서브메뉴
검색
Statistical Learning Theory of Deep Neural Networks: A Generalization Viewpoint
Statistical Learning Theory of Deep Neural Networks: A Generalization Viewpoint
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105531
- ISBN
- 9798263342708
- DDC
- 005.8
- 저자명
- Zhou, Tian-Yi.
- 서명/저자
- Statistical Learning Theory of Deep Neural Networks: A Generalization Viewpoint
- 발행사항
- [Sl] : Georgia Institute of Technology, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 268 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: B.
- 주기사항
- Advisor: Huo, Xiaoming.
- 학위논문주기
- Thesis (Ph.D.)--Georgia Institute of Technology, 2025.
- 초록/해제
- 요약Neural networks are computing systems with powerful applications in many disciplines, e.g.,data analysis and pattern and sequence recognition. In particular, deep neural networks with welldesigned structures, numerous trainable parameters, and massive-scale input data have outstandingperformances in function approximation [1, 2], classification [3, 4], regression [5], and featureextraction [6]. But beneath their advertised success, deep neural networks are brittle, black-boxsystems sensitive to observation errors and changes in data distribution. In everyday life, failuresof deep learning systems can be a nuisance: perhaps Siri misinterprets your question, or GoogleMaps underestimates the time you spend in traffic. However, these failures can jeopardize humanlives and infrastructure in critical applications like healthcare and cybersecurity. To ensure trusted,robust deployment of deep learning, we must characterize these failure modes by developing anunderstanding of model generalization and performance in realistic settings.This thesis contributes to the mathematical foundation of deep learning by establishing statistical guarantees on the efficiency of deep neural networks in regression and classification tasks.Specifically, it establishes theoretical guarantees as to when and how deep neural networks generalize unseen data well for these tasks. Our results rely on explicit constructions of neural networks,leading to fast convergence rates of excess risks, in that tools from approximation theory and statistical learning theory are adopted.Chapter 1: Learning Ability of Interpolating Deep Convolutional Neural Networks The firstchapter of this thesis studies an important family of neural networks known as convolutional neuralnetworks (CNNs). Given that neural networks, in general, are powerful and versatile, researchershave been working to improve their computational efficiency further. When the data dimension islarge, such as the AlexNet [3] of input dimension about 150, 000, fully-connected neural networksare not feasible. Structures are often imposed on neural networks to reduce the number of trainablefree parameters and get feasible deep learning algorithms for various practical tasks [7]. The structure we are interested in is induced by 1-D convolution, and the resulting networks are deepconvolutional neural networks [8]. The convolutional structure of CNN reduces the computationalcomplexity and is believed to capture local shift-invariance properties of image and speech data.Such features of CNNs contribute to their massive popularity in image processing and speechrecognition.It is frequently observed that overparameterized neural networks generalize well while achieving zero training error [9]. This phenomenon, known as benign overfitting, seems to confrontthe classical bias-variance trade-off in statistical theory. Such a mismatch between observationsand classical theory sparked avid research attempting to understand how benign overfitting occurs.Theoretical work studying benign overfitting was initiated in [10], where a linear regression settingwith Gaussian data and noise was considered. It presented conditions for minimum-norm interpolators to generalize well. In a non-linear setting induced by the ReLU activation function, benignoverfitting is previously verified for fully-connected networks in [11]. On top of that, a recentwork [12] shows that training shallow neural networks with shared weights by gradient descentcan achieve an arbitrarily small training error.Our work studies the learning rate (generalization performance) of CNNs under underparameterized and overparameterized settings, respectively. We use the term learning rates to refer tothe rates of convergence of the output function to the regression function. We study the benignoverfitting problem in a regression setting. Our goal is to show that, for any input data, there always exist some CNNs that generalize well (achieve a sufficient fast learning rate) and, at the sametime, perfectly fit the data. In other words, we intend to verify the possible occurrence of benignoverfitting among deep CNNs. This work is published in Applied and Computational HarmonicAnalysis, 2024.Chapter 2: Classification of Unbounded Data by Gaussian Mixture Models via deep ReLUNetworks The second chapter of this thesis studies the binary classification of unbounded datagenerated by a mixture of Gaussian distributions using neural networks. With input X ∈ Rdand output label Y ∈ {−1, 1}, we are interested in learning a binary classifier f : Rd → {−1, 1}.We assume our data in R dfollows a class of distribution largely used to model real-world data,namely the Gaussian Mixture Model (GMM). According to [13, Chapter 3], a GMM is a universalapproximator of densities. On top of that, many studies have shown that GMM is an effectivemodel for audio, speech, image, and text processing, e.g., see [14, 15, 16]. The universality ofGMM motivates us to study the classification problem under such distributional assumptions ondata.
- 일반주제명
- Cybersecurity
- 일반주제명
- Deep learning
- 일반주제명
- Voice recognition
- 일반주제명
- Neural networks
- 일반주제명
- Computer science
- 기본자료저록
- Dissertations Abstracts International. 87-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017360469
■00520260202105531
■006m o d
■007cr#unu||||||||
■020 ▼a9798263342708
■035 ▼a(MiAaPQ)AAI32309849
■035 ▼a(MiAaPQ)GeorgiaTech77871
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a005.8
■1001 ▼aZhou, Tian-Yi.
■24510▼aStatistical Learning Theory of Deep Neural Networks: A Generalization Viewpoint
■260 ▼a[Sl]▼bGeorgia Institute of Technology▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a268 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: B.
■500 ▼aAdvisor: Huo, Xiaoming.
■5021 ▼aThesis (Ph.D.)--Georgia Institute of Technology, 2025.
■520 ▼aNeural networks are computing systems with powerful applications in many disciplines, e.g.,data analysis and pattern and sequence recognition. In particular, deep neural networks with welldesigned structures, numerous trainable parameters, and massive-scale input data have outstandingperformances in function approximation [1, 2], classification [3, 4], regression [5], and featureextraction [6]. But beneath their advertised success, deep neural networks are brittle, black-boxsystems sensitive to observation errors and changes in data distribution. In everyday life, failuresof deep learning systems can be a nuisance: perhaps Siri misinterprets your question, or GoogleMaps underestimates the time you spend in traffic. However, these failures can jeopardize humanlives and infrastructure in critical applications like healthcare and cybersecurity. To ensure trusted,robust deployment of deep learning, we must characterize these failure modes by developing anunderstanding of model generalization and performance in realistic settings.This thesis contributes to the mathematical foundation of deep learning by establishing statistical guarantees on the efficiency of deep neural networks in regression and classification tasks.Specifically, it establishes theoretical guarantees as to when and how deep neural networks generalize unseen data well for these tasks. Our results rely on explicit constructions of neural networks,leading to fast convergence rates of excess risks, in that tools from approximation theory and statistical learning theory are adopted.Chapter 1: Learning Ability of Interpolating Deep Convolutional Neural Networks The firstchapter of this thesis studies an important family of neural networks known as convolutional neuralnetworks (CNNs). Given that neural networks, in general, are powerful and versatile, researchershave been working to improve their computational efficiency further. When the data dimension islarge, such as the AlexNet [3] of input dimension about 150, 000, fully-connected neural networksare not feasible. Structures are often imposed on neural networks to reduce the number of trainablefree parameters and get feasible deep learning algorithms for various practical tasks [7]. The structure we are interested in is induced by 1-D convolution, and the resulting networks are deepconvolutional neural networks [8]. The convolutional structure of CNN reduces the computationalcomplexity and is believed to capture local shift-invariance properties of image and speech data.Such features of CNNs contribute to their massive popularity in image processing and speechrecognition.It is frequently observed that overparameterized neural networks generalize well while achieving zero training error [9]. This phenomenon, known as benign overfitting, seems to confrontthe classical bias-variance trade-off in statistical theory. Such a mismatch between observationsand classical theory sparked avid research attempting to understand how benign overfitting occurs.Theoretical work studying benign overfitting was initiated in [10], where a linear regression settingwith Gaussian data and noise was considered. It presented conditions for minimum-norm interpolators to generalize well. In a non-linear setting induced by the ReLU activation function, benignoverfitting is previously verified for fully-connected networks in [11]. On top of that, a recentwork [12] shows that training shallow neural networks with shared weights by gradient descentcan achieve an arbitrarily small training error.Our work studies the learning rate (generalization performance) of CNNs under underparameterized and overparameterized settings, respectively. We use the term learning rates to refer tothe rates of convergence of the output function to the regression function. We study the benignoverfitting problem in a regression setting. Our goal is to show that, for any input data, there always exist some CNNs that generalize well (achieve a sufficient fast learning rate) and, at the sametime, perfectly fit the data. In other words, we intend to verify the possible occurrence of benignoverfitting among deep CNNs. This work is published in Applied and Computational HarmonicAnalysis, 2024.Chapter 2: Classification of Unbounded Data by Gaussian Mixture Models via deep ReLUNetworks The second chapter of this thesis studies the binary classification of unbounded datagenerated by a mixture of Gaussian distributions using neural networks. With input X ∈ Rdand output label Y ∈ {−1, 1}, we are interested in learning a binary classifier f : Rd → {−1, 1}.We assume our data in R dfollows a class of distribution largely used to model real-world data,namely the Gaussian Mixture Model (GMM). According to [13, Chapter 3], a GMM is a universalapproximator of densities. On top of that, many studies have shown that GMM is an effectivemodel for audio, speech, image, and text processing, e.g., see [14, 15, 16]. The universality ofGMM motivates us to study the classification problem under such distributional assumptions ondata.
■590 ▼aSchool code: 0078.
■650 4▼aCybersecurity
■650 4▼aDeep learning
■650 4▼aVoice recognition
■650 4▼aNeural networks
■650 4▼aComputer science
■690 ▼a0800
■690 ▼a0984
■71020▼aGeorgia Institute of Technology.
■7730 ▼tDissertations Abstracts International▼g87-05B.
■790 ▼a0078
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360469▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


