EMaP: Explainable AI with Manifold-based Perturbations
Authors: Minh Nhat Vu, Huy Quang Mai, My T. Thai
JMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that EMa P not only improves the explainers performance but also helps them overcome a recently developed attack against perturbation-based explanation methods. Keywords: black-box explanations, explainability, topological data analysis, adversarial robustness |
| Researcher Affiliation | Academia | Minh Nhat Vu EMAIL Theoretical Division, Los Alamos National Laboratory, Los Alamos, NM, 87545, USA... My T. Thai EMAIL Department of Computer and Information Science and Engineering University of Florida Gainesville, FL 32611, USA |
| Pseudocode | Yes | Algorithm 1 EMa P Input: Data to explain x0, a subset of training data (X, y), number of pivots per labels p, number of perturbations per pivot k, lower dimension V and noise level r. Output: Xr and Dr. Xr contains k(pl + 1) orthogonal perturbations locally around x0 and points in X. Dr contains the low-dimensional distances of points in Xr to x0 (l is the number of unique labels in y). 1: Initialized an EMa P sampler object M. 2: M.mapper Mapper to the manifold of dimension V of X 3: M.pivots 4: Include x0 to M.pivots 5: for each class l in y do 6: Include p samples of class l to M.pivots 9: for each data point x in M.pivots do 10: X, D M.gen perturbation(x, k, r) 11: Include X to Xr and include D to Dr 12: end for 13: return Xr and Dr. |
| Open Source Code | Yes | Our source code is attached in the supplementary material of this submission. |
| Open Datasets | Yes | The tabular data sets are the COMPAS (JeffLarson and Angwin, 2016), German Credit (Hofmann, 1994), and Communities and Crime (Redmond, 2011). The image data sets are MNIST (Le Cun and Cortes, 2010) and Fashion-MNIST (Xiao et al., 2017). |
| Dataset Splits | No | The paper refers to datasets and mentions 'test set accuracy' for image datasets, implying the existence of a test set, but it does not provide specific details on how the datasets were split into training, validation, or test sets (e.g., percentages, sample counts, or explicit methodology). It mentions following the experimental setup of previous work, but not the specific splits used in this paper. |
| Hardware Specification | Yes | Our experiments are conducted on a single GPU-assisted compute node that is installed with a Linux 64-bit operating system. The allocated resources include 32 CPU cores (AMD EPYC 7742 model) with 2 threads per core and 100GB of RAM. The node is also equipped with 8 GPUs (NVIDIA DGX A100 Super Pod model), with 80GB of memory per GPU. |
| Software Dependencies | No | The paper mentions using 'Pytorch (Paszke et al., 2019)' for convolutional networks and 'the Ripser Python library (Tralie et al., 2018)' for Bottleneck distances, and 'UMAP (Mc Innes et al., 2018)' for manifold learning. However, it does not specify version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | The noise vector used to generate perturbations for all applicable explainers has a radius 10^-3 for text data and 10^-4 for image data, which are in the range analyzed in the previous Bottleneck distance experiments in Fig. 6 and 7. The noise radius r_T used to approximate the local subspaces of EMa P (Algorithm 3) is chosen equal to the noise radius r for perturbation. For UMAP s hyper-parameters, we use n_components {2, 3} and min_dist = 0.1 (default value). Our source code is attached in the supplementary material of this submission. Finally, for a fair comparison, the number of perturbations used to generate the explanation of any reported methods is 1000. |