All-optical nonlinear activation function based on stimulated Brillouin scattering.

Area of Science:

• Photonics
• Artificial Intelligence
• Nonlinear Optics

Background:

• Optical neural networks (ONNs) promise to overcome computational bottlenecks in digital electronics.
• A key challenge in ONN development is the lack of efficient photonic activation functions.
• Current methods often involve opto-electronic conversion, negating photonic advantages like coherent and frequency-multiplexed encoding.

Purpose of the Study:

• To experimentally demonstrate a novel photonic nonlinear activation function.
• To overcome the limitations of opto-electronic conversion in optical neural networks.
• To enable the development of high-performing, deep optical neural networks.

Main Methods:

• Utilized stimulated Brillouin scattering to create a nonlinear optical effect.
• Implemented an all-optical tuning mechanism for the activation function's shape.
• Achieved net gain to compensate for insertion loss.

Main Results:

• Demonstrated a coherent and frequency-selective photonic activation function.
• Successfully tuned the activation function to exhibit LeakyReLU, Sigmoid, and Quadratic shapes.
• Achieved a net gain as high as 20 dB, compensating for insertion loss.

Conclusions:

• The demonstrated photonic nonlinear activation function addresses a critical bottleneck in optical neural networks.
• This all-optical approach preserves the advantages of photonics for information encoding.
• The technology paves the way for the realization of deep optical neural networks.