An alternative computing technique for machine learning

The artificial neural network (ANN), an information processing system designed after the human brain, has enabled machine learning to become the focus of recent technological transformations. However, ANNs require high computational power, which electronic integrated circuits can no longer satisfy.

Optical neural networks (ONNs) overcome this challenge. Instead of using electrons as the primary information carrier, ONNs use photons, which improve the architecture’s speed, power, and scalability. Inspired by its potential, Montes McNeil et al. compared the two main classes of ONNs and traditional machine learning models.

“Our review explores optical computing, which we believe offers significant benefits over today’s computing technologies,” said author Alex Montes McNeil. “Machine learning, specifically neural networks, provides an excellent platform for demonstrating the benefits of optical computing architectures.”

The authors are especially interested in free-space optical neural networks (FSONNs). During its operation, information is encoded onto the FSONN light source as the input to the neural network, preparing it for passive computation through the hidden layers. This allows for high speed and energy efficiency. Furthermore, different FSONN architectures, such as 3D printed layers, metasurfaces, and spatial light modulators, show promise in various specialized applications.

In the future, the authors are excited to design new free-space optical components and extend the FSONN platform beyond machine learning applications.

“Researchers have already demonstrated some great alternative applications, such as quantitative phase imaging, encryption, and seeing through random diffusers,” said Montes McNeil. “We anticipate that researchers will continue to push this platform far beyond these initial use cases and discover all sorts of novel free-space optical computing systems for the future.”

Source: “Fundamentals and recent developments of free-space optical neural networks,” by Alexander Montes McNeil, Yuxiao Li, Allen Zhang, Michael Moebius, and Yongmin Liu, Journal of Applied Physics (2024). The article can be accessed at https://doi.org/10.1063/5.0215752 .