Delocalized photonic deep learning on the internet's edge

Area of Science:

• Photonics
• Machine Learning
• Edge Computing

Background:

• Advanced machine learning models require significant power, processing, and memory, limiting their deployment on resource-constrained edge devices.
• Current edge devices cannot run complex AI models due to hardware limitations.

Purpose of the Study:

• Introduce Netcast, a new approach for machine learning inference using delocalized analog processing across networks.
• Enable ultra-efficient photonic inference on edge devices by streaming weight data from cloud-based smart transceivers.

Main Methods:

• Developed Netcast, a system utilizing cloud-based smart transceivers to stream weight data to edge devices.
• Implemented delocalized analog processing for machine learning inference.
• Conducted image recognition experiments using photonic inference.

Main Results:

• Achieved image recognition with ultralow optical energy (40 attojoules per multiply, <1 photon per multiply).
• Demonstrated high classification accuracy: 98.8% (93%) in lab and field trials.
• Reproducibly tested performance over 86 km of optical fiber with 3 THz bandwidth.

Conclusions:

• Netcast enables milliwatt-class edge devices to achieve teraFLOPS computing rates, previously exclusive to high-power cloud systems.
• This approach overcomes power, processing, and memory constraints for edge AI.
• Netcast paves the way for powerful AI applications on small, energy-efficient devices.