Fully forward mode training for optical neural networks.

Area of Science:

• Photonics
• Machine Learning
• Optical Computing

Background:

• Current machine learning training relies on digital computers, limiting speed and energy efficiency.
• In silico emulation poses significant constraints for complex optical computing models.

Purpose of the Study:

• To develop a novel method for efficient training of optical machine learning models.
• To implement the compute-intensive training process directly on the physical optical system.

Main Methods:

• Introduced fully forward mode (FFM) learning for optical systems.
• Experimentally demonstrated FFM learning in free-space and integrated photonics.
• Utilized optical neural networks with millions of parameters.

Main Results:

• Achieved state-of-the-art performance for optical neural networks of comparable size.
• Demonstrated all-optical focusing through scattering media with diffraction-limited resolution.
• Enabled parallel imaging of objects outside direct line of sight at kilohertz frame rates.
• Showcased high energy efficiency (5.40 × 10^18 ops/sec/watt) with low light intensity.
• Proved FFM learning can automatically search non-Hermitian exceptional points.

Conclusions:

• FFM learning significantly accelerates machine learning processes in optical systems.
• This method advances deep neural networks, ultrasensitive perception, and topological photonics.
• FFM learning overcomes numerical modeling constraints by performing computations on the physical system.