Advancing Post-Hoc Case-Based Explanation with Feature Highlighting

Authors: Eoin M. Kenny, Eoin Delaney, Mark T. Keane

IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Results demonstrate that the proposed approach appropriately calibrates a user s feelings of correctness for ambiguous classifications in real world data on the Image Net dataset, an effect which does not happen when just showing the explanation without feature highlighting.
Researcher Affiliation Academia Eoin M. Kenny1 , Eoin Delaney2,4 and Mark T. Keane2,3,4 1CSAIL, Massachusetts Institute of Technology 2University College Dublin 3Insight Centre for Data Analytics 4Vista Milk SFI Research Centre
Pseudocode Yes Algorithm 1 Latent-Based Require: f(.); CNN to-be-explained Require: I; Test Image Require: D; Training Dataset Require: m(.); Activation map algorithm (e.g., FAM) ... Algorithm 2 Superpixel-Based Require: f(.); ANN to-be-explained Require: I; Test Image Require: D; Training Dataset Require: S(.); Superpixel Algorithm ...
Open Source Code Yes 1Code available at https://github.com/EoinKenny/IJCAI-2023
Open Datasets Yes Two datasets CUB-200 [Welinder et al., 2010] and Image Net [Deng et al., 2009] were used
Dataset Splits Yes Tests used the first 500 validation images. ... The 24 misclassifications were randomly divided into two material sets (A-set and B-set) to counterbalance the experiment;
Hardware Specification Yes Gathering the data for Fig. 3 took two months on two Nvidia v100 GPUs.
Software Dependencies No The paper mentions software like 'ResNet34', 'ResNet50', 'LIME', 'CAM', 'FAMs' but does not provide specific version numbers for these or other software dependencies required for reproduction.
Experiment Setup Yes The purpose of this experiment is to isolate the best hyperparameter values for α and β in equations 3 and 4, respectively. ... For each hyperparamter value, the networks were finetuned for 2500 iterations and test-accuracy sampled every 50... For hyperparameters, latent-based CAM should use α=5... Superpixel segmentation of 30 and β= in superpixels is recommended as it generalizes best.