서브메뉴
검색
Exploring Hybrid Algorithms and Optimization Strategies in Non-Conventional Computing Architectures
Exploring Hybrid Algorithms and Optimization Strategies in Non-Conventional Computing Architectures
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152116
- ISBN
- 9798384340713
- DDC
- 004
- 서명/저자
- Exploring Hybrid Algorithms and Optimization Strategies in Non-Conventional Computing Architectures
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 158 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
- 주기사항
- Advisor: Pavone, Marco.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약The relentless increase in computing workloads coupled with the potential slowdown of Moore's law has spurred the development of both non-conventional computing architectures and algorithms that make use of these novel architectures. This research is primarily conducted by disjoint communities that rely heavily on algorithmic "primitives" to abstract away the intricacies of the other's work, and to serve as a buffer between them to enable concurrent development. As additional research builds on existing primitives, it creates a feedback cycle where those primitives become increasingly entrenched, creating further justification for continuing to work with them. This is particularly problematic because it is unclear whether the small number of primitives that dominate the literature today were deliberately designed with real-world applications in mind. More broadly, we advocate for taking a holistic view of hardware, algorithm and abstraction interactions, with the goal of promoting primitives that appropriately balance ease of hardware implementation with impact of downstream applications.In this thesis, we primarily focus on the interaction between algorithms and the Ising abstraction, with the objective of deriving complexity metrics that quantify their "fit". To do so, we will begin by introducing copositive and completely-positive optimization as a bridge between convex and non-convex optimization. We find that this is a helpful lens for complexity analysis of hybrid algorithms because it allows for convex optimization algorithms to be applied to some of the non-convex problems that we hope to address with non-conventional computing.We will examine neural network verification as an illustrative example to develop geometric intuition for how copositive/completely positive optimization can exactly represent mixed-binary quadratic programs, while closely related semi-definite programs result in relaxation gaps. While the completely positive representation of neural network verification is intractable, it is useful in its own right due to its similarity to semi-definite programming (SDP). This perspective offers insights for how to construct tighter SDP-based relaxations for verification and unifies existing SDP-based verification methods.With this intuition in place, we will then present our hybrid Ising-classical cutting-plane algorithm for copositive optimization. We will examine the algorithm's convergence guarantees, which it notably inherits directly from existing analysis of convex optimization algorithms. Concluding this section, we leverage our proposed methodology to justify revising the Ising abstraction.Finally, we turn our attention to the Coherent Ising Machine as an example of hardware whose native physics is more closely aligned with the modified Ising abstraction -- this variant of the CIM is referred to as the Continuous-Variable Coherent Ising Machine (CV-CIM). We will show how existing optimization techniques, such as momentum and Adam, can be readily incorporated into the CV-CIM. These modifications allow the device to inherit many of the favorable properties of these algorithms, including accelerated convergence, superior navigation of non-convex landscapes, and robustness to parameter tuning.
- 일반주제명
- Quantum computing
- 일반주제명
- Sample size
- 일반주제명
- Integer programming
- 일반주제명
- Computers
- 일반주제명
- Graphs
- 일반주제명
- Neural networks
- 일반주제명
- Plot (Narrative)
- 일반주제명
- Convex analysis
- 일반주제명
- Design
- 일반주제명
- Complexity theory
- 일반주제명
- Optimization algorithms
- 일반주제명
- Geometry
- 일반주제명
- Computer science
- 일반주제명
- Mathematics
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017162955
■00520250211152116
■006m o d
■007cr#unu||||||||
■020 ▼a9798384340713
■035 ▼a(MiAaPQ)AAI31460310
■035 ▼a(MiAaPQ)Stanfordkx590cv4574
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aBrown, Robin Alexandra.
■24510▼aExploring Hybrid Algorithms and Optimization Strategies in Non-Conventional Computing Architectures
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a158 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: B.
■500 ▼aAdvisor: Pavone, Marco.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aThe relentless increase in computing workloads coupled with the potential slowdown of Moore's law has spurred the development of both non-conventional computing architectures and algorithms that make use of these novel architectures. This research is primarily conducted by disjoint communities that rely heavily on algorithmic "primitives" to abstract away the intricacies of the other's work, and to serve as a buffer between them to enable concurrent development. As additional research builds on existing primitives, it creates a feedback cycle where those primitives become increasingly entrenched, creating further justification for continuing to work with them. This is particularly problematic because it is unclear whether the small number of primitives that dominate the literature today were deliberately designed with real-world applications in mind. More broadly, we advocate for taking a holistic view of hardware, algorithm and abstraction interactions, with the goal of promoting primitives that appropriately balance ease of hardware implementation with impact of downstream applications.In this thesis, we primarily focus on the interaction between algorithms and the Ising abstraction, with the objective of deriving complexity metrics that quantify their "fit". To do so, we will begin by introducing copositive and completely-positive optimization as a bridge between convex and non-convex optimization. We find that this is a helpful lens for complexity analysis of hybrid algorithms because it allows for convex optimization algorithms to be applied to some of the non-convex problems that we hope to address with non-conventional computing.We will examine neural network verification as an illustrative example to develop geometric intuition for how copositive/completely positive optimization can exactly represent mixed-binary quadratic programs, while closely related semi-definite programs result in relaxation gaps. While the completely positive representation of neural network verification is intractable, it is useful in its own right due to its similarity to semi-definite programming (SDP). This perspective offers insights for how to construct tighter SDP-based relaxations for verification and unifies existing SDP-based verification methods.With this intuition in place, we will then present our hybrid Ising-classical cutting-plane algorithm for copositive optimization. We will examine the algorithm's convergence guarantees, which it notably inherits directly from existing analysis of convex optimization algorithms. Concluding this section, we leverage our proposed methodology to justify revising the Ising abstraction.Finally, we turn our attention to the Coherent Ising Machine as an example of hardware whose native physics is more closely aligned with the modified Ising abstraction -- this variant of the CIM is referred to as the Continuous-Variable Coherent Ising Machine (CV-CIM). We will show how existing optimization techniques, such as momentum and Adam, can be readily incorporated into the CV-CIM. These modifications allow the device to inherit many of the favorable properties of these algorithms, including accelerated convergence, superior navigation of non-convex landscapes, and robustness to parameter tuning.
■590 ▼aSchool code: 0212.
■650 4▼aQuantum computing
■650 4▼aSample size
■650 4▼aInteger programming
■650 4▼aComputers
■650 4▼aGraphs
■650 4▼aNeural networks
■650 4▼aPlot (Narrative)
■650 4▼aConvex analysis
■650 4▼aDesign
■650 4▼aComplexity theory
■650 4▼aOptimization algorithms
■650 4▼aGeometry
■650 4▼aComputer science
■650 4▼aMathematics
■690 ▼a0389
■690 ▼a0800
■690 ▼a0984
■690 ▼a0405
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-03B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17162955▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


