Training large-scale optoelectronic neural networks with dual-neuron optical-artificial learning

Area of Science:

• Artificial Intelligence
• Optoelectronics
• Computational Science

Background:

• Optoelectronic neural networks (ONNs) offer potential for high-performance AI computing.
• Diffractive neural networks are promising but face training challenges due to computational costs.
• Modeling optical diffraction in large networks is computationally intensive.

Purpose of the Study:

• To introduce DANTE, a dual-neuron architecture for efficient training of large-scale ONNs.
• To overcome the computational and memory limitations of traditional diffractive network training.
• To enhance convergence speed and scalability in optical artificial learning.

Main Methods:

• Developed DANTE, a hybrid architecture with optical neurons for diffraction and artificial neurons for approximation.
• Implemented iterative global artificial-learning and local optical-learning steps for improved convergence.
• Utilized simulation experiments on ImageNet and CIFAR-10 datasets.

Main Results:

• Successfully trained large-scale ONNs with 150 million neurons on ImageNet, a previously unattainable scale.
• Achieved significant acceleration in training speeds on the CIFAR-10 benchmark compared to single-neuron methods.
• Demonstrated a two-layer physical ONN system based on DANTE for effective feature extraction.

Conclusions:

• DANTE offers a scalable and efficient solution for training large-scale optoelectronic neural networks.
• The dual-neuron approach effectively balances optical diffraction modeling with computational approximation.
• Physical implementation shows DANTE's capability in real-world image classification tasks.