All-Optical and Electro-Optic Nonlinearity for Photonic Artificial Neural Networks Towards Instantaneous Decision Making
Abstract
Machine-learning tasks performed by neural networks (NN) have demonstrated useful capabilities for producing reliable, repeatable intelligent decisions and results over large data. When trained, a NN can perform prediction tasks on unseen data with high throughput on optimized architectures able to parallelly and efficiently perform neuron functionalities, namely 1) dot-product multiplication, ‘weights’; 2) summation, addition; and 3) nonlinear activation function, or ‘thresholding’. However, electronic-based NN are limited with respect to both energy efficiency and signal delay when performing machine-learning tasks due to the dissipative (dis)charging wires and access to memory. In contrast, this project pursues optical NNs that enable short-delays towards near real-time processing capabilities when integrated photonic circuits are used. Once the ‘weights’ are SET, the photonic circuit functionals as an node-distributed, noniterative, i.e. O(1), non-Van Neumann analog processor. The project investigates fundamental questions, such as i) addressing the non-existence of an optical memory via heterogeneous integration of phase-change-materials into photonic circuits delivering non-volatility, iii) the impact of physical noise on NN training, inference accuracy, and NN cascadability, and iii) energy and speed-density scaling vectors for multiply-accumulate (MAC)-operations of photonic NNs, and combines these with device and photonic integrated-circuit prototyping and test. Towards practical scale-up and application potentials, the project will deliver insights into various basic questions and challenges. Towards transition pathways, the effort will coordinate with the DOD-R&D ecosystem. DOD-Impact: rapid and accurate decision-making processes are key in battlespace awareness which can be enhanced by real-time co-processors and relieve SWAP constrains of mission-critical systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2021
- Source ID
- FA95502010193
Entities
People
- Volker Sorger
Organizations
- Air Force Office of Scientific Research
- George Washington University
- United States Air Force