Optical neural networks with intensity‑based projection layers as effective nonlinear activations.

Area of Science:

• Optics
• Computer Science
• Artificial Intelligence

Background:

• Optical neural networks (ONNs) promise high speed and energy efficiency by utilizing light's parallelism.
• Implementing nonlinear activation functions in ONNs remains a significant design challenge.
• Existing methods often lack physical realizability or hardware efficiency.

Purpose of the Study:

• To propose and validate projection layers as effective nonlinear activation functions for ONNs.
• To demonstrate the robustness and scalability of this new approach.
• To extend the method to quaternion-valued representations for multi-dimensional feature extraction.

Main Methods:

• Utilizing optical intensity measurements as projections from complex/quaternion-valued fields to real-valued magnitudes.
• Employing amplitude or phase encoding for signal transmission to the next layer.
• Systematic evaluation across image classification, reconstruction, and physical parameter inference tasks.

Main Results:

• Projection layers effectively serve as nonlinear activation mechanisms in ONNs.
• The approach demonstrates robustness under noisy conditions.
• Successful application in a beta-Variational Autoencoder (β-VAE) for interpretable physical parameter learning.

Conclusions:

• Projection layers offer a scalable and hardware-efficient foundation for deep ONNs.
• This method leverages optical computing advantages for improved performance.
• Broad applicability in image recognition, signal processing, and scientific inference is established.