Fully Optical Integrated Mixture-of-Experts System

A fully optically integrated Mixture-of-Experts (MoE) system is introduced to address the explosive growth in computational power demands in the development of artificial intelligence technologies. Here, we highlights the potential of optical computing in solving communication challenges for MoE models and proposes a fully optically integrated MoE system. This system integrates on-chip subwavelength diffraction units (SWU) as expert models and uses Mach-Zehnder interferometers (MZI) as gating and tuning modules to achieve system integration. The system not only solves the technical issue of on-chip training difficulty in existing optical diffraction computing models but also overcomes the performance limitations of optical interference computing models. By combining the advantages of both, it enhances the performance of optical computing models and the scalability of optical systems. Finally, the article validates the feasibility of the system in visual tasks through simulation experiments and demonstrates the simulation results of signal input and output using MZI, proving the feasibility of the scheme.