MRI: Acquisition of a Purpose-Built Deep Learning Compute System to Advance Fundamental Research and Education at Penn State

MRI：收购专用深度学习计算系统以推进宾夕法尼亚州立大学的基础研究和教育

• 批准号: 2018280
• 负责人: Douglas Cowen
• 金额: $ 31.79万
• 依托单位: Pennsylvania State Univ University Park
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-08-15 至 2023-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2018280&HistoricalAwards=false
• 关键词: MRI; Acquisition; Purpose; Built; Deep

Successful application of deep learning to complex research problems often hinges on access to highly specialized computational hardware. This award will provide funds to purchase a system comprised of four highly effective, purpose-built deep learning nodes to be used for a wide array of deep learning applications. An additional purchase of two state-of-the-art, purpose-built deep learning nodes for high-dimensional, memory-hungry and/or data-rich applications is also enabled. The four “staging” nodes will permit development and testing prior to deployment on the two state-of-the-art nodes. The full six-node system will be used by members of the Deep Learning for Statistics, Astrophysics, Geoscience, Engineering, Meteorology and Atmospheric Science, Physical Sciences and Psychology (DL-SAGEMAPP) team and the broader community at the Pennsylvania State University. The team will provide access to the hardware to roughly 300 advanced undergraduate and graduate students each year, for use in courses in multiple departments that either focus on machine learning and artificial intelligence or include those topics in their curricula. The team will also host an annual multi-day hands-on workshop in deep learning. The workshop will welcome all students in the DL-SAGEMAPP team’s research areas but will advertise most heavily to students from underrepresented groups. Team members will also contribute deep learning sessions to one or more of the annual workshops for K-12 teachers held by Penn State’s Astronomy and Astrophysics Department, with the potential to convey deep learning ideas to grade-school students.The goal is to create a cutting-edge, shared resource that supports a diverse set of researchers, allowing them to transform problems that are currently impractical or impossible to solve with existing computational resources into proverbial “low-hanging fruit.” The DL-SAGEMAPP team will harness the power of deep learning to tackle some of the most challenging problems in their respective areas of research. With purpose-built hardware and popular open software, the team will apply deep learning methodologies to problems with high dimensionality, high data volume, and/or requiring very complex network topologies; problems that would take too long to run or would be too large to fit in on-board memory for most standard hardware configurations. This team aims to boost the search for multimessenger astrophysical signals by improving the sensitivity and response time of flagship high-energy neutrino, gravitational wave, and wide-field survey observatories. The team will apply deep learning to the simulation and analysis of satellite and aerial data, leading to greater predictive power for impending volcanic activity and the build-up of sea ice in the Arctic, and greater acuity with cloud-shrouded ground targets. Deep learning techniques will be used to increase the accuracy of flood forecasting, improve the accuracy and turnaround time of molecular-level simulations, delve more deeply into the process of protein synthesis by mRNA, tackle the analysis of increasingly large data volumes from brain-implanted electrodes, and sharpen researchers’ understanding of the human visual system.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.

深度学习在复杂研究问题上的成功应用往往取决于高度专业化的计算硬件。该合同将为购买一个由四个高效、专用深度学习节点组成的系统提供资金，用于广泛的深度学习应用。另外还为高维、内存消耗和/或数据丰富的应用程序购买了两个最先进的、专门构建的深度学习节点。这四个“分期”节点将允许在部署到两个最先进的节点之前进行开发和测试。完整的六节点系统将由宾夕法尼亚州立大学统计、天体物理学、地球科学、工程、气象和大气科学、物理科学和心理学（DL-SAGEMAPP）团队的深度学习成员以及更广泛的社区使用。该团队每年将向大约300名高级本科生和研究生提供硬件访问权限，用于多个部门的课程，这些课程要么专注于机器学习和人工智能，要么将这些主题纳入课程。该团队还将举办为期数天的年度深度学习实践研讨会。研讨会将欢迎DL-SAGEMAPP团队研究领域的所有学生，但将对来自代表性不足群体的学生进行最大量的宣传。团队成员还将在宾夕法尼亚州立大学天文学和天体物理系为K-12教师举办的一次或多次年度研讨会上贡献深度学习课程，有可能向小学生传达深度学习的理念。其目标是创建一个尖端的共享资源，支持不同的研究人员，使他们能够将目前不切实际或不可能用现有计算资源解决的问题转化为众所周知的“唾手可得的成果”。DL-SAGEMAPP团队将利用深度学习的力量来解决各自研究领域中一些最具挑战性的问题。借助专用硬件和流行的开放软件，该团队将把深度学习方法应用于高维、高数据量和/或需要非常复杂网络拓扑的问题；对于大多数标准硬件配置，运行时间太长或内存太大而无法容纳的问题。这个团队的目标是通过提高旗舰高能中微子、引力波和宽视场巡天天文台的灵敏度和响应时间，来促进对多信使天体物理信号的搜索。该团队将把深度学习应用于卫星和航空数据的模拟和分析，从而提高对即将到来的火山活动和北极海冰积聚的预测能力，并提高对云层覆盖的地面目标的敏锐度。深度学习技术将用于提高洪水预报的准确性，提高分子水平模拟的准确性和周转时间，更深入地研究mRNA合成蛋白质的过程，处理来自大脑植入电极的越来越大的数据量的分析，并提高研究人员对人类视觉系统的理解。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。

项目成果

A Multiscale Deep Learning Model for Soil Moisture Integrating Satellite and In Situ Data

• DOI: 10.1029/2021gl096847
• 发表时间: 2022-03
• 期刊: Geophysical Research Letters
• 影响因子: 5.2
• 作者: Jiangtao Liu;F. Rahmani;K. Lawson;Chaopeng Shen

Automatic atmospheric correction for shortwave hyperspectral remote sensing data using a time-dependent deep neural network

• DOI: 10.1016/j.isprsjprs.2021.02.007
• 发表时间: 2021-02-24
• 期刊: ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING
• 影响因子: 12.7
• 作者: Sun, Jian;Xu, Fangcao;Salvador, Mark
• 通讯作者: Salvador, Mark

Differentiable, Learnable, Regionalized Process‐Based Models With Multiphysical Outputs can Approach State‐Of‐The‐Art Hydrologic Prediction Accuracy

• DOI: 10.1029/2022wr032404
• 发表时间: 2022-03
• 期刊: Water Resources Research
• 影响因子: 5.4
• 作者: D. Feng;Jiangtao Liu;K. Lawson;Chaopeng Shen

Evaluating a global soil moisture dataset from a multitask model (GSM3 v1.0) with potential applications for crop threats

通过多任务模型 (GSM3 v1.0) 评估全球土壤湿度数据集以及作物威胁的潜在应用

• DOI: 10.5194/gmd-16-1553-2023
• 发表时间: 2023
• 期刊: Geoscientific Model Development
• 影响因子: 5.1
• 作者: Liu, Jiangtao;Hughes, David;Rahmani, Farshid;Lawson, Kathryn;Shen, Chaopeng
• 通讯作者: Shen, Chaopeng

Automatic Detection of Volcanic Surface Deformation Using Deep Learning

• DOI: 10.1029/2020jb019840
• 发表时间: 2020-09
• 期刊: Journal of Geophysical Research: Solid Earth
• 作者: Jian Sun;C. Wauthier;K. Stephens;M. Gervais;G. Cervone;P. L. La Femina;M. Higgins