서브메뉴
검색
Understanding the Mechanism of Pretraining Stabilization Heuristics: A Variance-Oriented Perspective
Understanding the Mechanism of Pretraining Stabilization Heuristics: A Variance-Oriented Perspective
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105703
- ISBN
- 9798263308032
- DDC
- 004
- 저자명
- Liu, Liyuan.
- 서명/저자
- Understanding the Mechanism of Pretraining Stabilization Heuristics: A Variance-Oriented Perspective
- 발행사항
- [Sl] : University of Illinois at Urbana-Champaign, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 112 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: B.
- 주기사항
- Advisor: Han, Jiawei.
- 학위논문주기
- Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2024.
- 초록/해제
- 요약Language model pretraining has been breaking the glass ceiling for various natural language processing tasks and has been viewed as one of the most significant successes of deep learning, continuously challenging our understanding of learning and cognition. Recently models, including GPT-4 and BART, fueled by an unprecedented scale of computing and data, exhibit unprecedented intelligence, that some even refer to as "sparks of artificial general intelligence". The success of large-scale pretraining hinges on intricate engineering heuristics. While the empirical benefits of these heuristics are evident, their underlying mechanisms remain elusive. This dissertation endeavors to demystify the mathematical principles underlying these pretraining heuristics, aiming to illuminate their mechanisms and potentially guide future algorithm developments. Adopting a variance-oriented perspective, my research rigorously inspects the heuristics that are pivotal to the stability of current pretraining practices, emphasizing learning rate warmup, model initialization, and gradient approximation. In this dissertation, I show that these pretraining stabilization heuristics can be coherently elucidated with a unified framework anchored in variance, a classical metric for stability. First, I analyze the variance of adaptive learning rate and model outputs, revealing that both learning rate warmup and model initialization function as variance modulators. Then, I move to explore the variance-bias tradeoff in the discrete variable gradient approximation, i.e., employing a numerical ODE framework, I unveil the underlying dynamics of the approximation bias, achieving second order precision with minimal computational overhead. Besides theoretical results, empirical verifications are conducted to verify the assumptions and applicability of the recognized principles. Building upon these insights, this dissertation introduces novel techniques designed to advance the pretraining practices, including RAdam for learning rate warmup, Admin for Transformer model initialization, ReinMax and SparseMixer for gradient approximation. Under the guidance of the recognized principles, all proposed methods require minimal trial-and-error configurations, thereby emerging as robust and high-perform tools for pretraining practices for adaptations.
- 일반주제명
- Computer science
- 일반주제명
- Applied mathematics
- 키워드
- Variance
- 기타저자
- University of Illinois at Urbana-Champaign Computer Science
- 기본자료저록
- Dissertations Abstracts International. 87-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2024 us c eng d■001000017361083
■00520260202105703
■006m o d
■007cr#unu||||||||
■020 ▼a9798263308032
■035 ▼a(MiAaPQ)AAI32409894
■035 ▼a(MiAaPQ)124230
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aLiu, Liyuan.
■24510▼aUnderstanding the Mechanism of Pretraining Stabilization Heuristics: A Variance-Oriented Perspective
■260 ▼a[Sl]▼bUniversity of Illinois at Urbana-Champaign▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a112 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: B.
■500 ▼aAdvisor: Han, Jiawei.
■5021 ▼aThesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2024.
■520 ▼aLanguage model pretraining has been breaking the glass ceiling for various natural language processing tasks and has been viewed as one of the most significant successes of deep learning, continuously challenging our understanding of learning and cognition. Recently models, including GPT-4 and BART, fueled by an unprecedented scale of computing and data, exhibit unprecedented intelligence, that some even refer to as "sparks of artificial general intelligence". The success of large-scale pretraining hinges on intricate engineering heuristics. While the empirical benefits of these heuristics are evident, their underlying mechanisms remain elusive. This dissertation endeavors to demystify the mathematical principles underlying these pretraining heuristics, aiming to illuminate their mechanisms and potentially guide future algorithm developments. Adopting a variance-oriented perspective, my research rigorously inspects the heuristics that are pivotal to the stability of current pretraining practices, emphasizing learning rate warmup, model initialization, and gradient approximation. In this dissertation, I show that these pretraining stabilization heuristics can be coherently elucidated with a unified framework anchored in variance, a classical metric for stability. First, I analyze the variance of adaptive learning rate and model outputs, revealing that both learning rate warmup and model initialization function as variance modulators. Then, I move to explore the variance-bias tradeoff in the discrete variable gradient approximation, i.e., employing a numerical ODE framework, I unveil the underlying dynamics of the approximation bias, achieving second order precision with minimal computational overhead. Besides theoretical results, empirical verifications are conducted to verify the assumptions and applicability of the recognized principles. Building upon these insights, this dissertation introduces novel techniques designed to advance the pretraining practices, including RAdam for learning rate warmup, Admin for Transformer model initialization, ReinMax and SparseMixer for gradient approximation. Under the guidance of the recognized principles, all proposed methods require minimal trial-and-error configurations, thereby emerging as robust and high-perform tools for pretraining practices for adaptations.
■590 ▼aSchool code: 0090.
■650 4▼aComputer science
■650 4▼aApplied mathematics
■653 ▼aTraining stability
■653 ▼aVariance
■653 ▼aModel initialization
■690 ▼a0984
■690 ▼a0364
■71020▼aUniversity of Illinois at Urbana-Champaign▼bComputer Science.
■7730 ▼tDissertations Abstracts International▼g87-05B.
■790 ▼a0090
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17361083▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


