Description
Artificial intelligence is increasingly deployed in safety-critical settings, from medical diagnosis to autonomous systems, where models must not only make accurate predictions but also recognize when their predictions are uncertain. However, conventional neural networks are largely deterministic and cannot reliably distinguish uncertainty caused by noisy or ambiguous inputs from uncertainty caused by inputs outside their training distribution. Bayesian neural networks provide a principled route to uncertainty-aware AI by treating model parameters as probability distributions, but their use is limited by the need for many stochastic forward passes and costly pseudo-random number generation on digital hardware.
In this talk, we present a photonic Bayesian machine that performs probabilistic neural-network inference by harnessing the intrinsic randomness of chaotic light. The system encodes stochastic weights in the spectrum of a broadband optical entropy source and computes probabilistic convolutions using dispersion-based photonic processing. Integrated with a 1.28 Tbit/s digital interface and a PyTorch-compatible workflow, the machine executes probabilistic convolutions in 37.5 ps, offloading the sampling-intensive operations of Bayesian neural networks to the optical domain.
We demonstrate the approach on image-classification tasks that require both prediction and uncertainty reasoning. In blood-cell microscopy, the system classifies known cell types while detecting erythroblast images excluded from training as out-of-domain. We further benchmark uncertainty disentanglement using MNIST, Ambiguous-MNIST, and Fashion-MNIST, showing that the system separates aleatoric uncertainty arising from ambiguous inputs from epistemic uncertainty arising from unfamiliar inputs. These results establish photonic hardware as a promising route toward low-latency, uncertainty-aware AI for real-time and safety-critical applications.
| I am the presenting author | Yes |
|---|