Orbital angular momentum-mediated machine learning for high-accuracy mode-feature encoding.

Area of Science:

• Optoelectronics
• Machine Learning
• Information Optics

Background:

• Optical neural networks offer high speed and low energy consumption by utilizing optical dimensions.
• Exploiting orbital angular momentum (OAM) states for neural network nodes remains a challenge due to limitations in feature extraction.
• Laguerre-Gaussian (LG) beam modes offer potential for increased degrees of freedom in optical information processing.

Purpose of the Study:

• To demonstrate machine learning using orbital angular momentum (OAM) with an all-optical convolutional neural network (CNN).
• To leverage OAM states for enhanced feature extraction and classification in optical neural networks.
• To explore applications in secure communication and anomaly detection using OAM-based optical machine learning.

Main Methods:

• Development of an all-optical CNN utilizing Laguerre-Gaussian (LG) beam modes.
• Implementation of a trainable OAM mode-dispersion impulse as a convolutional kernel for feature extraction.
• Utilizing deep-learning diffractive layers for classification and employing energy weighting coefficients of encoded OAM modes for detection.

Main Results:

• Achieved 97.2% accuracy on the MNIST database through OAM mode-dispersion selectivity for information mode-feature encoding.
• Demonstrated resistance to eavesdropping in point-to-point free-space transmission.
• Realized all-optical dimension reduction for anomaly detection with 85% accuracy using multiplexed OAM states.

Conclusions:

• OAM-mediated machine learning provides a novel approach for optical information processing.
• The proposed all-optical CNN architecture effectively extracts features and classifies data using OAM states.
• This research opens avenues for improved machine vision tasks and secure optical communication systems.