서브메뉴
검색
Trustworthiness : Revisiting the Foundations of Machine Learning
Trustworthiness : Revisiting the Foundations of Machine Learning
Detailed Information
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152121
- ISBN
- 9798384341949
- DDC
- 616
- 저자명
- Hu, Lunjia.
- 서명/저자
- Trustworthiness : Revisiting the Foundations of Machine Learning
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 239 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-03, Section: A.
- 주기사항
- Advisor: Charikar, Moses;Reingold, Omer.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약As we deploy machine learning models in more and more complex and critical tasks, can we trust that these models will work as intended in real-world settings? For example, a doctor might use a machine learning model to evaluate patients' risk of developing specific diseases and make consequential treatment decisions based on the model's predictions. There are significant barriers to building such trust-machine learning predictions are frustratingly inscrutable, concerns about unfair treatment of minority groups abound, and strong performance in idealized settings often breaks down on realworld data. To be trustworthy, a model must provide guarantees beyond the standard objectives of accuracy and loss minimization (i.e., minimizing a function that penalizes overall error). A trustworthy model should give interpretable explanations for its predictions, should be fair to patients from protected subpopulations, and should be robuston real-world data that does not satisfy idealized assumptions. These considerations have motivated a rapid growth of research on trustworthy machine learning.This dissertation presents my research that develops new mathematical theories to address fundamental problems in trustworthy machine learning, and also conversely, applies the insights from trustworthiness to discover new foundational theories for machine learning, computer science, and statistics.My research presented in this dissertation highlights the symbiotic relationship between studying practical issues of trustworthiness and building the theoretical foundations of machine learning. On one hand, trustworthiness notions exhibit varying levels of complexity and subtlety, needing mathematical sophistication to avoid ambiguity and provide principled guidance. Historically, theoretical computer science has achieved great success in formalizing seemingly vague concepts such as secrecy, rationality, learning, and privacy. Following this tradition, my research has developed new theories to answer pressing questions in trustworthy machine learning that are beyond the reach of empirical research alone. Chapter 1 presents a theory of calibration that has enabled an open-source Python package [B lasiok and Nakkiran, 2023] solving problems that caused common confusion and frustration in previous research. On the other hand, the versatility of trustworthiness considerations is a iv vibrant source of inspiration for new foundational theories. Several lines of my research have uncovered surprisingly richer theories than previously known for classic problems in machine learning and statistics: Chapter 2 presents new sample complexity theories generalizing the classic VC theory, whereas Chapter 3 presents new theories of omnipredictionwhich allows efficient training of a single prediction model that is easily adaptable to optimize a rich family of loss functions potentially with varying constraints.
- 일반주제명
- Cardiovascular disease
- 일반주제명
- Probability
- 일반주제명
- Computer science
- 일반주제명
- Distance learning
- 일반주제명
- Neural networks
- 일반주제명
- Educational technology
- 일반주제명
- Medicine
- 일반주제명
- Public health
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-03A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017162994
■00520250211152121
■006m o d
■007cr#unu||||||||
■020 ▼a9798384341949
■035 ▼a(MiAaPQ)AAI31460394
■035 ▼a(MiAaPQ)Stanfordzy428st6966
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a616
■1001 ▼aHu, Lunjia.
■24510▼aTrustworthiness : Revisiting the Foundations of Machine Learning
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a239 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: A.
■500 ▼aAdvisor: Charikar, Moses;Reingold, Omer.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aAs we deploy machine learning models in more and more complex and critical tasks, can we trust that these models will work as intended in real-world settings? For example, a doctor might use a machine learning model to evaluate patients' risk of developing specific diseases and make consequential treatment decisions based on the model's predictions. There are significant barriers to building such trust-machine learning predictions are frustratingly inscrutable, concerns about unfair treatment of minority groups abound, and strong performance in idealized settings often breaks down on realworld data. To be trustworthy, a model must provide guarantees beyond the standard objectives of accuracy and loss minimization (i.e., minimizing a function that penalizes overall error). A trustworthy model should give interpretable explanations for its predictions, should be fair to patients from protected subpopulations, and should be robuston real-world data that does not satisfy idealized assumptions. These considerations have motivated a rapid growth of research on trustworthy machine learning.This dissertation presents my research that develops new mathematical theories to address fundamental problems in trustworthy machine learning, and also conversely, applies the insights from trustworthiness to discover new foundational theories for machine learning, computer science, and statistics.My research presented in this dissertation highlights the symbiotic relationship between studying practical issues of trustworthiness and building the theoretical foundations of machine learning. On one hand, trustworthiness notions exhibit varying levels of complexity and subtlety, needing mathematical sophistication to avoid ambiguity and provide principled guidance. Historically, theoretical computer science has achieved great success in formalizing seemingly vague concepts such as secrecy, rationality, learning, and privacy. Following this tradition, my research has developed new theories to answer pressing questions in trustworthy machine learning that are beyond the reach of empirical research alone. Chapter 1 presents a theory of calibration that has enabled an open-source Python package [B lasiok and Nakkiran, 2023] solving problems that caused common confusion and frustration in previous research. On the other hand, the versatility of trustworthiness considerations is a iv vibrant source of inspiration for new foundational theories. Several lines of my research have uncovered surprisingly richer theories than previously known for classic problems in machine learning and statistics: Chapter 2 presents new sample complexity theories generalizing the classic VC theory, whereas Chapter 3 presents new theories of omnipredictionwhich allows efficient training of a single prediction model that is easily adaptable to optimize a rich family of loss functions potentially with varying constraints.
■590 ▼aSchool code: 0212.
■650 4▼aCardiovascular disease
■650 4▼aProbability
■650 4▼aComputer science
■650 4▼aDistance learning
■650 4▼aNeural networks
■650 4▼aEducational technology
■650 4▼aMedicine
■650 4▼aPublic health
■690 ▼a0984
■690 ▼a0800
■690 ▼a0710
■690 ▼a0564
■690 ▼a0573
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-03A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17162994▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
Preview
Export
ChatGPT Discussion
AI Recommended Related Books
Подробнее информация.
- Бронирование
- не существует
- моя папка
- Первый запрос зрения
- Non-Book Loan Application
- Nighttime Book Loan Application
Available after logging in.


