Representer Point Selection for Explaining Deep Neural Networks

Code release for Representer Point Selection for Explaining Deep Neural Networks at NeurIPS 2018

Instructions

Before running any code run the following script:

mkdir data output; cd influence-release-mod/scripts; ln -s ../../data data; ln -s ../../output output

This will create symbolic links to two directories data and output, which will be used by the influence function code inside the influence-release-mod folder.

data directory will store data files such as training/test data, training/test features, etc. that is used to compute the influence function values / representer values.

output directory will store the computed influence function values /representer values.

You can download the contents of the data and output directory used here

Both directory contains information used by the notebooks in experiments folder, which can be used to replicate the figures in the paper.

To calculate the representer values as in the paper, run

python compute_representer_vals.py --dataset Cifar
python compute_representer_vals.py --dataset AwA

in python 3 (for python 2, remove encoding = 'latin1' in load_data() )