In situ fine-tuning of in silico trained Optical Neural Networks

Optical Neural Networks (ONNs) offer ultrafast computation, high bandwidth, and low energy use by exploiting photonic hardware. Training, however, remains difficult since parameters learned on simplified in silico models often degrade when transferred to physical devices with noise and fabrication imperfections. We study how noise misspecification during in silico training impacts performance and propose Gradient-Informed Fine-Tuning (GIFT). This is a lightweight in situ algorithm that adapts pretrained parameters. GIFT requires no full retraining or elaborate setups and comes with mathematical guarantees on when it improves performance. Simulations on a five-layer ONN for MNIST classification show that GIFT yields up to 28% relative accuracy improvement under noise misspecification. Thus, GIFT offers a practical means to close the gap between digital models and physical ONNs.