Photonic unsupervised learning variational autoencoder for high-throughput and low-latency image transmission

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Data Transmission

Background:

• The exponential growth in global data necessitates high-throughput image transmission systems.
• Current electronic circuit-based methods face limitations in achieving desired transmission throughput.
• A need exists for advanced computational frameworks to enhance image transmission efficiency.

Purpose of the Study:

• To propose an end-to-end all-optical variational autoencoder for image transmission.
• To develop a photonic encoder-decoder (PED) for unsupervised optical computing.
• To integrate image compression, encryption, and error correction into an optical domain framework.

Main Methods:

• Developed an all-optical variational autoencoder (photonic encoder-decoder, PED).
• Modeled transmission noises as variations in an optical latent space.
• Implemented a large-scale, high-throughput unsupervised optical computing framework.

Main Results:

• Achieved computational latency reduction by over four orders of magnitude compared to state-of-the-art devices.
• Reduced transmission error ratio by 57% compared to on-off keying.
• Successfully integrated image compression, encryption, and error correction into the optical domain.

Conclusions:

• The photonic encoder-decoder (PED) offers a novel all-optical framework for image transmission.
• This approach significantly enhances computational speed and transmission accuracy.
• The work paves the way for AI-based physical systems and next-generation optical communications.