.png){kind=icon}
Return
Efficient hardware error correction with hybrid on-offline configuration algorithm for optical processor
.png){kind=icon}
OA
AI
DOI：10.1038/s42005-025-02247-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Photonic neural networks (PNNs) have emerged as a promising platform for high-speed, parallel, and low-latency computing by harnessing the linear propagation of optical signals. However, scaling up PNNs faces significant challenges due to hardware errors caused by fabrication variations and environmental factors. Traditional approaches, such as offline error correction and online training, either rely on complex control systems or suffer from local optima convergence issues, resulting in limited scalability and efficiency. Here, we propose a hybrid on-offline configuration (HOOC) algorithm for programmable optical processors. This innovative approach combines offline initial value presetting with online perturbed optimization iteration algorithm, enabling precise and highly efficient error correction. We benchmark the algorithm’s performance in complex-valued matrix configuration and classification tasks, demonstrating robust error correction capabilities, including high reconstruction fidelity (≥98%), rapid convergence (≤10 iterations), and reduced dependence on detection devices. Furthermore, numerical simulations of high-order coherent processors demonstrate that our HOOC algorithm effectively avoids local optima, a common limitation of the conventional in-situ training method, thus simultaneously improving the scalability and robustness. These results underscore the viability and efficiency of the HOOC algorithm for scalable and robust PNN implementations, paving the way for scalable optical computing in artificial intelligence applications. Scaling up photonic neural networks for AI hardware is hindered by manufacturing and environmental errors. This work introduces a hybrid algorithm that combines offline calibration with online optimization, achieving highly efficient hardware error correction with rapid convergence while avoiding common local optima issues.
Journal
IF:
5.8
Papers:
2.8K
Citations:
9.2K
Organization
Cited Papers
ShareSaveShareSaveA Compact Butterfly-Style Silicon Photonic-Electronic Neural Chip for Hardware-Efficient Deep Learning.png)
ACS PHOTONICSOAAI
ACS PHOTONICSIF6.7
OAAIShareSaveShareSaveShareSaveShareSaveThermal flux manipulation on the silicon photonic chip to suppress the thermal crosstalkAPL PHOTONICSOAAI
APL PHOTONICSIF5.3
OAAIShareSaveShareSave
View more