Quantifying Uncertainty in Neural Networks

This is the code used in the experiments described in the blog post Quantifying Uncertainty in Neural Networks.

To run the code, first download the CIFAR-10 and CIFAR-100 data sets. Extract them, and put them in a directory data as data/cifar-10-batches-py and data/cifar-100-python.

Configure the parameters in train.py (or leave them as the default) and create a directory nets for the learned weights to be written to. Run python inference.py to generate a directory creating the misclassifications from CIFAR-100. Finally, run python train.py to plot the image grids.

Those interested in further reading, should visit:

1. What My Deep Model Doesn't Know...

2. Bayesian Convolutional Neural Networks with Bernoulli Approximate Variational Inference