On Generating Plausible Counterfactual and Semi-Factual Explanations for Deep Learning

Authors: Eoin M. Kenny, Mark T Keane11575-11585

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Two controlled experiments compare this method to others in the literature, showing that PIECE generates highly plausible counterfactuals (and the best semi-factuals) on several benchmark measures.
Researcher Affiliation Academia Eoin M. Kenny*, Mark T. Keane University College Dublin, Dublin, Ireland Insight Centre for Data Analytics, UCD, Dublin, Ireland Vista Milk SFI Research Centre eoin.kenny@insight-centre.org, mark.keane@ucd.ie
Pseudocode Yes Algorithm 1: Modify exceptional features in x to produce x
Open Source Code Yes The code needed to reproduce our algorithm may be found at https://github.com/Eoin Kenny/AAAI-2021.
Open Datasets Yes Fig. 3 illustrates some of PIECE s plausible contrastive explanations for a CNN s classifications on the MNIST dataset (Le Cun, Cortes, and Burges 2010)
Dataset Splits No For MNIST, a test-set of 163 images classified by the CNN was used which divided into: (i) correct classifications (N=60) with six examples per number-class, (ii) close-correct classifications (N=62), that had an output Soft Max probability < 0.8, where the CNN just got the classification right,3 and (iii) incorrect classifications (N=41) by the CNN (i.e., every instance misclassified by the CNN). For CIFAR-10, the test-set was divided into: (i) correct classifications (N=30) with three examples per class, and (ii) incorrect classifications (N=30) with three examples per class. All instances were randomly selected, with the exception of MNIST s incorrect classifications, which were not randomly selected as there was only 41 of them.
Hardware Specification No The paper mentions 'utilizing a GPU' as a general possibility to reduce optimization time but does not provide specific hardware details (e.g., GPU models, CPU models, or memory specifications) used for running the experiments.
Software Dependencies No The paper does not explicitly provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup No Hyperparameter choices are presented in Section S2 of the supplementary material.