Related Experiment Video
Updated: Dec 10, 2025
10:18
Fiber-optic Implantation for Chronic Optogenetic Stimulation of Brain Tissue
Published on: October 29, 2012
45.6K
Reconfigurable all-optical nonlinear activation functions for neuromorphic photonics.
Optics Letters
|September 2, 2020
Summary
Researchers developed a reconfigurable photonic device for all-optical nonlinear activation functions, enabling programmable functions for photonic neural networks and achieving high accuracy in benchmark tasks.
Area of Science:
- Photonics
- Neuromorphic Computing
- Silicon Photonics
Background:
- Photonic neural networks require efficient nonlinear activation functions.
- Implementing these functions all-optically offers speed and energy advantages.
- Existing methods often lack reconfigurability.
Purpose of the Study:
- To demonstrate all-optical reconfigurable nonlinear activation functions.
- To integrate these functions into a silicon photonics platform.
- To showcase their utility in photonic neural network applications.
Main Methods:
- Utilized a cavity-loaded Mach-Zehnder interferometer.
- Employed the free-carrier dispersion effect for nonlinearity.
- Programmed various activation functions like sigmoid and ReLU.
Main Results:
- Achieved programmable generation of multiple nonlinear activation functions.
- Demonstrated 100% accuracy on XOR classification.
- Attained 94% accuracy on MNIST handwritten digit classification.
Conclusions:
- The developed device enables all-optical, reconfigurable nonlinear activation functions.
- This technology is suitable for nonlinear units in photonic neural networks.
- Flexible programming of the nonlinear transfer function can optimize neuromorphic tasks.
