EAGER: Collaborative Research: III: Exploring Physics Guided Machine Learning for Accelerating Sensing and Physical Sciences

EAGER：协作研究：III：探索物理引导机器学习以加速传感和物理科学

• 批准号: 2026704
• 负责人: Anish Arora
• 金额: $ 5.02万
• 依托单位: Ohio State University
• 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: 2020-05-01 至 2021-04-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2026704&HistoricalAwards=false
• 关键词: EAGER; Collaborative; Research; III; Exploring

As machine learning (ML) continues to revolutionize the commercial space including vision, speech, andtext recognition, there is great anticipation in the scientific community to unlock the power of ML foraccelerating scientific discovery. However, black-box ML models, which rely solely on training data andignore existing scientific knowledge have met with limited success in scientific problems, particularlywhen labeled data is limited, sometimes even leading to spectacular failures. This is because the blackbox ML models are susceptible to learning spurious relationships that do not generalize well outside thedata they are trained for. The emerging paradigm of physics-guided machine learning (PGML), whichleverages the unique ability of ML algorithms to automatically extract patterns and models from data withguidance of the knowledge accumulated in physics (or scientific theories), aims to address the challengesfaced by black box ML in scientific applications. Significant exploratory efforts are needed to formulate and assess sound PGML approaches for particular scientific problems.For data science, PGML has the potential to transform ML beyond black-box applications by enablingsolutions that generalize well even on unseen input-output distributions that are different from thoseencountered during training, by anchoring ML methods with the scientific body of knowledge. PGML makes a distinctdeparture from the conventional view that physics-based models and ML models are developed inisolation but seldom mixed together. The proposed project is fundamentally different from existing bodyof research that attempts to combine ML and domain sciences, e.g., by making use of domain-specificknowledge in ML algorithms in simplistic ways, or making use of data in the physics-based modelingprocess albeit without allowing data to change the functional forms of existing physics-based models. The tight interplay between data science and the domains of physics and sensing in the project lends itselfnaturally to diverse education activities that complement the research tasks outlined by our team. Over theduration of this one-year project, the team will develop an integrative course at the graduate level on "MLmeets Physics", which explores topical, emerging themes in this interdisciplinary area. Offerings of thecourse will draw upon course modules shared between the four universities, such as shared guest videosand case studies. The physics department at BU has a well-developed "Physics Outreach Project" thatannually performs science exhibitions for elementary schools in Binghamton metropolitan area, for whichthe team will create a new exhibition about neural networks and ML. In follow-on work, similar outreachevents will be replicated at schools (Robinson Middle School in Lowell and Metro STEM Middle Schoolin Columbus). The PIs are committed to increasing the diversity of involvement at various levels of thetraining ecosystem impacted by this project, and have planned various coordinated broader impactactivities for inclusion of female and underrepresented minority students as well as faculty.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.

随着机器学习（ML）继续彻底改变商业空间，包括愿景，语音，和文本识别，科学界有很大的期望，可以解锁ML强调科学发现的力量。但是，仅依靠培训数据和现有科学知识的黑盒ML模型在科学问题上取得了有限的成功，尤其是标记的数据是有限的，有时甚至导致了壮观的失败。这是因为BlackBox ML模型很容易受到学习的虚假关系，这些关系并不能很好地概括他们的训练之外。物理学引导的机器学习（PGML）的新兴范式却阐明了ML算法在物理学（或科学理论）中自动从数据中自动从数据中提取模式和模型的独特能力，旨在解决黑匣子ML在科学应用中提出的挑战。需要进行大量探索性努力来制定和评估针对特定科学问题的声音PGML方法。对于数据科学，PGML有可能通过AnablingSolutions将ML超越黑盒应用程序，从而通过与培训过程中的培训ML方法相同，这些方法与在培训期间与培训过程中不同的输入输出分布进行了很好的推广。 PGML从传统观点中脱颖而出，即基于物理学的模型和ML模型是不可分化的，但很少混合在一起。所提出的项目与现有的Bodyof研究根本不同，该研究试图通过简单的方式使用ML算法中的域规范性，或者在物理基于物理的模型中使用数据，尽管不允许数据更改现有物理学模型的功能形式，但使用ML算法中的域规范化。数据科学与物理领域和项目中的紧密相互作用与我们团队概述的研究任务的多样化的教育活动自然而然。在这个为期一年的项目中，该团队将在“ MLMeets Physics”上开发一门综合课程，该课程探讨了该跨学科领域的主题，新兴的主题。 TheCourse的产品将借鉴四所大学之间共享的课程模块，例如共享的来宾视频和案例研究。 BU的物理系有一个完善的“物理外展项目”，该项目旨在为宾厄姆顿都会区的小学进行科学展览，为此，该团队将创建有关神经网络和ML的新展览。在后续工作中，将在学校（洛厄尔的鲁滨逊中学和哥伦布大都会中学的罗宾逊中学）复制类似的外展文章。 PI致力于增加受该项目影响的各个级别的生态系统的参与度的多样性，并计划为包括女性和代表性不足的少数群体和教职员工的各种协调的更广泛的影响力以及该奖项的奖项反映了NSF的法定责任，并通过评估范围来进行评估，并反映了该奖项，并反映了企业的支持。

项目成果

Specialized Embedding Approximation for Edge Intelligence: A Case Study in Urban Sound Classification

• DOI: 10.1109/icassp39728.2021.9414287
• 发表时间: 2021-06
• 期刊: ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
• 作者: Sangeeta Srivastava;Dhrubojyoti Roy;M. Cartwright;J. Bello;A. Arora