All-optical doubly resonant cavities for energy-efficient ReLU function in nanophotonic deep learning.

Area of Science:

• Photonics
• Nonlinear Optics
• Artificial Intelligence Hardware

Background:

• Implementing efficient activation functions is crucial for optical neural networks.
• Existing optical activation functions often suffer from large footprints and high energy consumption.

Purpose of the Study:

• To present a novel, compact all-optical Rectified Linear Unit (ReLU) activation function.
• To demonstrate significant reductions in device size and energy usage compared to previous methods.
• To explore the adaptability of the proposed structure for other activation functions like ELU and GELU.

Main Methods:

• Utilizing compact doubly-resonant cavities (approx. 10μm).
• Leveraging nonlinear processes and phase-sensitive second-harmonic generation.
• Employing coupled-mode theory for theoretical framework and finite-difference time-domain simulations for validation.

Main Results:

• Achieved femtojoule-level activation energy, reducing device footprint by two orders of magnitude.
• Demonstrated an optical analog to the ReLU function with high accuracy.
• Showcased the ability to implement ELU and GELU functions with minor input adjustments.

Conclusions:

• The proposed compact optical activation functions offer a significant advancement for energy-efficient, high-density optical neural networks.
• This technology paves the way for next-generation artificial intelligence hardware.
• The demonstrated versatility in implementing multiple activation functions enhances its potential.