Using Metasurfaces for Multi-Task Learning in Optical Neural Networks
Author Information
Author(s): Behroozinia Sahar, Gu Qing
Primary Institution: North Carolina State University
Hypothesis
Can polarization and wavelength multiplexing in diffractive neural networks enable simultaneous classification of multiple tasks?
Conclusion
The study demonstrates that dual-channel and tri-channel diffractive neural networks can achieve high classification accuracies for multiple tasks using optical multiplexing techniques.
Supporting Evidence
- Dual-channel PM-DNN achieved 97.72% accuracy for MNIST and 88.01% for FMNIST.
- Tri-channel WM-DNN achieved 96.73% accuracy for MNIST, 80.9% for FMNIST, and 81.13% for KMNIST.
- The end-to-end joint training framework improved performance for FMNIST and KMNIST tasks.
Takeaway
This research shows how special surfaces can help computers learn to do many things at once using light, making them faster and more efficient.
Methodology
The study employed dual-channel and tri-channel diffractive neural networks using polarization and wavelength multiplexing to classify datasets.
Limitations
The performance of the tri-channel WM-DNN was lower for more complex tasks due to increased task competition.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website