Time-synthetic optical neural networks with stable programmable gain.

Area of Science:

• Photonics
• Artificial Intelligence
• Optical Computing

Background:

• Optical neural networks (ONNs) promise high speed and energy efficiency for AI.
• Passive optical components limit ONN depth due to signal loss and noise accumulation.
• Integrating optical gain is challenging due to instability from feedback and reflections.

Purpose of the Study:

• To overcome the depth limitations of passive optical neural networks.
• To develop a stable method for incorporating optical gain into ONNs.
• To enable deeper and more powerful photonic intelligence architectures.

Main Methods:

• Developed a time-synthetic optical neural network architecture.
• Integrated programmable optical gain into the temporal evolution of the network.
• Utilized a causal topology to suppress feedback and parasitic reflections.
• Conducted numerical simulations and in-situ experiments for validation.

Main Results:

• Achieved stable loss compensation by integrating programmable gain.
• Significantly extended the usable depth of the optical neural network.
• Demonstrated robust performance on image classification tasks.
• Validated the stability and scalability of the gain-assisted time-synthetic ONN approach.

Conclusions:

• Gain-assisted time-synthetic ONNs provide a stable and programmable pathway for deep photonic intelligence.
• This approach overcomes fundamental limitations of passive optical neural network architectures.
• The developed method enables scalable and robust optical artificial intelligence beyond current constraints.