Phase-limited quantization-aware training for diffractive deep neural networks

In recent years, all-optical diffractive deep neural networks (<i>D</i>²<i>N</i><i>N</i><i>s</i>) have demonstrated exceptional performances in many fields. Discretizing the grating height can reduce the complexity and enhance the network manufacturing efficiency. For this purpose, we propose a phase-limited quantization-aware training (PLQAT) method and construct an all-optical <i>D</i>²<i>N</i><i>N</i> to discretize the network for the MNIST image classification task. Our results indicate that the PLQAT method improves the classification performance of the <i>D</i>²<i>N</i><i>N</i> by 0.11-27.96% across different bit levels compared to the classical algorithm. Furthermore, we identified 3-bit quantization as the optimal choice, discretized the phase values in five layers to eight levels within [0,2<i>π</i>], and achieved a test accuracy of 96.22%. This method discretizes the height of gratings of <i>D</i>²<i>N</i><i>N</i><i>s</i>, effectively reducing the difficulty of grating etching while maintaining good network performance.