Fast Axiomatic Attribution for Neural Networks

Authors: Robin Hesse, Simone Schaub-Meyer, Stefan Roth

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

Reproducibility Variable Result LLM Response
Research Type Experimental Various experiments demonstrate the advantages of X-DNNs, beating state-of-the-art generic attribution methods on regular DNNs for training with attribution priors.
Researcher Affiliation Academia Robin Hesse1 Simone Schaub-Meyer1 Stefan Roth1,2 1Department of Computer Science, TU Darmstadt 2hessian.AI
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code and additional resources at https://visinf.github.io/fast-axiomatic-attribution/.
Open Datasets Yes For our experiments on models for image classification... we use the Image Net [24] dataset, containing about 1.2 million images of 1000 different categories. [...] To that end, we employ the public NHANES I survey data [17] of the CDC of the United States, containing 118 one-hot encoded medical attributes, e.g., age, sex, and vital sign measurements, from 13,000 human subjects.
Dataset Splits Yes We train on the training split and report numbers for the validation split. [...] we randomly subsample 200 training and validation datasets containing 100 data points from the original dataset.
Hardware Specification No The paper mentions 'Using a single GPU' but does not specify the model or other detailed hardware specifications.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or library versions).
Experiment Setup Yes For our experiments on models for image classification... we use the Image Net [24] dataset [...]. If not indicated otherwise, we assume numerical convergence for Integrated Gradients and Expected Gradients, which we found to occur after 128 approximation steps (see supplemental material). [...] A simple MLP with Re LU activations is used as the model. [...] we randomly subsample 200 training and validation datasets containing 100 data points from the original dataset.