Optical ReLU using membrane lasers for an all-optical neural network.

Area of Science:

• Photonics
• Optical Computing
• Semiconductor Devices

Background:

• All-optical neural networks (all-ONNs) offer potential for high-speed computation but face challenges in implementing efficient nonlinear activation functions.
• Traditional electronic activation functions consume significant power, limiting the scalability of neural network hardware.

Purpose of the Study:

• To propose and demonstrate low-power, programmable on-chip optical nonlinear units (ONUs) for all-optical neural networks.
• To utilize the inherent nonlinearity of semiconductor lasers as a rectified linear unit (ReLU) activation function.
• To assess the power efficiency and compatibility of these optical units with silicon photonics.

Main Methods:

• Constructed ONUs using a III-V semiconductor membrane laser.
• Leveraged the laser's input-output power relationship to emulate the ReLU activation function.
• Measured the optical response to validate the ReLU functionality.

Main Results:

• Achieved low-power operation for the optical nonlinear units.
• Successfully demonstrated the rectified linear unit (ReLU) activation function using the semiconductor laser.
• Confirmed the device's response characteristics matched the desired activation function.

Conclusions:

• The proposed low-power ONUs are a promising solution for implementing ReLU functions in optical circuits.
• The device's high compatibility with silicon photonics facilitates integration into future optical computing systems.
• This work advances the development of energy-efficient all-optical neural networks.