🔬 optics

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

This paper introduces small-scale photonic Kolmogorov-Arnold networks (SSP-KANs) built entirely from standard telecom components, demonstrating that fully optical, trainable nonlinear modules can achieve high-accuracy inference across diverse tasks while overcoming traditional optical-electrical bottlenecks and remaining robust to hardware impairments.

Original authors: Luca Nogueira Calçado, Sergei K. Turitsyn, Egor Manuylovich

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Luca Nogueira Calçado, Sergei K. Turitsyn, Egor Manuylovich

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a super-fast, energy-efficient computer that thinks using light instead of electricity. This is the dream of Photonic Neural Networks.

However, most current attempts have a major flaw: they are like a relay race where the runner (light) has to stop at every station to hand off a baton to a human (electronics) to do some thinking, and then hand it back to a runner. This "stop-and-go" process creates a traffic jam (bottleneck) that slows everything down and wastes energy.

This paper introduces a new way to build these light-based computers using a clever mathematical trick called a Kolmogorov-Arnold Network (KAN). Here is the simple breakdown of what they did and why it matters.

1. The Old Way vs. The New Way

The Old Way (MLP): Imagine a factory assembly line where a machine does a simple task, then a human steps in to add a "twist" (nonlinearity), then the machine does another task, and another human adds a twist. The humans are the electronics; they are slow and expensive to run.
The New Way (KAN): Imagine a team of artists. Instead of one big machine doing everything, you have many small artists. Each artist takes a single piece of raw material, adds their own unique "twist" (nonlinearity), and then everyone just puts their finished pieces into a big pile together. The pile is the final answer.
- The Magic: In this new system, the "twist" happens before the pieces are combined. This means we don't need the slow human (electronics) to stop the light. The light can do the twisting itself!

2. The "Magic Box" (The Optical Module)

The researchers needed a way for light to "twist" itself. They built a small, compact box using standard parts you can buy at any telecom store (like the ones used for internet cables).

Think of this box as a Light Sculptor:

The Splitter: It takes a beam of light and splits it into two paths.
The Amplifier (The Muscle): One path goes through a special glass amplifier (a Semiconductor Optical Amplifier). When you push enough light through it, it gets "tired" and stops amplifying linearly. It starts to squish and stretch the light in a complex, non-straight way. This is the "twist."
The Mixer: The two paths (the twisted one and the straight one) are smashed back together. Because light waves interfere with each other (like ripples in a pond), they create a complex pattern.

By adjusting four simple knobs on this box (how much electricity to the amplifier, how much to dim the light before/after, and the timing of the waves), they can make the box output almost any shape of curve they want.

3. The "Small-Scale" Surprise

Usually, to make a smart computer, you need millions of parameters (knobs). But the researchers found something surprising: You don't need a giant network to do smart things.

They built a tiny network with only 4 to 8 of these "Light Sculptor" boxes.

The Test: They asked this tiny network to solve puzzles that are impossible for simple straight-line logic (like separating two crescent moon shapes that are interlocked).
The Result: The tiny network solved it with 98.4% accuracy, almost as good as the massive, complex software programs running on supercomputers.

The Analogy: It's like realizing you don't need a massive orchestra to play a beautiful song; sometimes, a small quartet of talented musicians (the optical modules) playing together in harmony is enough to create something profound.

4. Why This Matters (The "Real World" Test)

The researchers didn't just simulate this on a computer; they modeled it using real, off-the-shelf hardware parts. They tested if the system would break if the light was a bit noisy or if the controls weren't perfectly precise (like using a cheap dial instead of a laser-precise one).

The Good News: The system is incredibly robust. Even if the signal is a bit "fuzzy" (noisy) or the controls are only 6-bit (like an old video game controller), it still works great.
The Bad News: If the noise gets too loud, the system reverts to a simple straight line, but it doesn't crash. It gracefully degrades.

5. The Big Picture

This paper is a roadmap. It says: "We don't need to invent exotic, expensive, custom-made lasers to build optical AI. We can use the standard parts already in our fiber-optic internet cables."

By using a specific mathematical structure (KANs) that fits perfectly with how light naturally behaves, they have opened a door to:

Ultra-fast AI: Processing data at the speed of light.
Low Energy: No electricity-hungry electronics slowing things down.
Real-World Use: Since they used standard telecom parts, this technology could be built in a lab tomorrow and eventually put into data centers to handle the massive data traffic of the future.

In a nutshell: They figured out how to make light "think" using simple, cheap tools and a clever mathematical recipe, proving that small, physically constrained networks can be surprisingly smart.

1. Problem Statement

Current Photonic Neural Networks (PNNs) largely follow the Multilayer Perceptron (MLP) paradigm, where linear optical matrix-vector multiplication is combined with electronic nonlinear activation functions. This hybrid approach creates Optical-Electrical-Optical (OEO) bottlenecks, reintroducing latency and energy consumption that negate the primary advantages of all-optical computing.

While Kolmogorov-Arnold Networks (KANs) offer a promising alternative by placing trainable univariate nonlinear functions on network edges (rather than fixed activations on nodes), previous photonic implementations have relied on:

Specialized resonator geometries (e.g., ring-assisted MZIs) requiring custom fabrication.
Free-space spatial light modulators.
Complex integration not easily compatible with mass-produced telecommunications infrastructure.

The Core Challenge: Can a photonic KAN be implemented using standard, off-the-shelf telecommunications components with limited tunable parameters (physically constrained) while still achieving performance comparable to software-based KANs and MLPs?

2. Methodology: SSP-KAN Architecture

The authors propose Small-scale Photonic Kolmogorov-Arnold Networks (SSP-KANs), implemented entirely with standard telecom hardware.

A. The Nonlinear Module (MZI-SOA-VOA)

The fundamental building block is a compact module consisting of:

Mach-Zehnder Interferometer (MZI): Splits input light into two arms.
Semiconductor Optical Amplifier (SOA): Placed in one arm. It provides gain saturation (a native nonlinearity in telecom) and a nonlinear phase shift via the linewidth enhancement factor.
Variable Optical Attenuators (VOAs): Two VOAs are used.
- VOA1 (Pre-SOA): Controls input power to the SOA, setting the operating point on the gain saturation curve.
- VOA2 (Post-SOA): Scales the amplified arm's contribution to the interference without altering saturation dynamics.

Trainable Parameters: Each module has exactly four tunable physical parameters:

SOA injection current ( $I$ ).
Input attenuation ( $\alpha_1$ ).
Output attenuation ( $\alpha_2$ ).
Interferometric phase shift ( $\phi$ ).

B. Network Structure

Following the KAN theorem, the network places a dedicated MZI-SOA-VOA module on every edge connecting input and output nodes.

Computation: Inputs are encoded as optical power. Each edge applies a nonlinear transfer function $\phi_{i,j}(x_i)$ . Outputs are computed by summing the powers of the transformed signals on a photodetector (incoherent summation).
Differentiability: The authors developed a fully differentiable physics-based model based on the Agrawal–Olsson formalism for SOA gain dynamics. This allows for end-to-end gradient optimization (backpropagation) of all physical parameters (current, attenuation, phase) using standard deep learning frameworks (PyTorch).

C. Simulation & Training

Hardware Model: Simulations use parameters from a commercially available SOA (Thorlabs BOA1554P).
Robustness Testing: The models are trained and tested under realistic hardware impairments:
- Optical Noise: Amplified Spontaneous Emission (ASE) modeled as signal-dependent noise.
- Quantization: Finite-resolution Digital-to-Analog Converters (DACs) for input encoding (tested from 12-bit down to 4-bit).

3. Key Contributions

Commodity Hardware Implementation: Demonstrated that standard telecom components (MZI, SOA, VOA) can function as trainable nonlinear units, removing the need for bespoke nanofabrication.
Physics-Aware Training: Established a framework for end-to-end optimization of physical parameters via a differentiable physics model, bridging the gap between simulation and experimental reality.
Parameter Efficiency: Showed that despite having only 4 parameters per edge (compared to hundreds in B-spline software KANs), the network achieves competitive performance through compositional depth (stacking layers) rather than just increasing width.
Coherent vs. Incoherent Analysis: Provided a detailed theoretical and empirical comparison of coherent (field) vs. incoherent (power) summation, revealing that incoherent summation is superior for shallow networks due to simpler optimization landscapes, while coherent summation offers benefits in deeper networks.

4. Key Results

A. Nonlinear Classification (Two Moons Dataset)

Task: Binary classification of interleaving crescents (nonlinearly separable).
Performance: A single-layer [2,2] SSP-KAN (4 modules, 16 parameters) achieved 98.4% accuracy.
Comparison:
- Linear Baseline: 89.2%.
- Software KAN (pyKAN): 99.9% (40 parameters).
- Insight: The SSP-KAN closed the gap to within 1.5% using fewer than half the parameters of the software baseline, proving that SOA gain saturation provides sufficient functional diversity.

B. Multivariate Regression (Yacht Hydrodynamics)

Task: Predicting ship resistance from 6 hull geometry features.
Performance:
- Single-layer [6,1] SSP-KAN: $R^2 = 0.869$ (limited by lack of feature interaction).
- Two-layer [6,1,1] SSP-KAN: $R^2 = 0.977$ .
Comparison: Matched the software KAN ( $R^2 = 0.977$ ) using 28 parameters vs. 60 for the software baseline. This confirms that cascading physically constrained nonlinearities captures multivariate interactions effectively.

C. Image Classification (MNIST & Fashion-MNIST)

MNIST: SSP-KAN [784,20,10] achieved 92.7% accuracy (vs. 91.0% for linear baseline).
Fashion-MNIST: SSP-KAN achieved 85.7% accuracy (vs. 83.3% for linear baseline).
Significance: The performance gap widens on Fashion-MNIST (a harder, more nonlinear task), confirming that optical nonlinearity provides meaningful benefits where feature interactions are critical.

D. Robustness Analysis

Quantization: Performance was negligible down to 6-bit DAC resolution. Low-resolution DACs are sufficient for input encoding.
Optical Noise: Performance degraded gracefully with Signal-to-Noise Ratio (SNR).
- At SNR = 14 dB, accuracy remained >96% (Two Moons) and $R^2 > 0.91$ (Yacht).
- At SNR = 6 dB, performance converged to a noise floor but did not collapse catastrophically.
Conclusion: Optical noise is the dominant impairment, not DAC resolution.

5. Significance and Future Outlook

Practical Pathway: This work establishes a realistic route to deployable all-optical computing systems using commodity telecom hardware, significantly lowering barriers to experimental validation compared to custom silicon photonics.
Latency-Critical Applications: The architecture is particularly suited for optical channel equalization in coherent communications, where electronic DSP struggles with latency at high data rates.
Scalability: While the current study uses small-scale networks (4–8 modules) realizable on an optical bench, the authors note that scaling to high-dimensional tasks (like MNIST) will require integrated photonic platforms for thermal stability and footprint reduction.
General Framework: The differentiable physics modeling approach is generalizable to other nonlinear mechanisms (e.g., Kerr, Brillouin scattering), providing a blueprint for co-designing photonic hardware and neural network architectures.

In summary, the paper proves that physically constrained, low-parameter photonic KANs are sufficient for meaningful nonlinear inference, offering a viable, energy-efficient, and ultrafast alternative to electronic AI for specific latency-critical tasks.