CoCoX: Generating Conceptual and Counterfactual Explanations via Fault-Lines

Authors: Arjun Akula, Shuai Wang, Song-Chun Zhu2594-2601

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive quantitative and qualitative experiments verify our hypotheses, showing that Co Co X significantly outperforms the state-of-the-art explainable AI models. We conducted extensive human subject experiments to quantitatively and qualitatively assess the effectiveness of the proposed fault-line explanations.
Researcher Affiliation Academia Arjun R. Akula,1 Shuai Wang,2 Song-Chun Zhu1 1UCLA Center for Vision, Cognition, Learning, and Autonomy 2University of Illinois at Chicago aakula@ucla.edu, shuaiwanghk@gmail.com, sczhu@stat.ucla.edu
Pseudocode Yes We outline our method in Algorithm 1. Algorithm 1: Generating Fault-Line Explanations
Open Source Code Yes Our implementation is available at https://github.com/arjunakula/Co Co X
Open Datasets Yes We used ILSVRC2012 dataset (Imagenet) (Russakovsky et al. 2015) and considered VGG-16 (Simonyan and Zisserman 2014) as the underlying network model.
Dataset Splits No The paper mentions '15 training images' and '5 test images' for the human subject study's familiarization and testing phases, but it does not provide explicit training, validation, or test dataset splits (e.g., percentages or sample counts) for the underlying CNN model (VGG-16) that was pre-trained.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments or train the models.
Software Dependencies No The paper mentions various techniques and models (e.g., VGG-16, Grad-CAM, TCAV, FISTA, K-means clustering) but does not list specific software packages or libraries with version numbers (e.g., Python, TensorFlow, PyTorch versions) that would be needed for replication.
Experiment Setup No The paper mentions that a 'pre-trained CNN (M) for image classification' and VGG-16 were used, but it does not specify hyperparameters (e.g., learning rate, batch size, number of epochs) or other system-level training settings for this model. The details provided are for the human subject study setup.