⚡ electrical engineering

Partially coherent high-order optical neural networks via pseudo-nonlinear encoding

This paper introduces and experimentally validates a high-order optical neural network framework under partially coherent illumination that overcomes the nonlinearity limitation of traditional diffractive systems, achieving performance comparable to electrical neural networks with significantly lower computational complexity and superior robustness.

Original authors: Weiqiang Ding, Bojie DONG, Xiaoxin Li, Zixuan Wang, Qi Jia, Bojian Shi, Shiyong Ren, Hanyi Zhang, Zhengjun Wang, Rui Feng, Fangkui Sun, Yongyin Cao, Liangcai Cao

Published 2026-08-03

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Weiqiang Ding, Bojie DONG, Xiaoxin Li, Zixuan Wang, Qi Jia, Bojian Shi, Shiyong Ren, Hanyi Zhang, Zhengjun Wang, Rui Feng, Fangkui Sun, Yongyin Cao, Liangcai Cao

Original paper licensed under CC BY 4.0 (https://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 a world where computers don't just crunch numbers with electricity, but dance with light. This is the realm of Optical Neural Networks (ONNs), a futuristic branch of science where beams of light, instead of electrons, carry information to solve complex problems like recognizing faces or diagnosing diseases. The most promising version of this technology is the Diffractive Optical Neural Network (DONN). Think of a DONN like a giant, invisible kaleidoscope: you shine a picture into it, and the light bounces off a series of special mirrors and filters, bending and interfering with itself to produce a new pattern that represents the answer.

However, there's a catch. Most of these light-based computers have been designed to work only under "perfectly coherent" light—like a laser beam where every photon marches in perfect lockstep, like soldiers in a parade. In the real world, though, light is rarely that perfect. Sunlight, lamp light, and even the light from a computer screen are "partially coherent," meaning the photons are a bit more chaotic, like a crowd of people walking through a busy market rather than a military parade. For years, scientists struggled to make these light-computers work well in this messy, real-world light. They also faced a second hurdle: light naturally wants to do math in a straight line (linear), but smart computers need to do tricky, curved math (nonlinear) to learn complex things. Usually, making light do this "curved math" requires dangerous, high-powered lasers or special materials that are hard to build.

This is where the new research comes in. A team of scientists has proposed a clever workaround called a Partially Coherent High-Order Optical Neural Network (HONN). Instead of fighting the messy nature of real-world light or using dangerous lasers, they found a way to use the chaos of the light itself, combined with a smart coding trick, to make the computer "think" nonlinearly. They call this "pseudo-nonlinear encoding." It's like teaching a chaotic crowd to dance in a specific, complex pattern just by how they enter the room, rather than forcing them to march in a line.

The team built a physical model of this idea using a visible light setup with spatial light modulators (essentially high-tech, programmable mirrors) and tested it on famous image datasets like Digit MNIST (handwritten numbers) and Fashion MNIST (clothing items). They also tried it on real-world facial expression data and medical images. The results were promising: their new system could successfully recognize images using partially coherent light, and the more "nonlinear" they made the system (by adjusting the layers of mirrors), the better it got. In fact, on some tests, their light-based computer performed just as well as traditional electrical computers but used less than 10% of the computational power.

Crucially, the researchers found that this system is surprisingly robust. Even if the light conditions in the real world didn't perfectly match the conditions the computer was trained on, it still worked well. The study suggests that by treating the "messiness" of light not as a bug, but as a feature, we can build optical computers that are ready for the real world. While the current experiments were done in a lab with simulated light conditions to prove the concept, the findings suggest a practical path forward for using light to power the next generation of machine vision and imaging, potentially making smart cameras and sensors faster and more energy-efficient than ever before.

Technical Summary: Partially Coherent High-Order Optical Neural Networks via Pseudo-Nonlinear Encoding

Problem Statement

Diffractive Optical Neural Networks (DONNs) offer a pathway to high-speed, energy-efficient computing but face a fundamental limitation: the absence of a general framework for nonlinear activation. Most existing DONNs operate under ideally coherent illumination and realize primarily linear mappings, which restricts their representational capacity and applicability in realistic environments where light sources possess partial coherence. While physical nonlinear effects (e.g., saturable absorption, second-harmonic generation) exist, they often require high optical intensities and specialized materials. Conversely, "pseudo-nonlinearity" (encoding strategies in linear systems) avoids high-intensity requirements but has not been sufficiently integrated with the physics of partially coherent light. The central challenge addressed is how to introduce effective, controllable nonlinearity into partially coherent optical neural networks in a physically consistent manner.

Methodology

The authors propose a High-Order Optical Neural Network (HONN) framework that operates under partially coherent illumination. This framework unifies two sources of nonlinearity:

Coherence-Dependent Detection Nonlinearity: Arising from the intensity response of optical detection under partially coherent light.
System-Level Pseudo-Nonlinearity: Achieved through optimized encoding strategies within the optical system.

Theoretical Framework

The model extends conventional linear-weight architectures by introducing higher-order interactions between input variables. The output $y$ is described by a high-order polynomial expansion where nonlinear terms arise from the interplay of the input amplitude, phase modulation, and the coherence matrix ( $\Gamma$ ).

Pseudo-Nonlinearity: Implemented by encoding input data into diffractive layers with phase modulation parameters ( $\Phi_i = a_i X + B_i$ ). By setting specific coefficients ( $a_i$ ) to zero, the system can be tuned to operate at different nonlinear orders (1st to 4th).
Partial Coherence Simulation: The system utilizes the Gaussian-Schell model. To simulate partially coherent propagation efficiently, the authors employ a superposition of coherent light fields with random spectra ( $M=20$ superimpositions), defined by a coherence length ( $l_c$ ).
Experimental Platform: A four-layer DONN was implemented using a visible-light spatial light modulator (SLM) setup (660 nm wavelength). The system uses two SLMs: one for input amplitude/phase encoding and a second for modulated phase loading to achieve specific nonlinear orders.

Training and Validation

The model was trained using the Rayleigh–Sommerfeld diffraction theory and the angular spectrum method, optimized via the Adam optimizer with cross-entropy loss. Validation was performed on:

Digit MNIST and Fashion MNIST: To analyze performance trends across varying coherence lengths ( $l_c$ ) and nonlinear orders.
RAF (Real-world Affective Faces) and MedMNIST: To demonstrate practical applicability on complex, real-world datasets.
Robustness Testing: Evaluating performance when test coherence lengths ( $l_{c-test}$ ) deviate from training coherence lengths ( $l_{c-train}$ ).

Key Contributions

Unified HONN Framework: The paper establishes a general framework for high-order optical inference that integrates pseudo-nonlinearity and partial coherence. It posits that physical nonlinearity and pseudo-nonlinearity are alternative implementations of a unified mathematical form.
Experimental Demonstration: The authors experimentally validated the HONN model in the visible spectrum, demonstrating that classification accuracy improves with increasing coherence length and nonlinear order.
Coherent Robustness: The study reveals that higher-order modulation strengthens effective nonlinear responses, compensating for performance degradation caused by reduced coherence. This suggests a trade-off between peak accuracy and robustness against coherence mismatch.
Efficiency: The proposed HONN achieves performance comparable to traditional electrical neural networks while utilizing less than 10% of their computational complexity.

Results

Performance Trends: On Digit MNIST, simulated accuracy increased from 79.25% to 94.83% (1st order) and 83.23% to 95.55% (3rd order) as coherence length moved from $8\Delta x$ to fully coherent. Experimental results followed similar trends, reaching up to 90.25% accuracy.
Real-World Datasets: On the RAF dataset, the partially coherent HONN achieved 72.68% accuracy, and on MedMNIST, 97.17%. When integrated with an electrical readout layer (optoelectronic hybrid), accuracy improved to 75.86% and 98.39%, respectively.
Scaling Behavior: A strong coupling between model capacity and coherence was observed. Under low coherence, over-parameterized models suffered from overfitting and optimization difficulties, whereas high coherence facilitated the training of high-capacity networks.
Robustness: Models trained at lower coherence levels ( $l_{c-train} = 8\Delta x$ ) maintained stable performance across a broader range of test coherence conditions compared to models trained at high coherence, which degraded rapidly as $l_{c-test}$ decreased.

Significance and Claims

The authors claim that this work fills a research gap by providing a unified framework for high-order optical neural computation under partially coherent illumination. They assert that partial coherence is not merely a limiting factor but a "physically meaningful degree of freedom" that shapes inference behavior.

The significance of the work lies in:

Practical Deployment: Providing a route for optical computing in machine vision, computational imaging, and sensing under natural light conditions (which are inherently partially coherent).
Generalizability: Establishing a framework that extends beyond traditional diffractive models, potentially applicable to metasurface-based systems and low-threshold nonlinear devices.
Efficiency: Demonstrating that optical inference can compete with electrical networks in accuracy while drastically reducing computational complexity.

The paper concludes that coherence, nonlinear order, and model capacity should be co-designed in partially coherent optical systems to optimize performance and robustness.