AttEXplore: Attribution for Explanation with model parameters eXploration

Authors: Zhiyu Zhu, Huaming Chen, Jiayu Zhang, Xinyi Wang, Zhibo Jin, Jason Xue, Flora D. Salim

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Large-scale experiments demonstrate that our Attribution method for Explanation with model parameter e Xploration (Att EXplore) outperforms other state-of-the-art interpretability methods. Moreover, by employing other transferable attack techniques, Att EXplore can explore potential variations in attribution outcomes. Our code is available at: https://github.com/LMBTough/ATTEXPLORE.
Researcher Affiliation Collaboration Zhiyu Zhu1, Huaming Chen1 , Jiayu Zhang2, Xinyi Wang3, Zhibo Jin1, Jason Xue4 & Flora D. Salim5 University of Sydney1, Su Zhou Yierqi2, Universiti Malaya3, CSIRO s Data614, University of New South Wales5
Pseudocode No The paper describes its methods mathematically and textually but does not include structured pseudocode or an algorithm block.
Open Source Code Yes Our code is available at: https://github.com/LMBTough/ATTEXPLORE.
Open Datasets Yes In this study, we employ Image Net dataset (Deng et al., 2009). We conduct experiments on a selection of 1000 samples from Image Net, guided by the principles outlined in NAA (Zhang et al., 2022), SSA (Long et al., 2022), and AGI (Pan et al., 2021).
Dataset Splits No The paper mentions using the ImageNet dataset and selecting 1000 samples but does not specify explicit training, validation, or test dataset splits, percentages, or counts.
Hardware Specification Yes All experiments are conducted using an AMD Ryzen Threadripper PRO 5955WX 16Core CPU, NVIDIA RTX6000 Ada GPU, and Ubuntu 22.04.
Software Dependencies No The paper mentions "Ubuntu 22.04" for the operating system, but does not provide specific version numbers for other ancillary software components like programming languages (e.g., Python), libraries (e.g., PyTorch, TensorFlow), or solvers.
Experiment Setup Yes Additionally, we apply the following general parameters setting: momentum set to 1.0, mask control parameter ρ set to 0.5, number of approximate features N set to 20, standard deviation of Gaussian noise (σ) set to 16, perturbation rate (ϵ) set to 48/255, and total attack iterations (num steps) set to 10.