Revolutionizing Machine Learning with Photonic Orbital Angular Momentum-based Neural Networks

Thursday - February 22 2024, 09:51 UTC - 9 months ago

The development of a new ONNs architecture using OAM-mediated machine learning protocols has opened up the use of orbital angular momentum (OAM) in machine learning. This approach enables the learning of data features of images in the OAM domain, achieving highly precise intelligent encoding. This has the potential to revolutionize the technology industry with its power efficiency, parallelism, and minimal latency.

In 2023, a new product took the world by storm - ChatGPT, a large-scale language model based on machine learning. This technology, derived from artificial neural networks, surpassed 100 million users in record time, showcasing its immense popularity and potential. As an extension of this technology, optical neural networks (ONNs) have been gaining traction due to their efficiency compared to traditional electronics .

Utilizing different physical dimensions of light, such as space, wavelength, amplitude, and phase, ONNs are capable of completing tasks like translating text, image recognition, and natural language understanding through pre-training.However, one crucial dimension of light, orbital angular momentum (OAM), has yet to be fully utilized in ONNs. OAM states have been widely used in optical information processing systems such as optical communication, digital spiral imaging, and quantum communication, thanks to their infinite orthogonality .

Still, they have not been adopted in ONNs due to the inability to extract information features in the OAM domain.Thanks to a new paper published in Light: Science & Applications, this will soon change. Led by Professor Min Gu and Prof. Xinyuan Fang from the Institute of Photonic Chip at the University of Shanghai for Science and Technology, a team of scientists has developed a new ONNs architecture using OAM-mediated machine learning protocols .

This groundbreaking approach utilizes OAM as signals for the nodes of the neural network.The key to this new architecture lies in the diffraction-based convolutional neural network (CNN) composed of two main parts. The first part utilizes convolution between the OAM mode spectrum of an image and a trainable OAM mode-dispersion impulse to extract mode-features. The second part, a classification block, consists of cascaded finite-aperture trainable diffraction layers that compress the mode-features by controlling the wide OAM spectrum distribution .

This results in specific one or a few OAM states, effectively achieving mode-feature OAM encoding. Through multi-task learning, the mode dispersion pulse and phase distribution of diffraction layers are commonly trained to achieve the required output OAM modes.This new approach enables the learning of data features of images in the OAM domain, leading to highly precise intelligent encoding of images into specific OAM states .

This intelligent OAM encoding can then be used in various applications, such as image classification, secure free-space image transmission with high throughput and low latency, and optical anomaly detection.In conclusion, with the development of this new ONNs architecture, the use of OAM in machine learning is set to revolutionize the technology industry. With its power efficiency, parallelism, and minimal latency, ONNs using OAM-based neural networks will open up exciting possibilities for more efficient and intelligent machine learning .