Photonic Mixture-of-Experts for scalable multi-task on-chip optical neural networks.

Area of Science:

• Photonic computing
• Artificial Intelligence (AI)
• Optical hardware architectures

Background:

• Photonic computing presents an energy-efficient, high-bandwidth alternative for AI.
• Current photonic AI faces scalability issues due to depth-dependent designs, optical losses, and high reconfiguration costs.
• Existing linear optical structures limit efficient multi-task processing capabilities.

Purpose of the Study:

• To introduce a novel scaling paradigm for photonic AI processors that overcomes current limitations.
• To implement a scalable Photonic Mixture-of-Experts (PMoE) architecture for efficient multi-task workloads.
• To demonstrate a photonic AI approach that expands network width instead of depth.

Main Methods:

• Developed a Photonic Mixture-of-Experts (PMoE) architecture utilizing parallel photonic cores as expert networks.
• Implemented dynamic input routing to expert networks, enabling efficient multi-task execution without altering optical weights.
• Fabricated a PMoE chip with three collaborative diffraction-based expert networks and 18 parallel kernels.

Main Results:

• The fabricated PMoE chip achieved an average accuracy of 97.1% in multi-domain image classification.
• The architecture demonstrated efficient multi-task workload execution by dynamically routing inputs.
• The PMoE approach reduced digital parameter overhead by 67% compared to conventional optical networks.

Conclusions:

• The Photonic Mixture-of-Experts (PMoE) architecture offers a scalable and efficient solution for next-generation photonic AI processors.
• Expanding network width, rather than depth, effectively leverages photonic parallelism to overcome scalability bottlenecks.
• This work highlights the potential of PMoE for high-performance, large-scale optical AI hardware.