CALIMOCHO

An implementation of Explanatory Active Learning (XAL) based on Self-explainable Neural Networks.

See:

Stefano Teso - Toward Faithful Explanatory Interactive Machine Learning with Self-explainable Neural Nets, Proceedings of the 3rd International Tutorial & Workshop on Interactive and Adaptive Learning (IAL'19).
Stefano Teso and Kristian Kersting - Explanatory Interactive Machine Learning, International Conference on AI, Ethics and Society, 2019 (pdf).

Dataset

Our preliminary experiments use the synthetic colors dataset from:

Andrew Ross, Michael C. Hughes, Finale Doshi-Velez - Right for the right reasons: Training differentiable models by constraining their explanations

The original data can be found on the rrr repo. We used the preprocessed dataset from the caipi repo.

Experiments

To run CALIMOCHO, use the main.py script. Type python main.py --help for the list of options.

To run the experiments on the colors dataset:

Download the toy_colors.npz file from the caipi repository linked above and place it into the data directory
Execute colors.sh in the shell

The code will save all results in the results directory in pickle format, and plot them in PNG format too.

Plots

To draw the final plots, unzip the zipped results files and run:

python draw.py lime-colors0-simplearch results-colors-lime/results/colors0__passive\=True__n\=None__k\=5__p\=0.2__T\=100__W\=101__P\=__e\=0.01__L\=0.9\,0.0__E\=1000__B\=None__s\=0__limer\=*.pickle -s q1 -m 10 11 12 13
python draw.py active-shallow-colors0-margin -s q2 results-colors-active-partialz/results/colors0__strategy\=margin__passive\=False__n\=None__k\=5__p\=0.0001__c\=*__T\=300__W\=101__P\=__e\=0.01__L\=0.1\,0.0__E\=100__B\=None__s\=0__trace.pickle results-colors-active-partialz/results/colors0__strategy\=margin__passive\=False__n\=None__k\=5__p\=0.0001__c\=1__T\=300__W\=101__P\=__e\=0.01__L\=0.0\,0.0__E\=100__B\=None__s\=0__trace.pickle
python draw.py active-deeper-colors0 -s q3 results-colors-active-partialz/results/colors0__strategy\=random__passive\=False__n\=None__k\=5__p\=0.0001__c\=1__T\=300__W\=*__P\=__e\=0.01__L\={0.0,0.1},*__E\=100__B\=None__s\=0__trace.pickle

Requirements

python >= 3.5
sklearn >= 0.21.0
tensorflow >= 1.13.1
lime >= 0.1.1

Older versions may also work.

Acknowledgements

This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. [694980] SYNTH: Synthesising Inductive Data Models).