Reconfigurable matrix multiplier with on-site reinforcement learning

Area of Science:

• Photonics and Artificial Intelligence
• Metasurface Optics
• Reconfigurable Computing

Background:

• Matrix multiplication is crucial for AI and information processing.
• Photonic matrix multipliers offer low energy and high speed but lack reconfigurability.
• Conventional designs use fixed Fourier optics and are difficult to generalize.

Purpose of the Study:

• To introduce a novel reconfigurable photonic matrix multiplier.
• To enable on-site customization of matrix multiplication functionalities.
• To address the limitations of fixed-function photonic multipliers.

Main Methods:

• Utilizing transmissive metasurfaces with varactor diodes as tunable dielectrics.
• Applying effective medium theory for dielectric tuning.
• Employing on-site reinforcement learning to drive reconfigurability.

Main Results:

• Demonstrated the viability of tunable dielectrics in metasurfaces.
• Validated the performance of matrix customization.
• Showcased a reconfigurable photonic matrix multiplier.

Conclusions:

• This work presents a new pathway for reconfigurable photonic matrix multipliers.
• The developed device offers on-site adaptability for diverse applications.
• It overcomes the limitations of static photonic computing components.