Hyperparameter tuning of optical neural network classifiers for high-order Gaussian beams

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Optical Communications

Background:

• High-order Gaussian beams are crucial for advanced free-space optical communications.
• Diffractive deep neural networks (D2NNs) offer fast beam classification but require optimization.
• Hyperparameter tuning, like interlayer distance, is critical for D2NN performance.

Purpose of the Study:

• To classify Hermite-Gaussian beams using a D2NN.
• To automatically tune the interlayer distance hyperparameter of the D2NN.
• To enhance the accuracy of D2NN-based beam classification.

Main Methods:

• Utilized a D2NN for classifying Hermite-Gaussian beams.
• Employed the tree-structured Parzen estimator for hyperparameter auto-tuning.
• Investigated the impact of interlayer distance optimization on classification accuracy.

Main Results:

• Achieved a classification accuracy of 98.8% for 16-mode Hermite-Gaussian beams, an improvement from 98.3%.
• Significantly enhanced classification accuracy for 36-mode beams from 84.9% to 94.9%.
• Demonstrated that accuracy gains from auto-tuning increase with a higher number of classification modes.

Conclusions:

• Automatic tuning of D2NN interlayer distance is effective for improving high-order Gaussian beam classification.
• The proposed method shows promise for robust optical communication systems.
• Further research can explore auto-tuning for other D2NN hyperparameters and beam types.