Physical neural networks using sharpness-aware training

Area of Science:

• Artificial Intelligence
• Machine Learning
• Hardware Acceleration

Background:

• Traditional hardware limits AI advancements, driving interest in physical neural networks (PNNs).
• Training PNNs faces challenges like model-reality mismatch (in silico) and device specificity (in situ).
• Post-deployment perturbations (e.g., thermal drift) degrade PNN performance, necessitating retraining.

Purpose of the Study:

• To introduce and evaluate sharpness-aware training (SAT) for improving PNN training and robustness.
• To address the limitations of current in silico and in situ PNN training methods.
• To enhance PNN resilience to post-deployment perturbations without requiring retraining.

Main Methods:

• Leveraged sharpness-aware minimization principles to connect loss landscape sharpness with generalization.
• Established a link between loss landscape sharpness and physical system robustness.
• Applied SAT to both in silico and in situ PNN training paradigms.

Main Results:

• SAT effectively mitigates model-reality gaps in PNNs.
• The proposed training method enables cross-device transfer of PNN models.
• SAT demonstrates strong resilience to post-deployment perturbations, eliminating the need for retraining.
• Demonstrated SAT's broad applicability across three PNN platforms and diverse tasks (classification, compression, reconstruction, generation).

Conclusions:

• Sharpness-aware training (SAT) offers a robust and versatile solution for training physical neural networks.
• SAT enhances PNN generalization, cross-device transferability, and resilience to real-world perturbations.
• This approach significantly advances the practical deployment and reliability of AI hardware.