서브메뉴
검색
Building More Reliable and Scalable AI Systems with Foundation Model Programming
Building More Reliable and Scalable AI Systems with Foundation Model Programming
Detailed Information
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104746
- ISBN
- 9798290652214
- DDC
- 400
- 서명/저자
- Building More Reliable and Scalable AI Systems with Foundation Model Programming
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 140 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: A.
- 주기사항
- Advisor: Manning, Christopher;Potts, Christopher;Zaharia, Matei.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약Language models (LMs), and more broadly foundation models, have made it much easier to prototype impressive AI demos, but turning these models into reliable and scalable AI systems remains challenging. Indeed, today's user-facing LMs often fabricate statements and make fundamental reasoning errors, while imposing enormous costs. We argue that the monolithic nature of LMs is what makes them hard to control, debug, and improve in modular and iterative ways when building AI systems.To address this, this dissertation introduces foundation model programming,a new paradigm centered on designing controllable systems that use LMs as modular fuzzy components and teach them the right system-level behavior via trial and error. To support this, we introduce (1) a new paradigm of information retrieval (IR) models by composing rich yet pruning-friendly interactions between token-level LM representations, (2) a new generation of scalable retrieval-based reasoning systems,and (3) a powerful compiler for natural language programming abstractions. We support these by proposing new algorithms for efficiently searching LM-based representations at scale, for weakly supervising natural language processing (NLP) systems with incomplete labels, and for optimizing the prompts of LMs inside arbitrary programs.This culminates in the ColBERTparadigm for neural IR and in the DSPyframework for natural language programming. We demonstrate through these contributions that making broad progress in AI is not restricted to training larger models, but can take the form of designing general toolsthat grant AI researchers and developers the capacity to controllably improve their systems via composition, to transparently ground their systems' responses in massive knowledge collections, and to scalably deploy their systems via new algorithms and new compositions of smaller LMs.
- 일반주제명
- Language
- 일반주제명
- Software
- 일반주제명
- Relevance
- 일반주제명
- Hallucinations
- 일반주제명
- Documents
- 일반주제명
- Information retrieval
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-01A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2024 us c eng d■001000017358751
■00520260202104746
■006m o d
■007cr#unu||||||||
■020 ▼a9798290652214
■035 ▼a(MiAaPQ)AAI32149759
■035 ▼a(MiAaPQ)Stanfordzc625xj7842
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a400
■1001 ▼aKhattab, Omar Ahmed Metwally.
■24510▼aBuilding More Reliable and Scalable AI Systems with Foundation Model Programming
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a140 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: A.
■500 ▼aAdvisor: Manning, Christopher;Potts, Christopher;Zaharia, Matei.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aLanguage models (LMs), and more broadly foundation models, have made it much easier to prototype impressive AI demos, but turning these models into reliable and scalable AI systems remains challenging. Indeed, today's user-facing LMs often fabricate statements and make fundamental reasoning errors, while imposing enormous costs. We argue that the monolithic nature of LMs is what makes them hard to control, debug, and improve in modular and iterative ways when building AI systems.To address this, this dissertation introduces foundation model programming,a new paradigm centered on designing controllable systems that use LMs as modular fuzzy components and teach them the right system-level behavior via trial and error. To support this, we introduce (1) a new paradigm of information retrieval (IR) models by composing rich yet pruning-friendly interactions between token-level LM representations, (2) a new generation of scalable retrieval-based reasoning systems,and (3) a powerful compiler for natural language programming abstractions. We support these by proposing new algorithms for efficiently searching LM-based representations at scale, for weakly supervising natural language processing (NLP) systems with incomplete labels, and for optimizing the prompts of LMs inside arbitrary programs.This culminates in the ColBERTparadigm for neural IR and in the DSPyframework for natural language programming. We demonstrate through these contributions that making broad progress in AI is not restricted to training larger models, but can take the form of designing general toolsthat grant AI researchers and developers the capacity to controllably improve their systems via composition, to transparently ground their systems' responses in massive knowledge collections, and to scalably deploy their systems via new algorithms and new compositions of smaller LMs.
■590 ▼aSchool code: 0212.
■650 4▼aLanguage
■650 4▼aSoftware
■650 4▼aRelevance
■650 4▼aNatural language processing
■650 4▼aHallucinations
■650 4▼aDocuments
■650 4▼aRetrieval performance measures
■650 4▼aInformation retrieval
■690 ▼a0679
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-01A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358751▼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.


