Experimental realization of arbitrary activation functions for optical neural networks

Area of Science:

• Photonics
• Optical Computing
• Artificial Intelligence Hardware

Background:

• Optical neural networks (ONNs) offer potential advantages in speed and energy efficiency over electronic counterparts.
• Implementing complex nonlinear activation functions, crucial for deep learning, remains a significant challenge in photonic systems.
• Existing optical nonlinearities are often limited in shape and introduce substantial signal loss.

Purpose of the Study:

• To experimentally demonstrate a novel on-chip electro-optic circuit for realizing arbitrary nonlinear activation functions in ONNs.
• To overcome the limitations of conventional optical nonlinearities and enable more sophisticated activation functions.
• To investigate the impact of this new activation function on the performance of ONNs for image classification tasks.

Main Methods:

• An on-chip electro-optic circuit was designed and fabricated.
• The circuit converts a portion of the input optical signal to an electrical signal for processing.
• This electrical signal modulates the intensity of the remaining optical signal to create the desired nonlinear activation function.
• Numerical simulations were performed to evaluate the circuit's performance in an ONN for MNIST image classification.

Main Results:

• The electro-optic circuit successfully implemented arbitrary optical-to-optical nonlinearities without requiring optical amplification.
• Simulations showed improved performance of the ONN utilizing this activation function on the MNIST dataset.
• The proposed method significantly reduces optical signal attenuation compared to traditional nonlinearities.

Conclusions:

• The developed on-chip electro-optic circuit provides a versatile platform for implementing diverse nonlinear activation functions in ONNs.
• This approach facilitates the creation of deeper and more powerful optical neural networks.
• The reduced signal loss opens new avenues for energy-efficient and high-performance optical computing.