Speaker
Description
Optical diffractive neural networks (ODNNs) perform computation through light propagation and diffraction, yet their internal physical decision mechanism remains difficult to interpret. In this work, we propose Gradient-based Physical Activation Mapping (GPAM), a method that quantifies the contribution of each diffractive neuron by back-propagating gradients through physical wave propagation. Beyond visualization, we establish a quantitative evaluation framework for optical neural network interpretability, including peak localization, foreground energy concentration, weakly-supervised segmentation, and faithfulness analysis based on mask ratio. Experiments on MNIST and Fashion-MNIST datasets show that GPAM relevance maps are spatially consistent with object regions and that masking high-relevance regions leads to significant confidence and accuracy drops. These results demonstrate that GPAM not only visualizes where the optical neural network focuses but also quantifies how accurate and faithful the explanations are. The proposed framework provides a practical tool for optical system analysis, alignment diagnosis, and interpretable optical neural network design.
Keywords: Optical diffractive neural networks, Interpretability, Optical computing, Saliency map, Optical neural networks