Scaling up for end-to-end on-chip photonic neural network inference

Area of Science:

• Photonics and Artificial Intelligence
• Integrated Optics
• Neuromorphic Computing

Background:

• Optical neural networks offer advantages in bandwidth and energy efficiency over electronic counterparts.
• Scaling challenges include limited network depth due to weak activation function cascadability and constrained input size by optical matrix scale.

Purpose of the Study:

• To propose a novel scaling strategy for on-chip optical neural networks.
• To enable greater network depth and larger input sizes for enhanced optical inference capabilities.

Main Methods:

• Introduced Partially Coherent Deep Optical Neural Networks (PDONNs) with an opto-electro-opto nonlinear activation function for positive net gain.
• Implemented convolutional layers for rapid dimensionality reduction to increase input size accommodation.
• Utilized partially coherent optical sources to reduce reliance on specialized lasers and coherent detection.

Main Results:

• Fabricated a monolithically integrated optical neural network with the largest input size (64) and deepest network depth to date.
• Achieved 96% accuracy for fashion image classification and 94% for handwritten digit classification.
• Demonstrated maintained performance with partially coherent illumination, highlighting robustness and accessibility.

Conclusions:

• PDONNs represent a significant advancement in scalable optical neural network design.
• The proposed architecture facilitates energy-efficient, widely accessible optical computing.
• This work paves the way for practical, large-scale optical inference systems.