Event

ANNOUNCEMENT: Ph.D. Dissertation Defense 

Name: Yingjie Li 

Committee: 

Prof. Cunxi Yu (Chair) 

Prof. Bahar Asgari 

Prof. Thomas E. Murphy 

Prof. Weilu Gao 

Prof. Dinesh Manocha (Dean’s Representative) 

Date/time: Tuesday, May 13, 2025 at 12pm – 2pm 

Location: AVW 1146 

Remote access: 

https://umd.zoom.us/j/6842341157?pwd=V27OFpe10cpKnyAsH82XYSxFTavMLM.1&omn=9753995 8886&jst=2 

Title: Bridging Light with Deep Learning: Algorithm, Compiler, and Applications 

Abstract: Deep neural networks (DNNs) have shown great potential in dealing with various intelligent tasks. However, conventional neural networks deployed on digital platforms have intrinsic limitations in their throughput, computation speed, and energy consumption, especially for the resource-constrained applications. A more scalable, faster, and energy-efficient method is needed for machine learning development: optical neural networks (ONNs), where information is encoded on the light signal. Among ONNs, free-space Diffractive ONNs (DONNs) process information-encoded light signal using trained passive devices, achieving high system throughput, light-speed computation, and energy efficiency due to no added energy costs from light propagation, diffraction, and phase modulation with passive optical devices. 

However, there are several critical challenges that limit the development and exploration of DONNs – (a) Existing physics engines for DONN emulation and training have high computational demands, and there’s no high performance computing (HPC) optimized language tailored for ONNs ensuring flexibility and maintainability. (b) There’s a lack of hardware-software co-design algorithms for practical DONNs from design to fabrication. (c) The absence of robust emulation frameworks limits the practical applications of DONNs, as the design and explorations of DONN systems require sufficient multi-disciplinary domain knowledge, posing critical technical barriers. Finally, (d) the accessibility of reproducing and advancing DONN research to broader communities is constrained, which posts great needs in developing an open-sourced design infrastructure. 

Targeting the improvements to the development of DONNs, my dissertation will include (1) Physics-aware differentiable co-design algorithm designed specifically for DONN systems, enabling the efficient and accurate system training and design automation; (2) Physics-aware optical adversary investigations, uncovering the unique optical security vulnerabilities in optical neural networks and offering insights into adversarial strategies applicable to other complex-domain systems; (3) An end-to-end agile design framework LightRidge designed specifically for DONN systems. It integrates with the proposed co-design algorithms and accurate and high-performance optimized computing kernels together with user-friendly domain-specific-language support, providing seamless design-to-deployment workflow for DONN systems and bridging the expertise gap for cross-disciplinary research; (4) various machine learning applications and explorations with DONNs including physics-aware multi-task learning, all-optical graph learning, and all-optical autonomous driving with DONN systems, demonstrating the capability of DONN systems in real-world applications and enriching the research for DONNs.