Reproduce “Physics-aware Differentiable Discrete Codesign for Diffractive Optical Neural Networks” on OpenResearch

About this paper

Yingjie Li, Ruiyang Chen, Weilu Gao, Cunxi Yu

Researchers at the University of Utah introduced a physics-aware differentiable discrete codesign framework utilizing Gumbel-Softmax for Diffractive Optical Neural Networks (DONNs). This framework directly integrates real-world discrete optical device characteristics into end-to-end training, achieving higher prediction accuracy and greater deployability on physical hardware compared to conventional quantization methods.

Key insights

• Gumbel-Softmax provides an effective differentiable approximation for discrete variables, which is crucial for end-to-end training of DONNs with real-world, non-ideal optical device responses.
• Directly embedding calibrated device characteristics into the training process ensures that learned parameters are inherently physically feasible and deployable, avoiding performance degradation from post-processing.
• Dynamic temperature annealing schedules for Gumbel-Softmax are critical for optimizing DONN performance, enabling both broad exploration and fine-tuning during training.

Summary by alphaXiv.