Blending Optimal Control and Biologically Plausible Learning for Noise-Robust Physical Neural Networks.

Area of Science:

• Computational neuroscience
• Physical systems for AI
• Neuromorphic computing

Background:

• Artificial intelligence (AI) demands are increasing computational needs.
• Physical neural networks (PNNs) leverage physical processes for efficient neuromorphic computing.
• Training PNNs is currently computationally expensive.

Purpose of the Study:

• To develop a cost-effective training approach for PNNs.
• To reduce the computational expense associated with training PNN weight parameters.
• To enable wider practical application of physical systems as PNNs.

Main Methods:

• A novel training approach merging optimal control for continuous-time dynamical systems.
• Integration with a biologically plausible training method: direct feedback alignment.
• Numerical and experimental verification in an optoelectronic delay system.

Main Results:

• Substantial reduction in the computational cost of training PNNs.
• Achieved robust information processing despite measurement errors and noise.
• Demonstrated effectiveness without requiring detailed system information.
• Extended the range of physical systems applicable as PNNs.

Conclusions:

• The proposed training method significantly lowers the barrier to PNN implementation.
• This approach enhances the practicality and robustness of physical neuromorphic computing.
• It paves the way for more efficient and versatile AI hardware.