서브메뉴
검색
Statistical Physics of Gradient Descent in High-Dimensions: From Coherent Ising Machines to Neural Networks
Statistical Physics of Gradient Descent in High-Dimensions: From Coherent Ising Machines to Neural Networks
Detailed Information
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105626
- ISBN
- 9798265428004
- DDC
- 530.13
- 서명/저자
- Statistical Physics of Gradient Descent in High-Dimensions: From Coherent Ising Machines to Neural Networks
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 183 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: B.
- 주기사항
- Advisor: Ganguli, Surya;Mabuchi, Hideo.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약Gradient descent and its variants are widely employed in optimization and machine learning tasks. A fundamental understanding of their behavior in high-dimensional loss or energy landscapes is essential for improving performance and efficiency. In the first part of this dissertation, we analyze the optimization dynamics of coherent Ising machines (CIMs)-optical recurrent neural networks designed to solve combinatorial optimization problems. By investigating their behavior on random combinatorial instances through statistical physics methods, we identify distinct phase transitions in the energy landscape of CIMs, from flat to rough to rigid. These transitions significantly influence gradient descent dynamics, providing valuable insights into optimal hyperparameter tuning for enhanced CIM performance.In the second part, we explore the dynamics of stochastic gradient descent (SGD) in deep neural networks, focusing on the fundamental question of why these networks exhibit remarkable generalization capabilities. We demonstrate that the stochastic noise intrinsic to SGD possesses a structured form determined by the architectural symmetries of neural networks. This structured noise guides SGD dynamics toward specific subsets of parameter space characterized by "simpler" functions, thereby mitigating overfitting and enhancing the model's ability to generalize to unseen data.
- 일반주제명
- Brownian motion
- 일반주제명
- Semiconductors
- 일반주제명
- Lasers
- 일반주제명
- Neural networks
- 일반주제명
- Symmetry
- 일반주제명
- Phase transitions
- 일반주제명
- Statistical physics
- 일반주제명
- Connectivity
- 일반주제명
- Eigenvalues
- 일반주제명
- Energy
- 일반주제명
- Eigenvectors
- 일반주제명
- Geometry
- 일반주제명
- Mathematics
- 일반주제명
- Optics
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017360838
■00520260202105626
■006m o d
■007cr#unu||||||||
■020 ▼a9798265428004
■035 ▼a(MiAaPQ)AAI32316558
■035 ▼a(MiAaPQ)Stanfordnx340sz2444
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a530.13
■1001 ▼aYamamura, Atsushi.
■24510▼aStatistical Physics of Gradient Descent in High-Dimensions: From Coherent Ising Machines to Neural Networks
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a183 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: B.
■500 ▼aAdvisor: Ganguli, Surya;Mabuchi, Hideo.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aGradient descent and its variants are widely employed in optimization and machine learning tasks. A fundamental understanding of their behavior in high-dimensional loss or energy landscapes is essential for improving performance and efficiency. In the first part of this dissertation, we analyze the optimization dynamics of coherent Ising machines (CIMs)-optical recurrent neural networks designed to solve combinatorial optimization problems. By investigating their behavior on random combinatorial instances through statistical physics methods, we identify distinct phase transitions in the energy landscape of CIMs, from flat to rough to rigid. These transitions significantly influence gradient descent dynamics, providing valuable insights into optimal hyperparameter tuning for enhanced CIM performance.In the second part, we explore the dynamics of stochastic gradient descent (SGD) in deep neural networks, focusing on the fundamental question of why these networks exhibit remarkable generalization capabilities. We demonstrate that the stochastic noise intrinsic to SGD possesses a structured form determined by the architectural symmetries of neural networks. This structured noise guides SGD dynamics toward specific subsets of parameter space characterized by "simpler" functions, thereby mitigating overfitting and enhancing the model's ability to generalize to unseen data.
■590 ▼aSchool code: 0212.
■650 4▼aBrownian motion
■650 4▼aSemiconductors
■650 4▼aLasers
■650 4▼aNeural networks
■650 4▼aSymmetry
■650 4▼aPhase transitions
■650 4▼aStatistical physics
■650 4▼aConnectivity
■650 4▼aEigenvalues
■650 4▼aEnergy
■650 4▼aTraveling salesman problem
■650 4▼aEigenvectors
■650 4▼aGeometry
■650 4▼aMathematics
■650 4▼aOptics
■690 ▼a0791
■690 ▼a0217
■690 ▼a0800
■690 ▼a0405
■690 ▼a0796
■690 ▼a0752
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-05B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360838▼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.


