서브메뉴
검색
Algorithmic Statistical Learning and Causality Pursuit Using Neural Networks
Algorithmic Statistical Learning and Causality Pursuit Using Neural Networks
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103521
- ISBN
- 9798280747371
- DDC
- 310
- 저자명
- Gu, Yihong.
- 서명/저자
- Algorithmic Statistical Learning and Causality Pursuit Using Neural Networks
- 발행사항
- [Sl] : Princeton University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 744 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
- 주기사항
- Advisor: Fan, Jianqing.
- 학위논문주기
- Thesis (Ph.D.)--Princeton University, 2025.
- 초록/해제
- 요약Statistics suffer from two fundamental problems, `the curse of dimensionality' -- the difficulty of accurately learning associations in high-dimensional data -- and `the curse of endogeneity', where the learned associations do not necessarily imply causation. Traditional solutions often rely heavily on prior knowledge, such as function structure and cause-effect direction, and may risk model misspecification. This dissertation contributes to (1) developing innovative algorithmic methods that can attain provably sample-efficient estimation blind to prior knowledge and (2) providing the corresponding theoretical insights.The first part of this dissertation contributes to statistical estimation using neural networks. In Chapter 2, we reveal another side of neural networks' super approximation ability, which makes them vulnerable to heavy-tailed noises. This message is rigorously delivered by establishing matching upper and lower error bounds, that are slower than those under sub-Gaussian counterparts, for neural network least squares and adaptive Huber estimators under heavy-tailed noise. In Chapter 3, we introduce the Factor Augmented Sparse Throughput model for high-dimensional regression and a corresponding structured neural network, which can achieve statistically efficient estimation with high-dimensional and highly correlated covariates. The second part proposes novel methods that utilize data from heterogeneous environments to learn causality rather than just association in a data-driven way. The key idea, which mimics humans' understanding of causality, is the causal law will be invariant across time, space, or more broadly environments to some extent. We develop provably sample-efficient invariance learning methods for linear and nonparametric models in Chapter 4 and Chapter 5, respectively. This marks a pioneering step in the invariance learning field. We also prove the intrinsic computational barriers in Chapter 6 and propose a remedy that balances computation and statistical accuracy on the one hand and trade-offs robustness and predictive power on the other hand.
- 일반주제명
- Statistics
- 키워드
- Causality
- 키워드
- Factor model
- 키워드
- Invariance
- 키워드
- Neural networks
- 기타저자
- Princeton University Operations Research and Financial Engineering
- 기본자료저록
- Dissertations Abstracts International. 86-12B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357501
■00520260202103521
■006m o d
■007cr#unu||||||||
■020 ▼a9798280747371
■035 ▼a(MiAaPQ)AAI32038683
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a310
■1001 ▼aGu, Yihong.
■24510▼aAlgorithmic Statistical Learning and Causality Pursuit Using Neural Networks
■260 ▼a[Sl]▼bPrinceton University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a744 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-12, Section: B.
■500 ▼aAdvisor: Fan, Jianqing.
■5021 ▼aThesis (Ph.D.)--Princeton University, 2025.
■520 ▼aStatistics suffer from two fundamental problems, `the curse of dimensionality' -- the difficulty of accurately learning associations in high-dimensional data -- and `the curse of endogeneity', where the learned associations do not necessarily imply causation. Traditional solutions often rely heavily on prior knowledge, such as function structure and cause-effect direction, and may risk model misspecification. This dissertation contributes to (1) developing innovative algorithmic methods that can attain provably sample-efficient estimation blind to prior knowledge and (2) providing the corresponding theoretical insights.The first part of this dissertation contributes to statistical estimation using neural networks. In Chapter 2, we reveal another side of neural networks' super approximation ability, which makes them vulnerable to heavy-tailed noises. This message is rigorously delivered by establishing matching upper and lower error bounds, that are slower than those under sub-Gaussian counterparts, for neural network least squares and adaptive Huber estimators under heavy-tailed noise. In Chapter 3, we introduce the Factor Augmented Sparse Throughput model for high-dimensional regression and a corresponding structured neural network, which can achieve statistically efficient estimation with high-dimensional and highly correlated covariates. The second part proposes novel methods that utilize data from heterogeneous environments to learn causality rather than just association in a data-driven way. The key idea, which mimics humans' understanding of causality, is the causal law will be invariant across time, space, or more broadly environments to some extent. We develop provably sample-efficient invariance learning methods for linear and nonparametric models in Chapter 4 and Chapter 5, respectively. This marks a pioneering step in the invariance learning field. We also prove the intrinsic computational barriers in Chapter 6 and propose a remedy that balances computation and statistical accuracy on the one hand and trade-offs robustness and predictive power on the other hand.
■590 ▼aSchool code: 0181.
■650 4▼aStatistics
■653 ▼aCausality
■653 ▼aFactor model
■653 ▼aHeavy-tailed noise
■653 ▼aInvariance
■653 ▼aNeural networks
■653 ▼aNonparametric estimation
■690 ▼a0463
■690 ▼a0796
■690 ▼a0800
■71020▼aPrinceton University▼bOperations Research and Financial Engineering.
■7730 ▼tDissertations Abstracts International▼g86-12B.
■790 ▼a0181
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357501▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
