Optical random phase dropout in a diffractive deep neural network.

Area of Science:

• Optics
• Machine Learning
• Deep Neural Networks

Background:

• Unitary learning, a backpropagation method, updates weights in complex-valued neural networks.
• Current methods face overfitting and poor generalization due to small-sample training with unitary weights.

Purpose of the Study:

• To address the overfitting and generalization issues in unitary learning.
• To introduce a novel method for enhancing the performance of diffractive deep neural networks.

Main Methods:

• Formulated and designed optical random phase dropout.
• Developed a synthetic mask by leveraging the equivalence between unitary forward and diffractive networks.
• Integrated computational modulation with a random sampling comb (dropout) filled with Bernoulli-distributed random phases.

Main Results:

• The proposed dropout method effectively alleviates overfitting in unitary learning.
• Massively parallel full connections with varied optical links during training enhance generalization.
• The introduction of a conjugated random phase comb signifies the importance of optical backpropagation.

Conclusions:

• Optical random phase dropout is a viable technique to improve generalization in unitary learning.
• This method offers a promising direction for developing more robust and efficient diffractive deep neural networks.
• The study highlights the significance of optical backpropagation in advanced neural network architectures.