Physical neural networks using sharpness-aware training

Abstract

Abstract Recent advances in AI are pushing the limits of traditional hardware, making physical neural networks (PNNs) a promising alternative. However, training PNNs remains challenging: in silico training suffers from model-reality mismatch, while in situ training produces device-specific models that do not transfer across fabrication variations. Both approaches are further compromised by post-deployment perturbations, such as thermal drift or misalignment, which invalidate trained models and require retraining. We address these challenges through sharpness-aware training (SAT), inspired by sharpness-aware minimization, which links loss landscape geometry to generalization. We establish a connection between loss landscape sharpness and robustness in physical systems and leverage it to improve PNN training. SAT is compatible with both in silico and in situ settings: it mitigates model-reality gaps, enables cross-device transfer, and provides strong resilience to post-deployment perturbations without retraining. We demonstrate SAT across three PNN platforms and multiple tasks, including classification, compression, reconstruction, and generation, showing its broad applicability.