What is Bayesian deep learning?

Deep neural networks are the most flexible function approximators we have — and famously overconfident. Give one an input far from anything it trained on and it will still answer with a straight face. Bayesian deep learning fixes that by putting distributions where point estimates used to be: a distribution over weights, or over functions, so every prediction arrives with a measure of how much the model actually knows. The network keeps its flexibility; it just stops pretending to be sure when it is not.

That uncertainty is not a nice-to-have — it is the whole point wherever a wrong, confident answer costs more than a cautious one: science, medicine, self-driving, any decision taken on a model's word. This page collects the best Learning Bayesian Statistics resources on the topic: the flagship episode with Maurizio Filippone on why Bayesian deep learning is so powerful, more episodes on Bayesian neural networks, foundation models and diffusion models, and hands-on guides as they appear.

Featured episode

With Maurizio Filippone

#144 Why is Bayesian Deep Learning so Powerful, with Maurizio Filippone

What makes Bayesian deep learning powerful? Maurizio Filippone explains priors over functions, deep Gaussian processes, and why principled uncertainty changes what deep models are good for.

Listen to the episode →

More episodes on Bayesian deep learning

#161 Amortized Inference & Neural Processes, with Luigi Acerbi

Listen →

#158 Bayesian Workflows & Foundation Models, with Stefan Radev

Listen →

#157 Amortized Inference & BayesFlow in Practice, with Stefan Radev

Listen →

#151 Diffusion Models for SBI in Python, with Jonas Arruda

Listen →

#150 Fast Bayesian Deep Learning, with David Rügamer, Emanuel Sommer & Jakob Robnik

Listen →

#142 Bayesian Trees & Deep Learning for Optimization & Big Data, with Gabriel Stechschulte

Listen →

#129 Bayesian Deep Learning & AI for Science with Vincent Fortuin

Listen →

#107 Amortized Bayesian Inference with Deep Neural Networks, with Marvin Schmitt

Listen →

#21 Gaussian Processes, Bayesian Neural Nets & SIR Models, with Elizaveta Semenova

Listen →

Frequently asked questions

What is Bayesian deep learning?

Bayesian deep learning is deep learning that quantifies its own uncertainty. A standard neural network learns a single set of weights and makes point predictions — confident even when it should not be. A Bayesian treatment instead holds a distribution over weights (or, equivalently, over functions), so predictions come with honest uncertainty: the model can say "I don't know" on inputs far from its training data. It combines the flexibility of deep models with the calibrated reasoning of Bayesian statistics.

Why are Bayesian neural networks hard to train?

The posterior over a neural network's weights lives in millions of dimensions and has no closed form, so exact inference is out of reach. In practice you approximate it — variational inference, MCMC, the Laplace approximation, deep ensembles or MC dropout — and each approximation trades accuracy for compute. Two things make it genuinely hard: the approximations themselves are expensive at deep-learning scale, and priors over weights do not map intuitively onto priors over functions, so choosing a prior that means what you think it means takes care.

What is the difference between a Bayesian neural network, a deep ensemble and MC dropout?

They are three rungs on the same ladder — approximating a posterior over networks. A Bayesian neural network is the principled version: a true (approximate) posterior over weights, but the most expensive to fit. A deep ensemble trains several ordinary networks from different starting points and averages their predictions — simple, strong in practice, and a coarse approximation to a posterior. MC dropout keeps dropout switched on at test time and averages many stochastic forward passes — the cheapest of the three, and formally an approximate variational inference scheme. Ensembles and dropout are what most people actually run; full Bayesian neural networks are what they approximate.

Where should I start learning Bayesian deep learning?

Start with the featured episode above: Maurizio Filippone explains what makes Bayesian deep learning powerful — priors over functions, deep Gaussian processes, and why principled uncertainty changes what deep models are good for. Then try it yourself: fit a small Bayesian neural network (a Laplace approximation or a tiny variational model), compare its uncertainty against a plain network on held-out data, and check whether the confident predictions are the correct ones.

New to Bayesian modeling itself? Start with the Bayesian workflow hub, see how to value your data with the Bayesian experimental design hub, explore all episodes or dive into the blog for more applied Bayesian modeling.