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

Area of Science:

• Optics
• Machine Learning
• Deep Neural Networks

Background:

• Unitary neural networks are foundational for active modulation diffractive deep neural networks.
• Optical implementation is key for manipulating complex network connections.

Purpose of the Study:

• To implement an optical random phase DropConnect for diffractive deep neural networks.
• To investigate the impact of micro-phase drilling on training convergence and statistical inference.
• To compare the performance of full-drilling versus part-drilling phenomena.

Main Methods:

• An optical random phase DropConnect was implemented on an optical weight.
• Micro-phase was integrated into Bernoulli holes to facilitate training convergence.
• Geometrical phase ray deflections were constantly reformulated during training epochs.
• The optical system utilized a projective imaging setup with random micro-phase-shift acting as a sparse griddle.

Main Results:

• Random micro-phase-shift part-drilling demonstrated superior performance over full-drilling.
• Enhanced performance was observed in both the training and inference phases.
• Part-drilling allows for a greater number of geometrical ray deflection recombinations.

Conclusions:

• Optical random phase DropConnect is effective for diffractive deep neural networks.
• Part-drilling strategies with micro-phase shifts improve network performance.
• The method enhances statistical inference capabilities through controlled optical manipulation.