Robustness of parallel subnetwork-filtered diffractive deep neural networks

Area of Science:

• Optical computing
• Deep learning hardware
• Photonic devices

Background:

• Diffractive deep neural networks (D2NNs) offer low latency and power consumption for optical computing.
• Hardware implementation of D2NNs is hindered by challenges in aligning multiple diffractive planes.
• System errors and misalignment reduce the practical deployability of D2NNs.

Purpose of the Study:

• To propose a novel parallel subnetwork-filtered diffractive deep neural networks (PSF-D2NNs) architecture.
• To enhance the robustness of D2NNs against multi-plane misalignment and experimental errors.
• To improve the practical feasibility of D2NNs in hardware implementations.

Main Methods:

• Introduced a dataset with random misalignment errors for modulation planes.
• Employed phase filtering factors to achieve modulation phase smoothing.
• Developed and simulated the PSF-D2NNs architecture.

Main Results:

• PSF-D2NNs demonstrated enhanced robustness against misalignment in simulations and experiments.
• Simulated misalignment tolerance reached 26 pixels (MSE: 5.441×10⁻⁵, SSIM: 0.7741).
• Experimental misalignment tolerance exceeded 13 pixels (MSE: 3.350×10⁻⁶, SSIM: 0.9760).

Conclusions:

• The proposed PSF-D2NNs architecture significantly improves tolerance to multi-plane misalignment.
• PSF-D2NNs offer a viable solution for robust optical computing hardware.
• The findings pave the way for more reliable and practical diffractive deep neural network implementations.