ReDON: Recurrent Diffractive Optical Neural Processor with Reconfigurable Self-Modulated Nonlinearity
Unlocking Next-Gen AI: The Power of ReDON
Diffractive optical neural networks (DONNs) have demonstrated unparalleled energy efficiency and parallelism by processing information directly in the optical domain. However, their computational expressivity is constrained by static, passive diffractive phase masks that lack efficient nonlinear responses and reprogrammability. To address these limitations, we introduce Recurrent Diffractive Optical Neural Processor (ReDON), a novel architecture featuring reconfigurable, recurrent self-modulated nonlinearity. This mechanism enables dynamic, input-dependent optical transmission through in-situ electro-optic self-modulation, providing a highly efficient and reprogrammable approach to optical computation. Inspired by the gated linear unit (GLU) in large language models, ReDON senses a fraction of the propagating optical field and modulates its phase or intensity via a lightweight, parametric function, enabling effective nonlinearity with minimal inference overhead. As a non-von Neumann architecture with the main weighting units (metasurfaces) being fixed, we substantially extend the DONN's nonlinear representational capacity and task adaptability via recurrent optical hardware reuse and dynamically tunable nonlinearity. We systematically investigate various self-modulation configurations to uncover the trade-offs between hardware efficiency and expressivity. On image recognition and segmentation tasks, ReDON improves test accuracy and mIoU by up to 20% over prior DONNs with optical or digital nonlinearities at comparable complexity and negligible power overhead. This work establishes a new paradigm for reconfigurable nonlinear optical computing, uniting the benefits of recurrence and self-modulation in non-von Neumann analog processors.
Authored by ZIANG YIN, QI JING, RAKTIM SARMA, ZHAORAN RENA HUANG, YU YAO, JIAQI GU
Executive Summary: ReDON's Transformative Impact
ReDON revolutionizes diffractive optical computing, overcoming fundamental limitations to deliver unprecedented performance and adaptability. This innovation is critical for next-generation AI hardware, enabling faster, more efficient, and versatile optical neural networks for demanding enterprise applications.
0
Accuracy Boost
0
Negligible Power Overhead
0
Peak Inference Speed
Dynamic
Modulation Reconfigurability
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
ReDON introduces a novel electro-optic self-modulated nonlinearity, inspired by GLU from LLMs. This dynamic, input-dependent mechanism provides strong, tunable, and reprogrammable nonlinear responses, overcoming the limitations of static diffractive elements. It allows for task-adaptive nonlinear functions, significantly improving expressivity over fixed or digitally added nonlinearities.
74.5%
CIFAR-10 Test Accuracy (R=2, N=4)
ReDON Self-Modulation Process Flow
Sense Optical Field Fraction (α)
→
Convert to Electrical Signal (|E|²)
→
Parametric Transformation (Ψ(·, Θ))
→
Modulate Downstream Metasurfaces (Phase/Amplitude)
| Mechanism | Nonlinearity Type | Programmability | Expressivity |
|---|---|---|---|
| Traditional DONNs | Weak, Fixed (square-law) | None | Low |
| All-Optical (e.g., saturable) | Strong (high power needed) | None | Limited |
| Structural (cavity feedback) | High-order, Fixed | None | Fixed |
| ReDON Self-Modulation | Strong, Tunable, Input-Dependent | High | Very High |
ReDON's hybrid architecture combines static, non-volatile metasurfaces (for main weighting) with lightweight electro-optic self-modulation for dynamic reconfigurability. Its recurrent inference mechanism reuses the same hardware, incrementally composing deep neural representations by updating parameters (Θ) across iterations, while keeping metasurface phases (Φ) fixed.
| Feature | Hybrid Optical-Electronic | ReDON |
|---|---|---|
| Computational Core | Static optical encoder | Recurrent, self-modulated optical core |
| Task Adaptation | Digital backend only | Optical + Digital backend (reconfigurable nonlinearity) |
| Weight Storage | Fixed metasurfaces | Fixed metasurfaces (Φ) + Dynamic SRAM (Θ) |
| Reconfigurability | Limited (mechanical/wavelength) | High (electro-optic self-modulation) |
ReDON Recurrence Mechanism
Input Encoded (E0)
→
Propagate & Modulate (Φ, Θγ)
→
Sense Fraction (α|E|²)
→
Update Θγ for Next Iteration
→
Recurrent Propagation (R iterations)
Hardware Reuse
via Recurrent Processing
ReDON demonstrates an average of 20% higher accuracy and mIoU over prior DONNs. Its self-modulation mechanism adds negligible power overhead (<1mW) while significantly boosting expressivity. The architecture supports high throughput, with over 400 FPS using commercial SLMs and potential for 1000+ FPS with advanced tunable metasurfaces.
+20%
Average Accuracy/mIoU Improvement
Enhanced Performance on Complex Tasks
On image recognition and segmentation tasks (e.g., CIFAR-10, QuickDraw-50, Stanford Background), ReDON improves test accuracy and mIoU by up to 20% over prior DONNs, even with a single block. For PDE solving (e.g., Darcy flow, Navier-Stokes), ReDON reduces the adapted solution error by 40%.
This significantly surpasses previous DONN limitations for real-world, demanding AI applications.
<1mW
Power Overhead for Self-Modulation
ReDON is designed for practical deployment, incorporating noise-aware training to maintain performance under physical non-idealities such as diffractive layer misalignment, readout noise, and fabrication errors. This approach ensures robust operation in real-world scenarios, with a minimal impact on noise-free accuracy.
91+%
Accuracy with Noise-Aware Training
ReDON Robustness Strategy
Model Non-Ideality Sources (MA, RN, FE)
→
Inject Combined System Noise (MA+RN+FE) During Training
→
Robust Model Learning
→
Maintain High Accuracy (91-92%) in Noisy Environments
| Training Strategy | Max Noise (MA 2 + RN 0.1 + FE 0.8) | Noise-Free Accuracy |
|---|---|---|
| Standard Training | 71.8% Test Accuracy | 95.1% Test Accuracy |
| Noise-Aware Training | 91.1% Test Accuracy | 92.9% Test Accuracy (slight reduction) |
Calculate Your Potential ROI with ReDON
Estimate the cost savings and efficiency gains your enterprise could achieve by integrating ReDON's advanced optical computing capabilities. This calculator provides a preliminary estimate; a custom analysis with our experts will provide precise figures.
Estimated Annual Savings
$0
Equivalent Hours Reclaimed
0
Your ReDON Implementation Roadmap
Our phased approach ensures a smooth and successful integration of ReDON into your existing AI infrastructure, maximizing impact with minimal disruption. We tailor each step to your enterprise's unique needs and technical landscape.
Phase 1: Discovery & Strategy
Comprehensive assessment of current systems, identification of key AI workloads, and collaborative strategy development for ReDON integration to meet specific enterprise objectives. This involves detailed analysis of your existing DONN or digital NN architectures and identifying optimal self-modulation configurations.
Phase 2: Hardware Customization & Prototyping
Design and prototyping of ReDON hardware tailored to your computational needs, including metasurface fabrication, SLM integration, and lightweight digital backend development. Focus on optimal power-performance trade-offs and ensuring reconfigurability.
Phase 3: Model Adaptation & Deployment
Fine-tuning of existing AI models or development of new ones for ReDON, leveraging its unique nonlinearity and recurrence. Rigorous testing and phased deployment into your production environment, with noise-aware training for robustness and real-world validation.
Phase 4: Optimization & Scalability
Continuous monitoring, performance optimization, and scaling of ReDON deployments across your enterprise. Exploration of multi-layer self-modulation and advanced tunable metasurfaces for future-proof scalability and enhanced expressivity, ensuring long-term value.
Ready to Transform Your AI Infrastructure?
Unlock unparalleled speed, efficiency, and adaptability for your enterprise AI with ReDON. Our experts are ready to guide you through every step of the journey, from initial consultation to full-scale deployment and optimization. Book a free consultation to start.
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