서브메뉴
검색
Understanding Dynamics in Intelligent Systems: Control and Learning in Neural Circuits and Artificial Neural Networks
Understanding Dynamics in Intelligent Systems: Control and Learning in Neural Circuits and Artificial Neural Networks
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105610
- ISBN
- 9798265427045
- DDC
- 574
- 저자명
- Chen, Feng.
- 서명/저자
- Understanding Dynamics in Intelligent Systems: Control and Learning in Neural Circuits and Artificial Neural Networks
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 263 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: B.
- 주기사항
- Advisor: Druckmann, Shaul;Ganguli, Surya.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약Intelligent systems, both neural circuits and artificial neural networks, rely on dynamical processes to enable computation, flexible adaptation and efficient learning. This dissertation explores these dynamical phenomena, revealing mechanisms that may subserve intelligence. In the first part, I will discuss neural dynamics recorded from biological experiments. A key subject of modern neuroscience is the adaptive control of behavior. Here, by analyzing and modeling neural dynamics in the orofacial motor cortex in Alston's singing mouse, we reveal a hierarchical mechanism of vocal production enabled by temporal scaling of the motor cortical dynamics. In complementary work on visual processing in Drosophila, we identify a recurrent neural circuit that dynamically sharpens temporal contrasts in visual inputs. Our modeling describes how recurrent dynamics can compensate for motion-induced blur, suggesting dynamic adaptation as an effective neural strategy for sensory enhancement. The second part examines iterative pruning in artificial neural networks through the lens of dynamics in the error landscape. We show that successive pruned networks remain connected by a linear path without error barriers, and that the subsequent retraining process can be interpreted as a form of thermalization that reshapes the weight distribution. The third part focuses on learning dynamics in artificial neural networks. We show that stochastic noise during training implicitly regularizes the learning trajectory towards simpler subnetworks, effectively reducing the model complexity and enhancing generalization. Additionally, we analyze how training dynamics impact performance in large language models. We find that conventional fine-tuning objective induces overconfidence, thereby limiting performance improvements when scaling test-time compute. We propose an alternative training method that aligns the training objective with the test-time strategy, achieving improved reasoning abilities. In summary, these studies highlight approaches to understand different critical roles of dynamics in intelligent systems and demonstrate how explaining the underlying mechanisms of dynamical processes can help advance our understanding of intelligence.
- 일반주제명
- Adaptation
- 일반주제명
- Insects
- 일반주제명
- Secondary schools
- 일반주제명
- Intelligent systems
- 일반주제명
- Teachers
- 일반주제명
- Neural networks
- 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■001000017360721
■00520260202105610
■006m o d
■007cr#unu||||||||
■020 ▼a9798265427045
■035 ▼a(MiAaPQ)AAI32316391
■035 ▼a(MiAaPQ)Stanfordff551mr4474
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a574
■1001 ▼aChen, Feng.
■24510▼aUnderstanding Dynamics in Intelligent Systems: Control and Learning in Neural Circuits and Artificial Neural Networks
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a263 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: B.
■500 ▼aAdvisor: Druckmann, Shaul;Ganguli, Surya.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aIntelligent systems, both neural circuits and artificial neural networks, rely on dynamical processes to enable computation, flexible adaptation and efficient learning. This dissertation explores these dynamical phenomena, revealing mechanisms that may subserve intelligence. In the first part, I will discuss neural dynamics recorded from biological experiments. A key subject of modern neuroscience is the adaptive control of behavior. Here, by analyzing and modeling neural dynamics in the orofacial motor cortex in Alston's singing mouse, we reveal a hierarchical mechanism of vocal production enabled by temporal scaling of the motor cortical dynamics. In complementary work on visual processing in Drosophila, we identify a recurrent neural circuit that dynamically sharpens temporal contrasts in visual inputs. Our modeling describes how recurrent dynamics can compensate for motion-induced blur, suggesting dynamic adaptation as an effective neural strategy for sensory enhancement. The second part examines iterative pruning in artificial neural networks through the lens of dynamics in the error landscape. We show that successive pruned networks remain connected by a linear path without error barriers, and that the subsequent retraining process can be interpreted as a form of thermalization that reshapes the weight distribution. The third part focuses on learning dynamics in artificial neural networks. We show that stochastic noise during training implicitly regularizes the learning trajectory towards simpler subnetworks, effectively reducing the model complexity and enhancing generalization. Additionally, we analyze how training dynamics impact performance in large language models. We find that conventional fine-tuning objective induces overconfidence, thereby limiting performance improvements when scaling test-time compute. We propose an alternative training method that aligns the training objective with the test-time strategy, achieving improved reasoning abilities. In summary, these studies highlight approaches to understand different critical roles of dynamics in intelligent systems and demonstrate how explaining the underlying mechanisms of dynamical processes can help advance our understanding of intelligence.
■590 ▼aSchool code: 0212.
■650 4▼aAdaptation
■650 4▼aInsects
■650 4▼aSecondary schools
■650 4▼aIntelligent systems
■650 4▼aTeachers
■650 4▼aNeural networks
■690 ▼a0800
■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=T17360721▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


