서브메뉴
검색
AI for Science: Graph Machine Learning as an Instrument for Understanding, Controlling, and Creating Physical Systems
AI for Science: Graph Machine Learning as an Instrument for Understanding, Controlling, and Creating Physical Systems
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151006
- ISBN
- 9798381975192
- DDC
- 540
- 저자명
- Thieme, Mattson.
- 서명/저자
- AI for Science: Graph Machine Learning as an Instrument for Understanding, Controlling, and Creating Physical Systems
- 발행사항
- [Sl] : Northwestern University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 198 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 85-10, Section: B.
- 주기사항
- Advisor: Liu, Han.
- 학위논문주기
- Thesis (Ph.D.)--Northwestern University, 2024.
- 초록/해제
- 요약We consider machine learning (ML) to be a programmatic analog of the scientific method. In this paradigm, a mapping encoded in the model weights (a theory) transforms input data (observations) into predictions, which are then compared against a target (validating the theory against reality). As training proceeds and the correspondence between out-of-sample predictions and reality grows, so too does our confidence that the learned mapping reflects some truth about the problem domain. If this agreement is sufficiently robust, an examination of the mapping may then yield insights into the rules and relationships governing the system under study. Inspired by the potential for ML to accelerate scientific discovery, we explore differentiable, graph-based algorithms that allow neural models to more efficiently extract and leverage latent relational information from raw data. Our focus on graph-based algorithms not only makes the work more general but also more readily applicable to the physical sciences, where graphs are the de facto representational structure. We introduce novel methods for 1) graph structure learning, where we fix the nodes and learn to retain/remove the edges of a graph; 2) graph partition learning, where we fix the edges and learn to cluster the nodes; and 3) molecular graph generation with a novel graph adaptation methodology, where we fix the edges and optimize graph-level properties by modifying node features. Finally, in the spirit of concretely accelerating scientific discovery with machine learning, we discuss our extensive work on AI for Science conducted in partnership with Fermilab, where we repurposed biomedical segmentation models for disentangling particle accelerator loss profiles, and recurrent sequence models for controlling high-frequency proton beam extractors.
- 일반주제명
- Chemistry
- 일반주제명
- Pharmacology
- 일반주제명
- Computer science
- 일반주제명
- Bioinformatics
- 키워드
- Drug discovery
- 키워드
- Graph pooling
- 키워드
- Graph structures
- 기타저자
- Northwestern University Computer Science
- 기본자료저록
- Dissertations Abstracts International. 85-10B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017160372
■00520250211151006
■006m o d
■007cr#unu||||||||
■020 ▼a9798381975192
■035 ▼a(MiAaPQ)AAI30994853
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a540
■1001 ▼aThieme, Mattson.
■24510▼aAI for Science: Graph Machine Learning as an Instrument for Understanding, Controlling, and Creating Physical Systems
■260 ▼a[Sl]▼bNorthwestern University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a198 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 85-10, Section: B.
■500 ▼aAdvisor: Liu, Han.
■5021 ▼aThesis (Ph.D.)--Northwestern University, 2024.
■520 ▼aWe consider machine learning (ML) to be a programmatic analog of the scientific method. In this paradigm, a mapping encoded in the model weights (a theory) transforms input data (observations) into predictions, which are then compared against a target (validating the theory against reality). As training proceeds and the correspondence between out-of-sample predictions and reality grows, so too does our confidence that the learned mapping reflects some truth about the problem domain. If this agreement is sufficiently robust, an examination of the mapping may then yield insights into the rules and relationships governing the system under study. Inspired by the potential for ML to accelerate scientific discovery, we explore differentiable, graph-based algorithms that allow neural models to more efficiently extract and leverage latent relational information from raw data. Our focus on graph-based algorithms not only makes the work more general but also more readily applicable to the physical sciences, where graphs are the de facto representational structure. We introduce novel methods for 1) graph structure learning, where we fix the nodes and learn to retain/remove the edges of a graph; 2) graph partition learning, where we fix the edges and learn to cluster the nodes; and 3) molecular graph generation with a novel graph adaptation methodology, where we fix the edges and optimize graph-level properties by modifying node features. Finally, in the spirit of concretely accelerating scientific discovery with machine learning, we discuss our extensive work on AI for Science conducted in partnership with Fermilab, where we repurposed biomedical segmentation models for disentangling particle accelerator loss profiles, and recurrent sequence models for controlling high-frequency proton beam extractors.
■590 ▼aSchool code: 0163.
■650 4▼aChemistry
■650 4▼aPharmacology
■650 4▼aComputer science
■650 4▼aBioinformatics
■653 ▼aGraph machine learning
■653 ▼aDrug discovery
■653 ▼aLearning on graphs
■653 ▼aGraph pooling
■653 ▼aGraph structures
■690 ▼a0800
■690 ▼a0984
■690 ▼a0419
■690 ▼a0715
■690 ▼a0485
■71020▼aNorthwestern University▼bComputer Science.
■7730 ▼tDissertations Abstracts International▼g85-10B.
■790 ▼a0163
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17160372▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
