Mutual learning knowledge distillation framework for adaptive optimization of optical diffractive neural network

This is the authors' abstract. We don't add key points for this paper.

The growing computational demands of artificial intelligence motivate the development of optical neural networks (ONNs), while their practical deployment remains constrained by optimization complexity and limited adaptability. We propose a mutual learning knowledge distillation framework for optical diffractive neural networks (ODNNs), synergizing photonic computing with adaptive training strategies. Our architecture enables bidirectional knowledge transfer between parameter-efficient optical student networks and electrical ResNet teachers, leveraging both labeled data and probabilistic guidance. The validation on Fashion-MNIST and Kuzushiji-MNIST demonstrates classification accuracy improvements of 5.9% and 13.88%, respectively, over conventional ODNNs training. In addition, we further developed the photonic to photonic mutual learning computational paradigm, in which an optical network with 0.98 × 106 parameters can bootstrap compact photonic classifiers with 0.19 × 106 parameters, achieving an accuracy of 79.5% on MNIST and yielding an absolute improvement of 9% relative to the noncooperative baseline. Even though the all-optical mutual learning paradigm still has a gap of about 19% with the state-of-the-art ODNNs, the number of parameters of the guided student networks is smaller than the state-of-the-art networks by several or even dozens of times. These results establish a paradigm for overcoming fundamental limitations in ONN optimization, paving the way for a new era in which ONNs can solve large-scale practical problems through photonic–electronic and photonic–photonic co-design.