서브메뉴
검색
Physical Neural Networks Using Acoustics and Photonics :
Physical Neural Networks Using Acoustics and Photonics :
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260311091507.5
- ISBN
- 9798315740162
- DDC
- 006
- 저자명
- Stein, Martin
- 서명/저자
- Physical Neural Networks Using Acoustics and Photonics : <html><head><meta name=ValidationSchema content=http://www.w3.org/2002/08/xhtml/xhtml1-strict.xsd/><title></title></head><body>Physikalische Neuronale Netze mit akustischen und photonischen Systemen</body></html> / Martin Stein
- 발행사항
- [Sl] : Cornell University, 2024
- 형태사항
- 1 electronic resource (168 pages)
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-11, Section: B.
- 주기사항
- Advisors: McMahon, Peter Committee members: Holmes, Natasha; Mueller, Erich.
- 학위논문주기
- - Ph.D. : Cornell University, 2024.
- 초록/해제
- 요약This thesis explores alternatives to the currently dominant approach of simulating artificial deep neural networks on digital electronic hardware. There is no serious alternative to computing with digital electronics at the moment, but no physical law singles it out as the superior approach. Strong arguments can be made for approaches using other physical processes such as optics or electro-chemistry. We introduce physical neural networks as an alternative to digital electronic simulation of neural networks. In analogy to deep neural networks using layers of fine-tuned mathematical transformations, physical neural networks harness the controllable transformations of layers of physical systems for computation. As a proof-of-concept, we trained a neural network that uses acoustic signals transformed by a vibrating metal plate mounted on an audio speaker to perform image classification. We also present the results from an optical and analog electronic system that experimentally perform audio and image classification. Physical neural networks may ultimately perform machine learning orders-of-magnitude faster and more energy-efficiently than digital electronic processors.We then turn to a photonic implementation of a physical neural network. On-chip photonic neural-network processors have potential benefits in both speed and energy efficiency but current approaches cannot reach the scale at which they can compete with digital electronic processors. Physics-aware training enabled us to pursue a different approach for on-chip photonic-neural-network processors in which the computation is performed by freely propagating waves in two dimensions. We propose and demonstrate a device whose refractive index as a function of space, n(x, z), can be rapidly reprogrammed, allowing arbitrary control over the wave propagation in the device. We used a prototype device with a functional area of 12 mm.
- 초록/해제
- 요약2 to perform neural-network inference with up to 49-dimensional input vectors in a single pass, achieving 96% accuracy on vowel classification and 86% accuracy on MNIST handwritten-digit classification, with no trained digital-electronic pre- or post-processing. This is a scale beyond that of previous photonic chips relying on discrete components, illustrating the benefit of the continuous-waves paradigm.In principle, with large enough chip area, the reprogrammability of the device's refractive index distribution enables the reconfigurable realization of any passive, linear photonic circuit or device. This promises the development of more compact and versatile photonic systems for a wide range of applications, including optical processing, smart sensing, spectroscopy, and optical communications.
- 언어주기
- English
- 일반주제명
- Physics
- 일반주제명
- Computer engineering
- 일반주제명
- Acoustics
- 키워드
- Lithium niobate
- 키워드
- Neural networks
- 기타저자
- Cornell University Physics
- 기본자료저록
- Dissertations Abstracts International. 86-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260311s2024 us eng d■001000017356532
■00520260311091507.5
■006m o d
■007cr|nu||||||||
■020 ▼a9798315740162
■040 ▼aMiAaPQD▼beng▼cMiAaPQD▼erda
■082 ▼a006
■1001 ▼aStein, Martin▼eauthor.▼0(orcid)0000-0001-6958-6925
■24510▼aPhysical Neural Networks Using Acoustics and Photonics ▼bhtmlheadmeta name='ValidationSchema' content='http://www.w3.org/2002/08/xhtml/xhtml1-strict.xsd'/title/title/headbodyPhysikalische Neuronale Netze mit akustischen und photonischen Systemen/body/html ▼cMartin Stein
■260 ▼a[Sl]▼bCornell University▼c2024
■264 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a1 electronic resource (168 pages)
■336 ▼atext▼btxt▼2rdacontent
■337 ▼acomputer▼bc▼2rdamedia
■338 ▼aonline resource▼bcr▼2rdacarrier
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-11, Section: B.
■500 ▼aAdvisors: McMahon, Peter Committee members: Holmes, Natasha; Mueller, Erich.
■5021 ▼bPh.D.▼cCornell University▼d2024.
■520 ▼aThis thesis explores alternatives to the currently dominant approach of simulating artificial deep neural networks on digital electronic hardware. There is no serious alternative to computing with digital electronics at the moment, but no physical law singles it out as the superior approach. Strong arguments can be made for approaches using other physical processes such as optics or electro-chemistry. We introduce physical neural networks as an alternative to digital electronic simulation of neural networks. In analogy to deep neural networks using layers of fine-tuned mathematical transformations, physical neural networks harness the controllable transformations of layers of physical systems for computation. As a proof-of-concept, we trained a neural network that uses acoustic signals transformed by a vibrating metal plate mounted on an audio speaker to perform image classification. We also present the results from an optical and analog electronic system that experimentally perform audio and image classification. Physical neural networks may ultimately perform machine learning orders-of-magnitude faster and more energy-efficiently than digital electronic processors.We then turn to a photonic implementation of a physical neural network. On-chip photonic neural-network processors have potential benefits in both speed and energy efficiency but current approaches cannot reach the scale at which they can compete with digital electronic processors. Physics-aware training enabled us to pursue a different approach for on-chip photonic-neural-network processors in which the computation is performed by freely propagating waves in two dimensions. We propose and demonstrate a device whose refractive index as a function of space, n(x, z), can be rapidly reprogrammed, allowing arbitrary control over the wave propagation in the device. We used a prototype device with a functional area of 12 mm.
■520 ▼a2 to perform neural-network inference with up to 49-dimensional input vectors in a single pass, achieving 96% accuracy on vowel classification and 86% accuracy on MNIST handwritten-digit classification, with no trained digital-electronic pre- or post-processing. This is a scale beyond that of previous photonic chips relying on discrete components, illustrating the benefit of the continuous-waves paradigm.In principle, with large enough chip area, the reprogrammability of the device's refractive index distribution enables the reconfigurable realization of any passive, linear photonic circuit or device. This promises the development of more compact and versatile photonic systems for a wide range of applications, including optical processing, smart sensing, spectroscopy, and optical communications.
■546 ▼aEnglish
■590 ▼aSchool code: 0058
■650 4▼aPhysics
■650 4▼aComputer engineering
■650 4▼aAcoustics
■653 ▼aEducation research
■653 ▼aIntegrated photonics
■653 ▼aIntroductory physics labs
■653 ▼aLithium niobate
■653 ▼aNeural networks
■7102 ▼aCornell University▼bPhysics.▼edegree granting institution.
■7201 ▼aMcMahon, Peter▼edegree supervisor.
■7730 ▼tDissertations Abstracts International▼g86-11B.
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17356532▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
