Probabilistic photonic computing based on chaotic Kerr optical frequency combs.

Abstract

Artificial neural networks (ANNs) have become powerful tools for deciphering complex data patterns. However, conventional ANNs can be viewed as point estimators that fail to capture predictive uncertainty, leading to overconfident predictions. Probabilistic models grounded in Bayesian inference offer a principled framework for characterizing uncertainty, yet they are generally ill-suited for conventional deterministic computing architectures. Here, we propose and experimentally demonstrate a probabilistic photonic computing link based on chaotic Kerr optical frequency combs (CK-PPC). The intrinsic randomness of chaotic Kerr optical frequency combs is exploited as a native entropy source, and a cascade of Mach-Zehnder modulators with a micro-ring resonator (MRR) array is employed to realize the probabilistic photonic computing link. Building upon this platform, a hybrid Bayesian convolutional network (HBNN) is constructed and applied to blood-cell microscopic image classification and out-of-distribution (OOD) detection. Compared with conventional convolutional neural networks, our approach improves OOD detection by 23% while maintaining high recognition accuracy. This scheme paves the way toward integrable, high-speed probabilistic optical processors for trustworthy artificial intelligence in real-world scenarios.