数据分析。这项工作结合了可扩展的深度学习算法和最先进的方法，这些算法使用数千个GPU/CPU在几分钟内使用TB大小的数据集进行训练，从而使确定性深度学习模型的预测具有完全的后验分布。该团队还将研究使用现场可编程门阵列(FPGA)来加速机器学习算法的低延迟推理，以最大限度地减少未来计算的需求，这是多信使天体物理和粒子物理的中心目标。将作为这些活动的一部分开发的开源工具将很容易与LIGO、LHC和LSST共享并作为核心数据分析算法采用，这些算法将显著提高现有算法的速度和深度，实现新的物理学，同时需要最少的计算资源进行实时推断分析。该团队将组织深度学习讲习班和训练营，培训学生和研究人员如何使用我们的框架并为其做出贡献，创建一个由关键科学任务的贡献者和开发人员组成的广泛网络。该团队将利用现有的开源和交互模型存储库，如Argonne的数据和科学学习中心(DLHub)，以接触到大量的社区，这些社区分析来自LIGO、LHC和LSST的开放数据集，并将受益于这些只需最少计算资源来完成推理任务的技术的使用。该项目由计算机与信息科学和信息科学理事会的高级网络基础设施办公室支持。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The cyberinfrastructure needs for gravitational wave astrophysics, high energy physics, and large-scale electromagnetic surveys have rapidly evolved in recent years. The construction and upgrade of the facilities used to enable scientific discovery in these disparate fields of research have led to a common pair of computational grand challenges: (i) datasets with ever-increasing complexity and volume; and (ii) data mining analyses that must be performed in real-time with oversubscribed computational resources. Furthermore, the convergence of gravitational wave astrophysics with electromagnetic and astroparticle surveys, the very birth of Multi-Messenger Astrophysics, has already provided a glimpse of the transformational discoveries that it will enable in years to come. Given the unique potential for scientific discovery with the Large Hadron Collider (LHC) and the combination of the Laser Interferometer Gravitational-wave Observatory (LIGO) and the Large Synoptic Survey Telescope (LSST) for Multi-Messenger Astrophysics, the community needs to accelerate the development and exploitation of deep learning algorithms that will outperform existing approaches. This project serves the national interest, as stated by NSF's mission, by promoting the progress of science. It will push the frontiers of deep learning at scale, demonstrating the versatility and scalability of these methods to accelerate and enable new physics in the big data era. Because these methods are also applicable to many other parts of our national and global economy and society, this work will positively impact many fields. The students and junior scientists to be mentored and trained in this research will interact closely with our industry partners, creating new career opportunities, and strengthening synergies between academia and industry. The team will share the algorithms with the community through open source software repositories, and through our tutorials and workshops the team will train the community regarding software credit and software citation.In this project, the PIs will build upon our recent work developing high quality deep learning algorithms for real-time data analytics of time-series and image datasets, as open source software. This work combines scalable deep learning algorithms, trained with TB-size datasets within minutes using thousands of GPUs/CPUs, with state-of-the-art approaches to endow the predictions of deterministic deep learning models with complete posterior distributions. The team will also investigate the use of Field Programmable Gate Arrays (FPGAs) to accelerate low-latency inference of machine learning algorithms to minimize the demands of future computing, which is a central goal for Multi-Messenger Astrophysics and particle physics. The open source tools to be developed as part of these activities will be readily shared with and adopted by LIGO, LHC, and LSST as core data analytics algorithms that will significantly increase the speed and depth of existing algorithms, enabling new physics while requiring minimal computational resources for real-time inferences analyses. The team will organize deep learning workshops and bootcamps to train students and researchers on how to use and contribute to our framework, creating a wide network of contributors and developers across key science missions. The team will leverage existing open source and interactive model repositories, such as the Data and Learning Hub for Science (DLHub) at Argonne, to reach out to a large cross-section of communities that analyze open datasets from LIGO, LHC, and LSST, and that will benefit from the use of these technologies that require minimal computational resources for inference tasks.This project is supported by the Office of Advanced Cyberinfrastructure in the Directorate for Computer & Information Science & Engineering and the Division of Physics in the Directorate of Mathematical and Physical Sciences.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.