Adaptive sigmoid-like and PReLU activation functions for all-optical perceptron

Area of Science:

• Photonics
• Optical Computing
• Semiconductor Lasers

Background:

• Nonlinear activation functions are crucial for neural networks.
• Optical implementations of perceptrons offer potential speed advantages.
• Controlling activation function profiles in all-optical systems is challenging.

Purpose of the Study:

• To present a novel method for generating adaptive nonlinear activation functions in all-optical perceptrons.
• To demonstrate the tunability of these activation functions.
• To provide a universal fitting function for the generated optical activation functions.

Main Methods:

• Exploiting the bistability of an injection-locked Fabry-Perot semiconductor laser.
• Tailoring the activation function profile by adjusting parameters such as injection-locked side-mode order, frequency detuning, Henry factor, and bias current.

Main Results:

• Successfully generated adaptive sigmoid-like and PReLU nonlinear activation functions using an all-optical approach.
• Demonstrated precise control over the activation function's profile through parameter adjustments.
• Presented a universal fitting function applicable to both sigmoid-like and PReLU optical activation functions.

Conclusions:

• The proposed method enables the creation of flexible and adaptive nonlinear activation functions for all-optical perceptrons.
• This work contributes to the development of advanced optical computing hardware.
• The tunability of the optical activation functions opens possibilities for novel optical neural network designs.