Look at the Variance! Efficient Black-box Explanations with Sobol-based Sensitivity Analysis
Authors: Thomas FEL, Remi Cadene, Mathieu Chalvidal, Matthieu Cord, David Vigouroux, Thomas Serre
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We run extensive experiments to demonstrate the benefits of the proposed method on several recent complementary benchmarks including the Pointing Game and Deletion. |
| Researcher Affiliation | Collaboration | 1Carney Institute for Brain Science, Brown University, USA 2Sorbonne Université, CNRS, France 3Artificial and Natural Intelligence Toulouse Institute, Université de Toulouse, France 4 Institut de Recherche Technologique Saint-Exupery, France 5Valeo.ai |
| Pseudocode | Yes | Algorithm 1: Pythonic implementation of the Total Order indices ( ˆ STi) calculation. |
| Open Source Code | Yes | Our code is freely available: github.com/fel-thomas/ Sobol-Attribution-Method. |
| Open Datasets | Yes | For our vision experiments, we compared the plausibility of the explanations produced on the Pointing Game [36] benchmark. We evaluate the fidelity of our explanations using the Deletion metric for 4 representative models... trained on ILSVRC-2012 [41]. For our NLP experiments, we fine-tuned a Bert model and trained a bi-LSTM on the IMDB sentiment analysis dataset [42]... |
| Dataset Splits | Yes | We evaluate the fidelity of our explanations using the Deletion metric for 4 different pretrained models... on 2,000 images sampled from the Image Net validation set. |
| Hardware Specification | Yes | The reported execution time is an average over 100 runs on Res Net50 using an Nvidia Tesla P100 on Google Colab and a batch size of 64. |
| Software Dependencies | No | Tensor Flow [54] and the Keras [55] API were used to run the models. No version numbers are specified for these or for PyTorch/Torch Ray. |
| Experiment Setup | Yes | For the experiments involving images, the masks were generated at a resolution of d = 11 11 pixels, then upsampled with a nearest-neighbor interpolation method before being applied with the Inpainting perturbation function. Finally, N was set to 32 which is equivalent to 3, 936 forward passes... The reported execution time is an average over 100 runs on Res Net50 using an Nvidia Tesla P100 on Google Colab and a batch size of 64. |