Optical micro-phase-shift dropvolume in a diffractive deep neural network.

Area of Science:

• Optics
• Artificial Intelligence
• Deep Learning

Background:

• Active modulation diffractive deep neural networks require robust inference through parallel subnetworks.
• Existing methods may involve complex mathematical derivations for phase-only modulation masks.

Purpose of the Study:

• To introduce a novel encoding strategy for enhancing inference in active modulation diffractive deep neural networks.
• To develop a method that avoids complex mathematical derivations while maintaining network nonlinearity.

Main Methods:

• Embedding a random micro-phase-shift dropvolume with five-layer statistically independent dropconnect arrays into unitary backpropagation.
• Implementing a drop-block strategy for structured-phase patterns to create a macro-micro phase dropvolume.
• Utilizing macro-phase dropconnects with fringe griddles encapsulating sparse micro-phase patterns.

Main Results:

• Numerical validation confirms the effectiveness of the macro-micro phase encoding strategy.
• The proposed method successfully integrates structured-phase encoding within the dropvolume.
• The approach maintains the nonlinear nested characteristics of neural networks.

Conclusions:

• Macro-micro phase encoding is an effective strategy for diffractive deep neural networks.
• The developed method offers a simplified approach to achieving robust inference.
• This technique provides a pathway for structured-phase encoding within dropvolumes.