Abstract
Photonic neural networks (PNNs) based on silicon photonic integrated circuits (Si-PICs) offer significant advantages over microelectronic counterparts, including lower energy consumption, higher bandwidth, and faster computing speeds. However, the analog nature of optical signal in PNNs makes Si-PIC solutions highly sensitive to device noise, especially when using fixed-value deterministic models, which are not robust to hardware fluctuation. Furthermore, current PNNs are unable to handle data uncertainty, a critical factor in applications such as autonomous driving, medical diagnostics, and financial forecasting. Herein, a photonic Bayesian neural network (PBNN) architecture that incorporates Bayesian principles to enhance robustness and address uncertainty is proposed. In the PBNN, device noise is leveraged through photonic-noise-based random number generators, which combine Mach-Zehnder interferometers and micro-ring resonators to independently control output mean and standard deviation. Based on modelling with experimentally extracted data, the PBNN achieves a classification accuracy of up to 98% for handwritten digit recognition, matching full-precision models on conventional computers. Beyond classification, the PBNN excels in multimodal data processing, regression, and outlier detection. This scalable, energy-efficient architecture transforms photonic noise into computational value, addressing the limitations of deterministic PNNs and enabling uncertainty-aware computing for real-world applications.