서브메뉴
검색
Algorithms for Robust Learning, Gradient Flows, and Diffusion Generation of Rare Events
Algorithms for Robust Learning, Gradient Flows, and Diffusion Generation of Rare Events
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152747
- ISBN
- 9798342107211
- DDC
- 510
- 저자명
- Hua, Xinru.
- 서명/저자
- Algorithms for Robust Learning, Gradient Flows, and Diffusion Generation of Rare Events
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 125 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-04, Section: B.
- 주기사항
- Advisor: Mancilla, Jose Blanchet;Ma, Tengyu.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약As machine learning systems become an integral part of our daily lives, particularly with the remarkable advancements of foundation models, we must assess their trustworthiness, fairness, and also exploring methods to improve these crucial aspects. This dissertation explores the measurement and enhancement of a machine learning model's performance, robustness, and fairness. Furthermore, we explore the design and deployment of these systems in novel application fields, with these metrics as crucial objectives.This thesis aims to make machine learning more trustable and powerful in general. The first topic is on evaluating the robustness and fairness of machine learning models and strategies to enhance them. Our method combines distributionally robust optimization (DRO) and human imperceptible adversarial attacks to simultaneously improve both model robustness and fairness. By analyzing the robustness and fairness of the machine learning systems, we intuitively bring machine perception closer to human perception. In addition to robustness and fairness, we study the gradient flow method to alleviate data scarcity problem and improve classification systems' performance in few-shot learning settings. We prove the gradient flow method converge globally and the downstream transfer learning tasks demonstrate its ability to generate useful data samples. Lastly, we focus on an innovative application of machine learning algorithms in the field of material science. Specifically, we design machine learning systems to accelerate the sampling of rare events in molecular simulations. Our approach yields substantial speed improvements compared to traditional sampling methods, along with robust estimations of the probability of these rare events.Throughout the thesis, we demonstrate that machine learning can be improved in many aspects, including fairness and robustness. We also showcase its power in traditional applications, such as mechanical simulations. Future work will extend these systems to tackle more complex and higher-dimensional challenges. With the ongoing efforts, the thesis contributes to the development of more reliable and potent machine learning systems.
- 일반주제명
- Sample size
- 일반주제명
- Deep learning
- 일반주제명
- Neural networks
- 일반주제명
- Design
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-04B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017163742
■00520250211152747
■006m o d
■007cr#unu||||||||
■020 ▼a9798342107211
■035 ▼a(MiAaPQ)AAI31520314
■035 ▼a(MiAaPQ)Stanfordsz944xc0403
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a510
■1001 ▼aHua, Xinru.
■24510▼aAlgorithms for Robust Learning, Gradient Flows, and Diffusion Generation of Rare Events
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a125 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-04, Section: B.
■500 ▼aAdvisor: Mancilla, Jose Blanchet;Ma, Tengyu.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aAs machine learning systems become an integral part of our daily lives, particularly with the remarkable advancements of foundation models, we must assess their trustworthiness, fairness, and also exploring methods to improve these crucial aspects. This dissertation explores the measurement and enhancement of a machine learning model's performance, robustness, and fairness. Furthermore, we explore the design and deployment of these systems in novel application fields, with these metrics as crucial objectives.This thesis aims to make machine learning more trustable and powerful in general. The first topic is on evaluating the robustness and fairness of machine learning models and strategies to enhance them. Our method combines distributionally robust optimization (DRO) and human imperceptible adversarial attacks to simultaneously improve both model robustness and fairness. By analyzing the robustness and fairness of the machine learning systems, we intuitively bring machine perception closer to human perception. In addition to robustness and fairness, we study the gradient flow method to alleviate data scarcity problem and improve classification systems' performance in few-shot learning settings. We prove the gradient flow method converge globally and the downstream transfer learning tasks demonstrate its ability to generate useful data samples. Lastly, we focus on an innovative application of machine learning algorithms in the field of material science. Specifically, we design machine learning systems to accelerate the sampling of rare events in molecular simulations. Our approach yields substantial speed improvements compared to traditional sampling methods, along with robust estimations of the probability of these rare events.Throughout the thesis, we demonstrate that machine learning can be improved in many aspects, including fairness and robustness. We also showcase its power in traditional applications, such as mechanical simulations. Future work will extend these systems to tackle more complex and higher-dimensional challenges. With the ongoing efforts, the thesis contributes to the development of more reliable and potent machine learning systems.
■590 ▼aSchool code: 0212.
■650 4▼aSample size
■650 4▼aDeep learning
■650 4▼aNeural networks
■650 4▼aDesign
■690 ▼a0389
■690 ▼a0800
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-04B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17163742▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


