Description:
- Optical computing
- Neural networks
Abstract
USC inventors propose PT-symmetric ONN, a novel architecture based on parity-time (PT) symmetric couplers. The architecture uses optical gain-loss in III-V semiconductors or other gain materials, providing a performance comparable to passive optical systems with phase shifters even at low/moderate levels of gain-loss. PT-symmetric ONN uses a cascading structure to ensure a large number of free parameters, making the network sufficiently expressive to distinguish patterns. This approach has the potential to significantly reduce energy consumption, increase training speed, and lower the footprint in on-chip ONNs.
Benefits
- Addresses problems of current ONNs
- Performs comparably to passive optical systems with phase shifters
- Decreased energy consumption and on-chip footprint
- Faster training speed
Market Application
The computing power of modern electronics is inherently limited by their electronic Von Neumann architecture. Neuromorphic approaches and optical platforms have emerged as a solution to these bottlenecks; past work shows that optical neural networks (ONNs) utilizing an array of cascaded Mach-Zehnder interferometers (MZIs) can offer higher energy efficiency and computational speed than their electronic counterparts. However, current ONNs suffer from limitations in phase shifters.
Publications
Deng, Haoqin, and Mercedeh Khajavikhan. "Parity–time symmetric optical neural networks." Optica 8.10 (2021): 1328-1333. https://doi.org/10.1364/OPTICA.435525
Stage of Development
- Proof of concept demonstrated
- Simulation tested
- Available for licensing