Probabilistic photonic computing

We present a novel photonic architecture that leverages broadband chaotic-light as an entropy source together with an incoherent photonic crossbar array to enable high-speed probabilistic computation. By encoding input means and variances in the optical domain and performing parallel wavelength-division-multiplexed sampling, we execute probabilistic convolutions and embed a Bayesian neural network capable of both classification and uncertainty estimation. Our prototype processes at an effective sampling rate of tens of gigasamples per second and demonstrates high accuracy on the MNIST task while reliably detecting out-of-distribution inputs. This work paves the way for ultrafast hardware-native probabilistic inferencing.