서브메뉴
검색
Adaptation and Regularization of Deep Neural Networks Under Temporal Smoothness Assumption
Adaptation and Regularization of Deep Neural Networks Under Temporal Smoothness Assumption
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103135
- ISBN
- 9798311957151
- DDC
- 511.5
- 서명/저자
- Adaptation and Regularization of Deep Neural Networks Under Temporal Smoothness Assumption
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 86 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
- 주기사항
- Advisor: Wall, Dennis.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약Deep neural networks have demonstrated exceptional performance across various machine learning tasks over the past decade, yet their deployment in resource-constrained and dynamically shifting real-world environments remains a challenge. While largescale models excel in accuracy, their computational demands often render them impractical for edge devices and real-time applications. In contrast, lightweight models, although efficient, frequently suffer from decreased robustness and generalization, particularly in dynamic settings that involve domain shifts. This trade-off is especially problematic in mobile healthcare applications, where privacy, efficiency, and reliability are critical constraints.This dissertation introduces TempT (Temporal Consistency for Test-Time Adaptation), a novel test-time adaptation (TTA) approach designed to enhance the robustness of deep neural networks without requiring labeled data. TempT leverages temporal coherence as a self-supervised learning signal, enforcing smoothness constraints on model predictions across sequential inputs. By minimizing high-frequency fluctuations, the method not only improves prediction stability but also enhances model performance and robustness in unseen environments. This adaptation technique is particularly effective in video-based learning tasks, such as facial expression recognition or video object detection, where maintaining consistency across frames is crucial.Furthermore, we explore the application of topological data analysis (TDA), particularly Persistent Homology, as a mechanism for quantifying model behavior during adaptation. By analyzing the topological features of intermediate network activations, we develop a selective adaptation strategy, enabling the model to determine when adaptation is beneficial and when it might degrade performance. Additionally, we propose a novel regularization technique based on temporal consistency, which mproves model generalization, and robustness against domain shifts.We evaluate our frameworks on real-world datasets, including AffWild2, SHIFT, and CIFAR100P. Experimental results demonstrate that TempT not only outperforms existing test-time adaptation methods, but also enables computationally efficient models to achieve performance levels comparable to larger, more expensive architectures. This work bridges the gap between domain adaptation and self-supervised learning, offering a robust, privacy-conscious, and scalable solution for deep learning in constrained environments.
- 일반주제명
- Graphs
- 일반주제명
- Video recordings
- 일반주제명
- Neural networks
- 일반주제명
- Adaptation
- 일반주제명
- Fault tolerance
- 일반주제명
- Optimization algorithms
- 일반주제명
- Entropy
- 일반주제명
- Computer science
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-12B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357123
■00520260202103135
■006m o d
■007cr#unu||||||||
■020 ▼a9798311957151
■035 ▼a(MiAaPQ)AAI31974587
■035 ▼a(MiAaPQ)Stanfordbp098kt2063
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a511.5
■1001 ▼aMutlu, Onur Cezmi.
■24510▼aAdaptation and Regularization of Deep Neural Networks Under Temporal Smoothness Assumption
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a86 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-12, Section: B.
■500 ▼aAdvisor: Wall, Dennis.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aDeep neural networks have demonstrated exceptional performance across various machine learning tasks over the past decade, yet their deployment in resource-constrained and dynamically shifting real-world environments remains a challenge. While largescale models excel in accuracy, their computational demands often render them impractical for edge devices and real-time applications. In contrast, lightweight models, although efficient, frequently suffer from decreased robustness and generalization, particularly in dynamic settings that involve domain shifts. This trade-off is especially problematic in mobile healthcare applications, where privacy, efficiency, and reliability are critical constraints.This dissertation introduces TempT (Temporal Consistency for Test-Time Adaptation), a novel test-time adaptation (TTA) approach designed to enhance the robustness of deep neural networks without requiring labeled data. TempT leverages temporal coherence as a self-supervised learning signal, enforcing smoothness constraints on model predictions across sequential inputs. By minimizing high-frequency fluctuations, the method not only improves prediction stability but also enhances model performance and robustness in unseen environments. This adaptation technique is particularly effective in video-based learning tasks, such as facial expression recognition or video object detection, where maintaining consistency across frames is crucial.Furthermore, we explore the application of topological data analysis (TDA), particularly Persistent Homology, as a mechanism for quantifying model behavior during adaptation. By analyzing the topological features of intermediate network activations, we develop a selective adaptation strategy, enabling the model to determine when adaptation is beneficial and when it might degrade performance. Additionally, we propose a novel regularization technique based on temporal consistency, which mproves model generalization, and robustness against domain shifts.We evaluate our frameworks on real-world datasets, including AffWild2, SHIFT, and CIFAR100P. Experimental results demonstrate that TempT not only outperforms existing test-time adaptation methods, but also enables computationally efficient models to achieve performance levels comparable to larger, more expensive architectures. This work bridges the gap between domain adaptation and self-supervised learning, offering a robust, privacy-conscious, and scalable solution for deep learning in constrained environments.
■590 ▼aSchool code: 0212.
■650 4▼aGraphs
■650 4▼aVideo recordings
■650 4▼aNeural networks
■650 4▼aAdaptation
■650 4▼aFault tolerance
■650 4▼aOptimization algorithms
■650 4▼aEntropy
■650 4▼aComputer science
■690 ▼a0984
■690 ▼a0800
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-12B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357123▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


