Exact Bayesian Layers 2026
New preprint: “Exact Bayesian layers via characteristic functions”, with Guillaume St-Onge, Laurent Hébert-Dufresne, and Jean-Gabriel Young. We introduce CF-Prop, which propagates characteristic functions to get the exact, sampling-free output distribution of a Bayesian ReLU layer, giving a closed-form ELBO for Gaussian, Poisson, and Gamma likelihoods; the resulting CF-VI has the lowest calibration error of any tested method on CIFAR-100 and captures heavy-tailed uncertainty that other methods underestimate.