Research map / Architectures and efficiency / RNNs and normalization
Physics-Constrained Comprehensive Optical Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a physics-constrained learning framework for optical neural networks that combines an intensity-gap loss for laser instability, CCD exposure adjustment, and a lightweight physics-prior error-compensation DNN, raising experimental MNIST accuracy from 61.6% to 96.5%.
Problem
Experimental optical neural networks lose accuracy because of a gap between ideal simulated models and physical systems, caused by laser instability, CCD exposure mismatch, and other unmeasured systematic errors.
Approach
Injects quantifiable physical priors (laser intensity fluctuation via an intensity-gap loss, and CCD exposure calibration so simulated and measured overall intensity match) into a residual error-compensation DNN on top of an ideal Fresnel-diffraction 4f model, iterated with SLM phase retraining.
Key results
- Experimental MNIST accuracy rises from 61.6% (original ONN) to 96.5% with the proposed framework.
- QuickDraw16 experimental accuracy improves from 63.0% to 85.7%.
- Fashion MNIST experimental accuracy improves from 56.2% to 77.5% (text elsewhere reports 77.3%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add quantifiable physical information to error-compensation training (ablation) | Optical neural network · MNIST | accuracy: 93.5% -> 96.5% | Helped |
| Add the error compensation DNN (ablation) | ONN with quantitative physical information · MNIST | accuracy: 83.7% -> 96.5% | Helped |
| Add quantifiable physical information and error compensation | Optical neural network · QuickDraw16 | accuracy: 63.0% -> 85.7% | Helped |
| Add quantifiable physical information and error compensation | Optical neural network · Fashion MNIST | accuracy: 56.2% -> 77.5% | Helped |
| Optimize using the proposed method | Optical neural network · CIFAR-10 | accuracy: 30% -> 57% | Helped |
Surprising findings
- Adjusting CCD exposure alone to match overall intensity raised MNIST accuracy to 83.7% even without an error compensation network.
Limitations
- The intensity-gap loss may not transfer to other AI tasks beyond image classification.
- Evaluation uses a single physical optical 4f setup and only 1,000 test images per dataset.
- CIFAR-10 results (30% to 57%) are reported without comparative benchmarks.
Experimental setup
Datasets: MNIST QuickDraw16 Fashion MNIST CIFAR-10
Benchmarks: MNIST QuickDraw16 Fashion MNIST
Metrics: classification accuracy
Baselines: Hybrid training CNN error compensation DAT PAT Original ONN (no compensation)
Builds on: physics-informed neural networks hardware-in-the-loop training error compensation neural network
Closest papers by meaning
- Neural Tangent Knowledge Distillation for Optical Convolutional Networks (NeurIPS 2025 · similarity 0.88)
- Layered-Parameter Perturbation for Zeroth-Order Optimization of Optical Neural Networks (AAAI 2025 · similarity 0.85)
- All-Optical Nonlinear Diffractive Deep Network for Ultrafast Image Denoising (CVPR 2025 · similarity 0.81)
- Gradient Alignment in Physics-informed Neural Networks: A Second-Order Optimization Perspective (NeurIPS 2025 · similarity 0.80)
- PACE: Pacing Operator Learning to Accurate Optical Field Simulation for Complicated Photonic Devices (NeurIPS 2024 · similarity 0.80)
- Challenges in Training PINNs: A Loss Landscape Perspective (ICML 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.