서브메뉴
검색
Learning with and Without Human Feedback
Learning with and Without Human Feedback
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105556
- ISBN
- 9798265406668
- DDC
- 404
- 서명/저자
- Learning with and Without Human Feedback
- 발행사항
- [Sl] : Georgia Institute of Technology, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 201 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: B.
- 주기사항
- Advisor: Davenport, Mark.
- 학위논문주기
- Thesis (Ph.D.)--Georgia Institute of Technology, 2024.
- 초록/해제
- 요약The development of contemporary machine learning (ML) models is driven, in part, by the availability and volume of labeled training data. Labels provided by humans play a central role in this training pipeline, offering models ground truth annotations from which to extract patterns. However, collecting such feedback from humans is a challenging and time-consuming task. As a result, practitioners must be intentional both in how they choose to query humans for feedback and in the problem settings for which they request feedback.This thesis explores learning from human feedback along two fundamental directions. The first part of the thesis focuses on how we can more effectively learn from and collect human feedback from a mathematically grounded perspective. In Chapter 2, we consider the paired comparison, a simple mechanism for collecting human feedback, and show that paired comparison responses are capable of estimating a much richer parametrization of user preferences than previously established [1]. In Chapter 3, we propose a new mechanism for collecting human feedback called the perceptual adjustment query [2] designed to balance informativeness and cognitive burden. We apply perceptual adjustment queries to a human perception model parametrized by a low-rank metric and rigorously prove estimation error bounds.The second part focuses on how we can leverage pretrained models to avoid collecting additional human feedback. In Chapter 4, we consider the cold-start phase of a recommender system, where no user relevance feedback is available to train a retrieval model. Using the generative abilities of large language models, we design a retrieval framework capable of retrieving relevant text for users without any human relevance feedback [3]. In Chapter 5, we improve synthetic image dataset generation by removing the need for humans-in-the-loop [4]. Existing methods require human annotators to repeatedly label synthetically generated images; our proposed framework leverages tools from image editting to re-use existing labeled images, bypassing the need for human annotators.
- 일반주제명
- Multilingualism
- 일반주제명
- Large language models
- 일반주제명
- Visualization
- 일반주제명
- Semantics
- 일반주제명
- Bilingual education
- 기본자료저록
- Dissertations Abstracts International. 87-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2024 us c eng d■001000017360620
■00520260202105556
■006m o d
■007cr#unu||||||||
■020 ▼a9798265406668
■035 ▼a(MiAaPQ)AAI32315896
■035 ▼a(MiAaPQ)GeorgiaTech75217
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a404
■1001 ▼aXu, Austin Shiyi.
■24510▼aLearning with and Without Human Feedback
■260 ▼a[Sl]▼bGeorgia Institute of Technology▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a201 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: B.
■500 ▼aAdvisor: Davenport, Mark.
■5021 ▼aThesis (Ph.D.)--Georgia Institute of Technology, 2024.
■520 ▼aThe development of contemporary machine learning (ML) models is driven, in part, by the availability and volume of labeled training data. Labels provided by humans play a central role in this training pipeline, offering models ground truth annotations from which to extract patterns. However, collecting such feedback from humans is a challenging and time-consuming task. As a result, practitioners must be intentional both in how they choose to query humans for feedback and in the problem settings for which they request feedback.This thesis explores learning from human feedback along two fundamental directions. The first part of the thesis focuses on how we can more effectively learn from and collect human feedback from a mathematically grounded perspective. In Chapter 2, we consider the paired comparison, a simple mechanism for collecting human feedback, and show that paired comparison responses are capable of estimating a much richer parametrization of user preferences than previously established [1]. In Chapter 3, we propose a new mechanism for collecting human feedback called the perceptual adjustment query [2] designed to balance informativeness and cognitive burden. We apply perceptual adjustment queries to a human perception model parametrized by a low-rank metric and rigorously prove estimation error bounds.The second part focuses on how we can leverage pretrained models to avoid collecting additional human feedback. In Chapter 4, we consider the cold-start phase of a recommender system, where no user relevance feedback is available to train a retrieval model. Using the generative abilities of large language models, we design a retrieval framework capable of retrieving relevant text for users without any human relevance feedback [3]. In Chapter 5, we improve synthetic image dataset generation by removing the need for humans-in-the-loop [4]. Existing methods require human annotators to repeatedly label synthetically generated images; our proposed framework leverages tools from image editting to re-use existing labeled images, bypassing the need for human annotators.
■590 ▼aSchool code: 0078.
■650 4▼aMultilingualism
■650 4▼aLarge language models
■650 4▼aVisualization
■650 4▼aSemantics
■650 4▼aBilingual education
■690 ▼a0800
■690 ▼a0282
■71020▼aGeorgia Institute of Technology.
■7730 ▼tDissertations Abstracts International▼g87-05B.
■790 ▼a0078
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360620▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
