Photonic network-in-network with on-chip optical diffraction to enable high-order tensor inputs

Key Points

• The architecture achieves a 55.16% reduction in backend parameters, enhancing efficiency without compromising performance.
• Competitive results in machine vision benchmarks indicate the effectiveness of the proposed optical neural network.
• Integration of on-chip diffraction facilitates advanced tensor fusion and channel-wise pooling for improved data handling.
• This novel approach supports high-order tensor inputs, paving the way for more efficient photonic computing.

Abstract