Event
Ph.D. Research Proposal Exam: Yingjie Li
Tuesday, September 17, 2024
1:30 p.m.-3:30 p.m.
AVW 1146
Maria Hoo
301 405 3681
mch@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Yingjie Li
Committee:
Prof. Cunxi Yu (Chair)
Prof. Bahar Asgari
Prof. Thomas E. Murphy
Date/time: Tuesday, September 17, 2024 at 1:30pm – 3:30pm EST
Location: AVW 1146
Title: Bridging Light with Deep Learning: Algorithm, Compiler, and
Applications
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 overcome these challenges, a more scalable, faster,
and energy-efficient approach is required for the advancement of
machine learning: optical neural networks (ONNs), which utilize light
signals instead of electrical ones. Among ONNs, Free-space
Diffractive Optical Neural Networks (DONNs) offer high throughput,
light-speed computation, and energy efficiency by processing
information-encoded light using trained passive devices, without
incurring additional energy costs due to light propagation, diffraction,
and phase modulation through passive optical components.
However, the development and application of DONNs are hindered by
several key challenges: (a) Current physics engines for DONN
emulation and training are computationally intensive, and there is no
High-Performance Computing (HPC)-optimized language tailored
specifically for ONNs that balances flexibility and maintainability. (b)
There is a lack of hardware-software co-design algorithms that can
facilitate the practical implementation of DONNs from design to
fabrication. (c) The absence of a robust emulation framework limits
the practical applications of DONNs, including the ability to address
security threats and discrepancies between emulation and real-world
implementation. (d) The accessibility of DONN research is limited,
necessitating the development of an open-source design infrastructure
to enable broader community involvement.
To address these challenges and contribute to the advancement of
DONNs, my thesis will focus on the following: (1) the development of a
physics-aware differentiable co-design algorithm specifically tailored
for DONN systems; (2) the creation of an end-to-end, physics-aware,
agile design framework, LightRidge, for DONN systems, which
integrates the proposed co-design algorithms with optimized
computing kernels and provides the full-stack support through a
user-friendly domain-specific language; (3) the exploration and
application of DONNs across various domains using the framework,
including multi-task learning with physics-aware DONN systems, the
investigation of optical AI adversaries, and hardware implementation,
among others.
