Small-scale photonic Kolmogorov-Arnold networks using standard telecom nonlinear modules
Researchers built tiny photonic neural networks from off-the-shelf telecom parts that do nonlinear computation entirely in light — no electronic detours required.
The Thesis
Most photonic neural networks cheat: they do the linear math in light but convert back to electrons for every nonlinear operation (such as a sigmoid or ReLU activation), creating costly signal bottlenecks. This paper proposes a way to keep the whole computation optical, using a new architecture called a photonic Kolmogorov-Arnold network, or KAN — a neural network design where the learnable functions live on the connections between nodes rather than at the nodes themselves. The components are all standard telecom gear: Mach-Zehnder interferometers (devices that split and recombine light beams to control phase and amplitude), semiconductor optical amplifiers, and variable attenuators — hardware already manufactured at scale for fiber-optic communications. A network of just four such modules reached 98.4% accuracy on a nonlinear classification task that linear-only optical systems cannot solve at all. The catch is that these are simulated results, not physical hardware demonstrations, and the architecture's expressivity is deliberately constrained by the four-parameter physics of each module.
Catalyst
Kolmogorov-Arnold networks were only formally introduced to the machine learning community in 2024, giving the photonics community a new architecture template that maps unusually well onto optical hardware — each trainable edge function corresponds naturally to a physical transfer function. Simultaneously, telecom component manufacturers have driven down the cost and improved the reliability of the exact parts this design requires, making a simulation-to-hardware translation more credible than it would have been five years ago.
What's New
Prior photonic neural network architectures, such as systems based on unitary optical meshes (matrices of beam splitters and phase shifters), handle nonlinearity by converting the optical signal to an electrical one, applying a nonlinear function digitally, then converting back to light — a sequence called an optical-electrical-optical bottleneck that erases much of the speed and energy advantage of photonic computing. This paper replaces that pattern entirely: the nonlinearity comes from gain saturation in a semiconductor optical amplifier, meaning the light signal is transformed nonlinearly while staying optical. The authors claim this enables end-to-end differentiable training (gradient-based optimization directly on the physical parameters) without any electronic nonlinearity stage.
The Counter
Every result in this paper is a simulation. The authors model hardware impairments — quantization noise, signal-to-noise ratio degradation — but they have not built the physical system, and the gap between a differentiable physics model and actual fiber-optic bench behavior is notoriously wide. The networks tested are tiny: four modules, benchmarked on toy classification tasks like XOR variants and small image patches, not real-world inference workloads. A four-module network achieving 98.4% on a nonlinear classification benchmark sounds impressive until you realize that any cheap microcontroller running a two-layer MLP does the same task with less engineering complexity. The Kolmogorov-Arnold network architecture itself is still being stress-tested by the broader ML community and has not yet demonstrated decisive advantages over standard multilayer perceptrons at scale. Finally, the telecom components cited — semiconductor optical amplifiers in particular — are analog, temperature-sensitive, and notoriously difficult to calibrate precisely; real-world drift could undermine the tight parameter control the training procedure assumes.
Longs
- LITE — Lumentum, a maker of photonic integrated circuit components including MZIs and optical amplifiers directly relevant to this architecture
- IIVI (now Coherent, COHR) — major telecom photonics component supplier whose catalog aligns with the hardware this design uses
- NPTN — NeoPhotonics / acquired by II-VI, but segment exposure to advanced optical components
- BOTZ (robotics/automation ETF) — indirect exposure to edge inference hardware trends
- VIAVI Solutions (VIAV) — optical test and measurement equipment needed to validate photonic AI hardware
Shorts
- Companies building photonic AI chips around linear optical meshes with electronic nonlinearities (e.g., early Lightmatter Mercury architecture) — if all-optical nonlinearity proves practical, their hybrid OEO approach becomes a transitional dead end
- Nvidia (NVDA) — not immediately threatened, but any viable ultrafast optical inference at the edge reduces the addressable market for GPU-based inference servers in latency-sensitive applications
Enablers (Picks & Shovels)
- Standard telecom MZI module suppliers (Lumentum, Coherent) — the paper's design explicitly uses commodity components from this supply chain
- PyTorch / JAX autodifferentiation frameworks — the end-to-end differentiable physics model in the paper is implemented in software and depends on these tools
- KAN open-source implementations (pyKAN on GitHub) — the software KAN framework the photonic design is derived from
- Silicon photonics foundries (GlobalFoundries, IMEC) — eventual fabrication pathway for integrated versions of this architecture
Private Watchlist
- Lightmatter — photonic AI chip startup building optical matrix processors
- Luminous Computing — photonic neural network inference hardware
- Xanadu — photonic quantum and classical computing, relevant photonics fabrication expertise
- Salience Labs — integrated photonics for AI inference
Resources
The Paper
Photonic neural networks promise ultrafast inference, yet most architectures rely on linear optical meshes with electronic nonlinearities, reintroducing optical-electrical-optical bottlenecks. Here we introduce small-scale photonic Kolmogorov-Arnold networks (SSP-KANs) implemented entirely with standard telecommunications components. Each network edge employs a trainable nonlinear module composed of a Mach-Zehnder interferometer, semiconductor optical amplifier, and variable optical attenuators, providing a four-parameter transfer function derived from gain saturation and interferometric mixing. Despite this constrained expressivity, SSP-KANs comprising only a few optical modules achieve strong nonlinear inference performance across classification, regression, and image recognition tasks, approaching software baselines with significantly fewer parameters. A four-module network achieves 98.4\% accuracy on nonlinear classification benchmarks inaccessible to linear models. Performance remains robust under realistic hardware impairments, maintaining high accuracy down to 6-bit input resolution and 14 dB signal-to-noise ratio. By using a fully differentiable physics model for end-to-end optimisation of optical parameters, this work establishes a practical pathway from simulation to experimental demonstration of photonic KANs using commodity telecom hardware.