Content area

Abstract

Deep neural networks (DNNs) have demonstrated significant potential in addressing a wide range of intelligent tasks. However, traditional neural networks deployed on digital platforms face inherent limitations in terms of throughput, computational speed, and energy efficiency, particularly in resource-constrained environments. To address these challenges, a more scalable, faster, and energy-efficient approach is needed for the advancement of deep learning. Optical neural networks (ONNs), which utilize light signals instead of electrical ones as the information carrier for computation, offer such a promising alternative. Among ONNs, free-space diffractive optical neural networks (DONNs) stand out for their high throughput, light-speed computation, and exceptional energy efficiency. They enable all-optical computing at near-light speed by manipulating information-encoded light signals through optical phenomena such as propagation, diffraction, and phase modulation. By leveraging trained passive optical elements, DONNs perform computation without additional energy consumption during all-optical inference.

However, the development and practical deployment of DONNs face several critical challenges. First, the lack of hardware-software co-design algorithms impedes the seamless realization of DONNs, from conceptual design to physical fabrication with analog optical components. Second, the absence of robust emulation frameworks limits system-level applications of DONNs, as designing and exploring DONNs require extensive cross-disciplinary expertise, posing significant technical barriers. Third, current computing engines for DONN emulation and training are computationally intensive, lacking both optimized computing kernels and domain-specific language (DSL) support tailored to ONNs that balances flexibility and maintainability. Fourth, the accessibility of DONN research is limited, necessitating the development of an open-source design infrastructure to facilitate broader community engagement and innovation.

Targeting the improvements and contributions to the development of DONNs, this dissertation presents four key contributions. First, we propose a physics-aware differentiable co-design algorithm designed specifically for DONN systems, enabling the efficient and accurate system training and design automation. Second, we conduct physics-aware optical adversarial investigations, which uncover unique optical security vulnerabilities in ONNs and provide insights into adversarial robustness applicable to other complex-domain systems. Third, we develop an open-source, end-to-end agile design framework, LightRidge, for DONN systems. This framework integrates efficient co-design algorithms, accurate yet high-performance optimized computing kernels, and user-friendly DSL support. It offers a seamless design-to-deployment workflow, bridging the expertise gap for cross-disciplinary research for DONNs. Fourth, we explore DONNs across diverse deep learning applications, including physics-aware multi-task learning, optical-inspired graph learning, and optical-inspired image processing. These applications demonstrate the capability of DONNs for real-world applications and enrich the research landscape of DONNs.

Details

1010268
Business indexing term
Title
Bridging Light With Deep Learning: Algorithm, Compiler, and Applications
Author
Number of pages
184
Publication year
2025
Degree date
2025
School code
0117
Source
DAI-B 87/3(E), Dissertation Abstracts International
ISBN
9798293834532
Advisor
Committee member
Gao, Weilu; Manocha, Dinesh; Murphy, Thomas E.; Qu, Gang
University/institution
University of Maryland, College Park
Department
Electrical Engineering
University location
United States -- Maryland
Degree
Ph.D.
Source type
Dissertation or Thesis
Language
English
Document type
Dissertation/Thesis
Dissertation/thesis number
32114699
ProQuest document ID
3250260563
Document URL
https://www.proquest.com/dissertations-theses/bridging-light-with-deep-learning-algorithm/docview/3250260563/se-2?accountid=208611
Copyright
Database copyright ProQuest LLC; ProQuest does not claim copyright in the individual underlying works.
Database
ProQuest One Academic