Synthetic Nonlinearity in Diffractive Optical Networks via Input‐Adaptive Modulation for Intelligent Vision
TL;DR
This work proposes an adaptive modulation strategy driven by the initial input that synthesizes effective nonlinearity within a standard linear optical framework, and reveals that shallow nonlinear architectures can outperform significantly deeper linear counterparts, demonstrating that computational nonlinearity can effectively substitute physical depth.
Abstract
Optical computing offers a promising avenue for high‐speed, low‐power information processing but is fundamentally constrained by the inherent linearity of standard diffractive architectures, which severely restricts representational capacity. Existing solutions often rely on exotic nonlinear materials, imposing fabrication complexity and limiting hardware compatibility. Here, we propose an adaptive modulation strategy driven by the initial input that synthesizes effective nonlinearity within a standard linear optical framework. By directly mapping the original input instance to the phase modulation parameters, we break the linearity constraint without requiring exotic media or optical gain. We demonstrate this paradigm through end‐to‐end optical edge extraction, achieving SSIM values consistently exceeding 0.9. Notably, our results reveal that shallow nonlinear architectures can outperform significantly deeper linear counterparts, demonstrating that computational nonlinearity can effectively substitute physical depth. The proposed system exhibits exceptional robustness against occlusion and noise, alongside scalable parallelism with minimal inter‐channel crosstalk. When deployed as an optical front‐end, it consistently enhances downstream classification performance, establishing a pathway for physically realizable, high‐performance intelligent optical sensors.