Optical neural network via loose neuron array and functional learning.

Area of Science:

• Deep learning
• Physical neural networks
• Optical computing

Background:

• Training non-differentiable hardware presents significant challenges in modeling, control, and calibration.
• Existing digital neural networks face power and bandwidth limitations.

Purpose of the Study:

• To introduce functional learning (FL), a novel deep-learning paradigm for training physical neural networks.
• To enable end-to-end training of non-differentiable and modeless physical neurons.
• To develop a programmable incoherent optical neural network.

Main Methods:

• Functional learning (FL) paradigm for training loose neuron arrays.
• Implicit gradient propagation for non-differentiable physical neurons.
• Numerical and physical verification using a light field neural network (LFNN).

Main Results:

• Demonstrated a programmable incoherent optical neural network.
• Achieved light-speed, high-bandwidth, and power-efficient neural network inference.
• Validated FL's capability for hardware without handcrafted design or precise assembly.

Conclusions:

• Functional learning provides a new methodology for hardware design, chip manufacturing, and system control.
• Optical neural networks offer a promising alternative to digital networks for specific applications.
• FL paves the way for brain-inspired optical computation and advanced optical devices.