Interpretability for sequence generation models 🐛 🔍

Explainable AI in Julia.

Code for our ICML '19 paper: Neural Network Attributions: A Causal Perspective.

On Explaining Your Explanations of BERT: An Empirical Study with Sequence Classification

Attribution (or visual explanation) methods for understanding video classification networks. Demo codes for WACV2021 paper: Towards Visually Explaining Video Understanding Networks with Perturbation.

surrogate quantitative interpretability for deepnets

Easy-to-use MIRAGE code for faithful answer attribution in RAG applications. Paper: https://arxiv.org/abs/2406.13663

Code for the paper: Towards Better Understanding Attribution Methods. CVPR 2022.

Metrics for evaluating interpretability methods.

The source code for the journal paper: Spatio-Temporal Perturbations for Video Attribution, TCSVT-2021

Hacking SetFit so that it works with integrated gradients.

Source code for the GAtt method in "Revisiting Attention Weights as Interpretations of Message-Passing Neural Networks".

squid repository for manuscript analysis

Code for our AISTATS '22 paper: Improving Attribution Methods by Learning Submodular Functions.