Probabilistic Photonic Computing with Chaotic Light

Biological neural networks excel at solving complex computational problems and predicting outcomes from noisy, incomplete data. In contrast, conventional artificial neural networks often struggle with uncertainty estimation and are vulnerable to adversarial inputs outside the training distribution [1]. Bayesian neural networks (BNNs) offer a promising solution by using weight distributions obtained via Bayesian inference, rather than relying solely on point estimates [2]. However, evaluating BNNs requires probabilistic computing and sampling, which significantly degrades performance on traditional hardware.