Small-scale photonic Kolmogorov-Arnold networks using standard telecom nonlinear modules

Area of Science:

• Photonics
• Artificial Intelligence
• Optical Computing

Background:

• Photonic neural networks (PNNs) offer high-speed inference but often rely on electronic nonlinearities, creating optical-electrical-optical conversion bottlenecks.
• Existing PNN architectures struggle to integrate efficient nonlinear processing directly within the optical domain.

Purpose of the Study:

• To develop and demonstrate a novel photonic neural network architecture based on Kolmogorov-Arnold networks (KANs) that integrates nonlinear processing directly into optical modules.
• To enable end-to-end differentiable training of these photonic KANs using standard telecom components.

Main Methods:

• Implemented KANs where each network edge is a trainable module comprising a Mach-Zehnder interferometer, semiconductor optical amplifier, and variable optical attenuators.
• Developed a fully differentiable physics-based model for end-to-end optimization of the optical modules.
• Evaluated network performance on nonlinear classification and regression tasks with varying input precision and signal-to-noise ratios.

Main Results:

• A four-module photonic KAN achieved 94.3% accuracy on nonlinear classification tasks.
• A seven-module network demonstrated R² = 0.986 for a six-input regression task.
• The networks exhibited robustness to 6-bit inputs and a 14 dB signal-to-noise ratio.

Conclusions:

• Photonic KANs offer a practical pathway for realizing high-performance optical neural networks by integrating trainable nonlinearities.
• End-to-end differentiable training of standard telecom modules facilitates the transition from simulation to experimental photonic computing.
• This architecture minimizes optical-electrical-optical bottlenecks, paving the way for light-speed AI inference.