Guardado en:
| Autor principal: | |
|---|---|
| Formato: | Recurso digital |
| Lenguaje: | |
| Publicado: |
Zenodo
2026
|
| Materias: | |
| Acceso en línea: | https://doi.org/10.5281/zenodo.18147809 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
Tabla de Contenidos:
- <p>This paper presents a controlled empirical evaluation of three widely used learning paradigms—transfer learning, centralized model-agnostic meta-learning (MAML), and federated meta-learning—for few-shot photonic inverse design under structured physical domain shifts and privacy constraints.</p> <p>Using a large-scale synthetic dataset of 500,000 physics-consistent grating coupler simulations, the study constructs a non-IID learning scenario by partitioning data across federated clients based exclusively on a physically meaningful parameter (grating period). All methods are evaluated within a unified experimental framework that enforces identical neural architectures, optimization settings, data splits, and statistically stabilized evaluation protocols, enabling a fair and reproducible comparison.</p> <p>Results show that conventional transfer learning achieves the lowest absolute prediction error and strong zero-shot generalization across period-shifted regimes. In contrast, both centralized and federated MAML exhibit limited few-shot adaptation, with modest improvements as the number of support samples increases, but consistently underperforming the transfer baseline in absolute terms. Importantly, federated meta-learning achieves performance parity with centralized meta-learning across all evaluated settings, with no statistically significant degradation despite operating without access to raw client data.</p> <p>Rather than proposing new algorithms, this work focuses on clarifying the practical behavior and limitations of established few-shot and privacy-preserving learning paradigms in a physics-driven inverse regression setting. By explicitly reporting both positive and negative results under tightly controlled conditions, the study provides a reproducible benchmark for future research in photonic inverse design, scientific machine learning, and federated learning under physically structured data heterogeneity.</p>