A standard neural network uses fixed weights and biases: its predicted class probability can express class ambiguity, but not uncertainty about those parameter values. A Bayesian neural network instead represents weights and biases with probability distributions. A prior describes plausible parameter values before training; the data update those distributions through Bayes’ theorem. section 11.4.3 section 11.4.3
The posterior over parameters combines the prior P(θ) with the likelihood P(D | θ), normalized by P(D). section 11.4.3
A pedestrian prediction
A fixed-parameter network’s 0.68 probability that a pedestrian is present describes its prediction under those parameter values; it does not say how certain the network is that those values are correct. section 11.4.3

