I agree this paper is more like a tutorial, which I like, but that 73% accuracy is from their implementation of a standard (non-Bayesian) Alexnet, so no excuse

I agree this paper is more like a tutorial, which I like, but that 73% accuracy is from their implementation of a standard (non-Bayesian) Alexnet, so no excuse there.