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

Abstract

Abstract Photonic neural networks promise inference at the speed of light, yet most architectures combine linear optical meshes with electronic nonlinearities, reintroducing optical-electrical-optical bottlenecks. Kolmogorov-Arnold networks place trainable nonlinear functions on network edges, concentrating expressivity into a few structured modules. Each edge here is a single module built from a Mach-Zehnder interferometer, a semiconductor optical amplifier, and variable optical attenuators, giving a four-parameter transfer function set by gain saturation and interferometric mixing. A four-module network attains 94.3% accuracy (±3.9% s.d. over ten seeds) on nonlinear classification, and a seven-module network reaches R 2  = 0.986 ± 0.015 on six-input regression, remaining robust to 6-bit inputs and 14 dB signal-to-noise ratio. Here, we show that a fully differentiable physics model enables end-to-end optimization of these standard telecom modules, giving a practical route from simulation toward experimental demonstration of photonic Kolmogorov-Arnold networks.