Robustness criterion for diffractive deep neural networks using normalized cutoff frequency

Area of Science:

• Optics and photonics
• Artificial intelligence
• Machine learning

Background:

• The real-world deployment of diffractive deep neural networks (D²NNs) is hindered by an insufficient understanding of their robustness.
• Evaluating D²NN robustness is critical for practical applications in various fields.

Purpose of the Study:

• To propose a novel criterion for evaluating D²NN robustness.
• To establish a physically interpretable and widely applicable principle for D²NN robustness assessment.

Main Methods:

• Utilizing normalized cutoff frequency (NCF) as a criterion to evaluate D²NN robustness.
• Analyzing the spatial frequency propagation characteristics of the light field within D²NNs.
• Demonstrating the principle across different D²NN architectures for classification and regression tasks.

Main Results:

• D²NN robustness is determined by the spatial frequency propagation characteristics of the light field.
• Equal NCF values correlate with identical robustness in D²NNs, irrespective of architecture.
• The discovered relationship enables efficient and accurate robustness prediction for large-scale D²NN models.

Conclusions:

• The normalized cutoff frequency (NCF) serves as a reliable metric for D²NN robustness.
• The findings are applicable to both linear and nonlinear D²NN models.
• This research facilitates the development of more robust D²NNs for real-world applications.