Characterizing and Improving Electro-Optical Convolution Acceleration forAchieving Peta-Scale to Exa-Scale Performance for Deep-Network Training and Inference
Abstract
This grant will construct a micro-mirror-based electro-optical deep-learning accelerator (EODLA) that can greatly accelerate forward and backward inference for deep learning. We also want to characterize the EODLA#s behavior to iterate on designs and post-processing in an inexpensive, quick manner.Task I: Micro-Mirror-based EODLA: We will fabricate a DMD version of our EODLA and compare its performance against our micro-display EODLA prototypes. This is predominantly an engineering problem. We have already addressed how to perform convolution with tensors on attenuating surfaces for coherent and incoherent light in our existing ONR grant. We have alsoaddressed how to perform both multi-layer inference and training.These processes carry over, with few modifications, from the micro-display case to the micro-mirror-array case. On the research side, we will consider novel approaches for knowledge distillation to create compact networks and push much of the task-relevant computation to the passive stage of the EODLA. We will also explore the design of novel deep network architectures that can exploit unique properties of micro-mirror-based convolution.Task II: Kernel-basedEODLA Physics Modeling: We will unify our work in kernel-based functional filtering and kernel-based epistemic uncertainty quantification into a framework to emulate physical phenomena. Instead of applying the methodology to solve machine learning problems, we plan to apply the approach to approximately, but sufficiently, characterize the physics of an EODLA and create a simulation environment to facilitate several tasks. One example is cross-device calibration. Such a problem needs to be addressed to ensure that models ported from one EODLA to another will behave similarly. Another example is convolution contrast enhancement. In both cases, we anticipate that the same kernel methods used in the creation of the physics simulator will facilitate the development of one-step, non-iterative processes to align EODLA results with their expected ones from simulation.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 29, 2023
- Source ID
- N000142312571
Entities
People
- José PrÃncipe
Organizations
- Office of Naval Research
- United States Navy
- University of Florida