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

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Nanophotonics

Background:

• All-optical diffractive deep neural networks (D2NNs) show great promise in various applications.
• Manufacturing complexity and efficiency are key challenges for D2NN implementation.
• Discretizing grating structures can simplify fabrication and reduce costs.

Purpose of the Study:

• To propose a novel method for discretizing D2NNs.
• To enhance the manufacturing efficiency of D2NNs while maintaining performance.
• To optimize the quantization strategy for all-optical D2NNs.

Main Methods:

• Developed a phase-limited quantization-aware training (PLQAT) method.
• Constructed an all-optical D2NN for MNIST image classification.
• Investigated the impact of different bit levels on network performance.

Main Results:

• The PLQAT method improved D2NN classification performance by 0.11-27.96% compared to classical algorithms.
• 3-bit quantization was identified as optimal, discretizing phase values to eight levels.
• Achieved a test accuracy of 96.22% for MNIST classification with the optimized D2NN.

Conclusions:

• The PLQAT method effectively discretizes D2NN grating heights, simplifying fabrication.
• This approach significantly reduces manufacturing difficulty while preserving high network performance.
• The study highlights the potential of quantization-aware training for practical D2NN realization.