Bayesian-Neural-Networks

Why we included this project

Researchers and students learning how Bayesian uncertainty estimation works inside deep networks will find a lot to work with here. The repo implements the main approximate inference families in PyTorch: variational approaches such as Bayes by Backprop and the local reparameterisation trick, sampling methods like SGLD and Hamiltonian Monte Carlo, plus cheaper shortcuts such as MC dropout and the Kronecker-factored Laplace approximation. Since the same toy and UCI regression benchmarks and MNIST classification tasks are used across methods, you can compare how each one behaves without rebuilding the setup, and the Colab notebooks mean nothing needs to be installed locally. Teams that want calibrated confidence scores or out-of-distribution awareness in classifiers and regressors can treat these implementations as a starting point. The code stays close to the original papers, which makes it a genuinely useful way to work through the theory rather than just copying.