Residual calibration for high-precision optical neural networks.

Optical processors have emerged as promising platforms for accelerating matrix-vector multiplications (MVMs), offering significant advantages in energy efficiency and low latency for applications such as optical neural networks (ONNs). However, errors in existing optical analog architectures limit computational accuracy and scalability, posing a critical challenge in optical computing. In this work, we propose a residual calibration method that iteratively refines optical computations using multiple low-precision multiplications to achieve high-precision matrix products. Theoretical analysis demonstrates that the method reduces computational errors at an exponential rate, contingent on the condition that the maximum singular value of the deviation matrix remains below unity. Experimental validation conducted on fabricated optical processors has confirmed the effectiveness of the proposed residual calibration, achieving a significant error reduction across successive iterations. Additionally, we demonstrate the tangible benefits of the residual calibration method through applications in ONNs performing semantic segmentation tasks. A single calibration iteration restores ONN performance to levels comparable to digital implementations, resulting in a 24% improvement in mean intersection-over-union and a 22% enhancement in pixel accuracy. This work provides a flexible and scalable solution to the persistent challenge of achieving high-precision computations on optical platforms, significantly advancing the feasibility of practical deployments in demanding computational scenarios.