Training hybrid neural networks with multimode optical nonlinearities using digital twins.

Area of Science:

• Artificial Intelligence
• Optical Physics
• Computational Science

Background:

• The increasing size of neural networks demands significant energy and computational resources.
• Complex physical events can be integrated as efficient computation modules to reduce trainable layer complexity.

Purpose of the Study:

• To develop a hybrid artificial intelligence architecture that leverages physical systems for efficient computation.
• To reduce the energy and computational demands of training large neural networks.

Main Methods:

• Utilized ultrashort pulse propagation in multimode fibers for large-scale nonlinear transformations.
• Developed a neural model to differentiably approximate the optical system for training.
• Employed a training algorithm that backpropagates error signals through the optical proxy.

Main Results:

• Achieved state-of-the-art image classification accuracies.
• Demonstrated high simulation fidelity and exceptional resistance to experimental drifts.
• Successfully integrated low-energy physical systems into neural network architecture.

Conclusions:

• This hybrid approach enables scalable and energy-efficient AI models.
• Reduced computational demands for AI training by incorporating physical computation modules.
• Offers a pathway towards more sustainable and powerful artificial intelligence.