In which I show that Bayesian deep learning is nothing special: priors on the weights, explained with pictures and PyMC3 code.
Deep learning is, at its core, matrix multiplications with weights learned by gradient descent. Bayesian deep learning is what you get when you put priors on those weights. That's the whole mystery! In this PyData NYC 2017 talk, I build the intuition with mostly pictures and PyMC3/Theano code, with as little math as possible, and show how to construct Bayesian neural networks and visualize the uncertainty in their predictions.
Slides are available here.