/uncertainty-attribution

Scripts to reproduce results within the following manuscript: Perez, I., Skalski, P., Barns-Graham, A., Wong, J. and Sutton, D. (2022) Attribution of Predictive Uncertainties in Classification Models, 38th Conference on Uncertainty in Artificial Intelligence (UAI), Eindhoven, Netherlands, 2022.

Primary LanguagePythonMIT LicenseMIT

Uncertainty attributions

This is the official tensorflow implementation of the following manuscript:

Perez, I., Skalski, P., Barns-Graham, A., Wong, J. and Sutton, D. (2022) Attribution of Predictive Uncertainties in Classification Models, 38th Conference on Uncertainty in Artificial Intelligence (UAI), Eindhoven, Netherlands, 2022.

OpenReview link available here.

Contact: iker [dot] perez [at] featurespace [dot] co [dot] uk

Running the code

For the scripts to work correctly, you need to install the uncertainty_library and necessary dependencies. We recommend to do that by running

$ pip install .

in a virtual environment with python 3.8.

The scripts folder contains python scripts to train the models and reproduce results for the MNIST, FashionMNIST and CelebA datasets.