Towards the Unification and Robustness of Perturbation and Gradient Based Explanations

Authors: Sushant Agarwal, Shahin Jabbari, Chirag Agarwal, Sohini Upadhyay, Steven Wu, Himabindu Lakkaraju

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we empirically validate our theory using extensive experimentation on both synthetic and real world datasets.
Researcher Affiliation Academia 1David R. Cheriton School of Computer Science, University of Waterloo, Waterloo, ON, Canada 2Department of Computer Science, Harvard University, Cambridge, MA, USA 3School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not explicitly state that source code for the methodology is provided or link to a repository.
Open Datasets Yes We generate a synthetic dataset and use 2 real world classification datasets from the UCI Machine Learning Repository (Dua & Graff, 2017).
Dataset Splits Yes We follow the standard 80/20 dataset split, i.e., 80% of the data was used for training the model while 20% was used for testing.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments (e.g., GPU/CPU models, memory).
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'ELU activation function' but does not specify version numbers for these or other software dependencies.
Experiment Setup Yes The models are trained using Adam optimizer using a cross-entropy loss function. Our best performing models achieve a testing accuracy of 99.50%, 96.30%, and 99.8% using 15, 60, and 100 training epochs for the Simulated, Bankruptcy, and Online Shopping datasets, respectively. We also train models using fewer than the aforementioned training epochs to assess the the impact of model accuracy on our equivalence and robustness guarantees. Consistent with our theory, for any input point x, for both C-LIME and Smooth Grad we generate perturbations from a local neighborhood of x by sampling points from N(x, σ2I). We study the effect of the number of perturbations and the value of σ2 in our experiments.