Broadband, Compact, and Training-Free Optical Processors for Parallel Image Classification

Area of Science:

• Photonics
• Optical Computing
• Artificial Intelligence

Background:

• Growing demand for faster, energy-efficient computing in AI.
• Limitations of current optical computing: bulky, wavelength-specific, complex training.
• Need for scalable and parallel optical processing solutions.

Purpose of the Study:

• Demonstrate a compact, training-free optical processor for parallel image classification.
• Overcome limitations of existing optical computing implementations.
• Explore broadband operation and multi-channel computation capabilities.

Main Methods:

• Utilized wavy diffractive structures (Fourier surfaces) for optical processing.
• Developed a compact device with a 40 × 40 μm² footprint.
• Implemented all-optical classification and post-processing linear matrix operations.

Main Results:

• Achieved all-optical classification accuracies of 76% (digits) and 59% (fashion items).
• Boosted accuracies to 84% (digits) and 66% (fashion items) with linear matrix operations.
• Demonstrated broadband operation with up to 6 simultaneous computations due to wavelength separation.

Conclusions:

• The Fourier surface processor offers a compact, efficient, and scalable solution for optical image classification.
• The passive system's ability to handle multiple wavelengths enables parallel computation.
• Future work includes inverse-designed extensions and on-chip photonic integration for enhanced performance.