Coherent optical neural network chip with novel computing model for large-scale matrix-vector multiplication.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computer Engineering

Background:

• Optical neural networks (ONNs) offer potential for high-speed computation but face challenges in complexity and efficiency.
• Existing ONN designs often require complex phase compensation and encoding/decoding steps.

Purpose of the Study:

• To propose an innovative optical neural network (ONN) chip utilizing a coherent detection structure.
• To enhance computational efficiency and accuracy in ONNs by simplifying the operational process.

Main Methods:

• Developed a novel ONN chip architecture based on coherent detection.
• Established a mathematical model to integrate transformation functions into activation functions, eliminating phase compensation.
• Reduced encoding and decoding complexity.

Main Results:

• Achieved 97.28% accuracy on the MNIST dataset, comparable to conventional computer-based results.
• Demonstrated high computing efficiency and maintained accuracy.
• Showcased strong robustness and generalizability across various network architectures.

Conclusions:

• The proposed ONN chip design offers a significant advancement in optical computing.
• The integrated approach simplifies ONN operation while preserving high performance.
• This technology holds promise for efficient and accurate image classification and broader AI applications.