LB for BNNs
This is the official GitHub repository for the paper "Fast Predictive Uncertainty For Classification Bayesian Deep Networks" by me (Marius Hobbhahn), Agustinus Kristiadi, and Philipp Hennig. The paper can be found on arxiv at: https://arxiv.org/abs/2003.01227. I have also written a blog post, which is an easy-to-understand summary of the paper.
All files are supposed to be stand-alone notebooks so you can look at different experiments independent of each other.
This is a 2D toy example. We show that the naiv version of the LB is overconfident in some cases but this is fixed by our correction. The corrected version shows comparable results to the sampled version while being much faster.
The OOD experiment from the main paper compares the mean of the Dirichlet from the Laplace Bridge (and the norm version) to the MC integral of 100 samples. To run all experiments use the python scripts. To look at individual results and play around, check out the notebooks.
If you want to play around with the different methods yourself, you can simply comment the call of the train() method back in and start running the notebook. The prior variances (called var0 in the code) of the Gaussians have been chosen such that their in-dataset MMC differs only around 5% from that of the MAP estimate but you can choose different ones as well if you wish.
These are all the timing experiments. Since one of the major selling points of the LB is that it is much faster, we want to investigate exactly how much faster it is.
The experiment on ImageNet is conducted here. If you don't have the training and test data for ImageNet yourself, you can't validate the experiment and, unfortunately, I am not allowed to share them for legal (copyright) reasons.
Applying the Laplace Bridge to other use cases
If you want to apply the Laplace Bridge to your own datasets this repository should include all necessary components. Applying a last-layer Laplace approximation to a network is very simple since most classifiers have a linear last layer. There are many references in the original paper showing that a last-layer Laplace approximation is nearly as good as a full layer one at the fraction of a cost. Just use the Exp2_MNIST or CIFAR10 notebooks to guide you.
I'm always looking for feedback. If you have problems understanding the content don't hesitate to contact me.
If you have any further questions or suggestions about this repo you can write me at marius.hobbhahn[at]gmail.com