Enterprise AI Research Analysis
Small-scale photonic Kolmogorov-Arnold networks using standard telecom nonlinear modules
Authors: Luca Nogueira Calçado, Sergei K. Turitsyn, Egor Manuylovich
Aston Institute of Photonic Technologies (AiPT), Aston University, Birmingham, B4 7ET, United Kingdom.
This paper introduces 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 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 94.3% accuracy on nonlinear classification; a seven-module network attains R² = 0.986 on six-input regression. Performance remains robust under realistic hardware impairments, maintaining high accuracy down to 6-bit input resolution and 14 dB signal-to-noise ratio. Using a fully differentiable physics model for end-to-end optimisation, this work establishes a practical pathway from simulation to experimental demonstration of photonic KANs using commodity telecom hardware.
Key Enterprise Impact Metrics
Leveraging standard telecom components, this research pioneers practical, high-performance photonic AI with significant implications for speed, energy efficiency, and interpretability.
0
Nonlinear Classification Accuracy
0
High-Dimensional Regression Performance
0
Parameters Per Trainable Module
0
Energy Per Inference (at 1 GHz)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
The Photonic KAN Architecture
Explore the novel MZI-VOA-SOA-VOA module that forms the core of our small-scale photonic Kolmogorov-Arnold Networks (SSP-KANs). This architecture places trainable univariate nonlinear functions on network edges, leveraging gain saturation and interferometric mixing for all-optical computation, circumventing OEO conversion bottlenecks.
Enterprise Process Flow
Optical Input Encoding
→
MZI-VOA-SOA-VOA Module
→
Gain Saturation & Interferometric Mixing
→
Nonlinear Transfer Function
→
Optical Summation on PDs
→
Nonlinear Decision Boundary
Each network edge in an SSP-KAN employs a trainable nonlinear module comprising a Mach-Zehnder interferometer, a semiconductor optical amplifier (SOA), and variable optical attenuators (VOAs). This setup provides a four-parameter transfer function derived from gain saturation and interferometric mixing.
The Role of Trainable Parameters
Each MZI-VOA-SOA-VOA module provides four independently trainable parameters: SOA injection current (I), input attenuation (α₁), output attenuation (α₂), and interferometric phase (φ). These parameters control the small-signal gain, saturation nonlinearity, operating point, output scaling, and interference conditions, respectively. This four-parameter family of transfer functions spans a rich space of nonlinear input-output relationships, allowing the network to adapt to diverse tasks.
Impact: Despite this constrained expressivity, SSP-KANs achieve strong nonlinear inference, demonstrating that physically constrained photonic nonlinearities can act as effective trainable activation functions within Kolmogorov-Arnold Networks.
Benchmarking Across AI Tasks
Discover how SSP-KANs achieve strong nonlinear inference performance across classification, regression, and image recognition tasks. Our results demonstrate that even with a few optical modules, SSP-KANs can approach software baselines with significantly fewer trainable parameters.
94.3%
Achieved on Two Moons (4-module SSP-KAN)
The small-scale photonic KAN (SSP-KAN) achieved strong nonlinear inference performance, demonstrating its capability to handle complex classification tasks with significantly fewer parameters than traditional MLPs.
R² = 0.986
Attained on 6-input Regression (7-module SSP-KAN)
For higher-dimensional tasks like yacht hydrodynamics regression, the SSP-KAN demonstrated excellent R² performance, highlighting its ability to learn complex multivariate functions.
| Model | Accuracy | Trainable Parameters |
|---|---|---|
| SSP-KAN [2,2] | 94.3% | 16 |
| Deeper SSP-KAN [2,2,2] | 99.1% | 32 (8 modules) |
| Linear Baseline | 89.2% | 6 |
| Software KAN (pyKAN, G=1) | 99.9% | 40 |
The SSP-KAN approaches software baselines with significantly fewer parameters, demonstrating parameter-efficient performance in nonlinear classification.
| Model | R² Score | Trainable Parameters |
|---|---|---|
| SSP-KAN [6,1,1] | 0.986±0.015 | 28 |
| Single-layer SSP-KAN [6,1] | 0.921±0.017 | 24 |
| Linear Regression | 0.664 | 6 |
| MLP [6→4→1] | 0.952±0.103 | 33 |
| Software KAN (pyKAN, G=1) | 0.977 | 60 |
Compositional depth significantly improves performance, allowing the deeper SSP-KAN to exceed software KAN baselines with fewer parameters.
Real-world Practicality & Scalability
Examine the robustness of SSP-KANs under realistic hardware impairments, including finite input resolution and optical noise. This section also covers the practical pathways for experimental demonstration using commodity telecom hardware and the energy efficiency potential.
14 dB SNR
Maintained high accuracy down to 6-bit input resolution
Performance remains robust even under realistic hardware impairments, including finite digital-to-analogue converter (DAC) resolution and optical noise, making it practical for real-world deployment.
Experimental Realization with Telecom Components
The SSP-KAN architecture is designed for practical implementation using off-the-shelf telecommunications components. Key elements include a Mach-Zehnder interferometer (MZI), a semiconductor optical amplifier (SOA), and variable optical attenuators (VOAs). Each module is controlled by four physical parameters: SOA injection current, input attenuation, output attenuation, and interferometric phase. All simulations are based on manufacturer-specified parameters for a commercially available SOA (Thorlabs BOA1554P), establishing a direct pathway from simulation to experimental demonstration.
Impact: This approach significantly lowers the barriers to experimental realization of photonic KANs, leveraging mass-produced, well-characterized, and optimized telecom hardware.
~200 pJ
Per inference at 1 GHz (with low-power SOAs)
While the initial BOA1554P SOAs are for high-output power, a low-power SOA variant (Gmax ~10-15 dB) could significantly reduce energy draw to ~200 pJ per inference at 1 GHz, making it comparable to bare electronic MACs.
Quantify Your AI Transformation ROI
Estimate the potential annual cost savings and productivity gains by implementing advanced AI solutions in your enterprise.
Your AI Implementation Roadmap
We guide enterprises through a structured journey from initial concept to full-scale AI deployment, ensuring seamless integration and measurable results.
Component Characterization & Model Calibration
Accurate measurement of SOA device-level constants (saturation power, maximum small-signal gain, transparency current, linewidth enhancement factor) and calibration of the physics-based model. This initial step ensures high fidelity between simulation and physical hardware.
MZI Phase Stabilization
Active phase stabilization for Mach-Zehnder interferometers using standard thermo-electric cooler (TEC) modules and low-bandwidth feedback. This maintains the integrity of the trained transfer function against environmental drifts.
Small-Scale Prototype Assembly
Assembly of a 4-7 module fibre-coupled prototype on an optical bench using commercially available telecom components. This direct, cost-effective approach bypasses custom fabrication or cleanroom access requirements.
Simulation-to-Hardware Validation
Transfer of simulation-trained optical parameters to the physical prototype. This validates the fidelity of the physics model and the end-to-end training framework, fine-tuning parameters for optimal hardware performance.
Performance Benchmarking & Optimization
Rigorous evaluation of the physical SSP-KAN on classification, regression, and image recognition tasks, including robustness testing against hardware impairments like noise and quantization. Iterative optimization ensures peak real-world performance.
Scalability & Integration Roadmap
Planning for larger-scale deployments, transitioning to integrated photonic platforms for compactness and thermal stability. Exploration of wavelength- or space-division multiplexing for parallel processing and future AI co-design strategies.
Ready to Transform Your Enterprise with Photonic AI?
Our experts are prepared to discuss how small-scale photonic KANs can accelerate your AI inference, reduce energy costs, and provide transparent, explainable models. Schedule a personalized consultation to explore tailored strategies for your business.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat