In a groundbreaking development in the field of machine learning, scientists from the Institute of Photonic Chip at the University of Shanghai for Science and Technology have pioneered a new architecture for optical neural networks (ONNs) using orbital angular momentum (OAM)-mediated machine learning protocol. This innovative approach harnesses the unique dimension of light—OAM—to enhance the information capacity and efficiency of machine learning processes.
The team, led by Professor Min Gu and Prof. Xinyuan Fang, has successfully integrated OAM signals into the nodes of a neural network, enabling the high-precision encoding of image data into specific OAM states. By learning data features in the OAM domain, the researchers have unlocked a host of applications, including image classification, secure free-space image transmission, and optical anomaly detection.
The core of this new ONNs architecture lies in a diffraction-based convolutional neural network (CNN) comprising two key components:
1. **Convolution Part:** Utilizes convolution to extract mode-features by densifying the input OAM mode combs through the use of a trainable OAM mode-dispersion impulse.
2. **Classification Block:** Consists of trainable diffraction layers that compress mode-features, controlling the output of specific OAM states.
By training the OAM mode-dispersion impulse and phase distribution of diffraction layers, the CNN can effectively extract OAM features from input data, leading to significant dimensionality reduction and enhanced predictive accuracy.
The practical applications of this OAM-mediated machine learning technique are far-reaching. The team achieved impressive results in a classification task involving handwritten digits, with a classification accuracy of 96%. Additionally, they successfully encoded images for wireless optical communication, achieving an accuracy of 93.3% in transmitting information using OAM states.
Furthermore, the OAM encoding method demonstrated exceptional anti-eavesdropping capabilities, providing a secure means of communication. By integrating OAM multiplexing holograms as decoding systems, the researchers showcased an end-to-end switchable image display, highlighting the potential for all-optical information encoding, transmission, and display.
Looking ahead, this innovative approach to intelligent OAM encoding holds promise for a wide array of applications, offering a new avenue for machine learning tasks in the optical domain. The seamless integration of OAM states with machine learning protocols paves the way for high-capacity, high-security optical neural networks, revolutionizing machine vision tasks and unlocking new possibilities for the future of information processing.
