Explainability of Point Cloud Neural Networks Using SMILE

Abstract

This study explores the implementation of SMILE for Point Cloud offering enhanced robustness and interpretability, particularly when Anderson-Darling distance is used. The approach demonstrates superior performance in terms of fidelity loss, R² scores, and robustness across various kernel widths, perturbation numbers, and clustering configurations.

Additionally, a stability analysis using the Jaccard index establishes a benchmark for model stability in point cloud classification, identifying dataset biases crucial for safety-critical applications like autonomous driving.

BibTeX

      @article{aslansefat2024pointcloud,
        title={Explainability of Point Cloud Neural Networks Using SMILE: Statistical Model-Agnostic Interpretability with Local Explanations},
        author={Aslansefat, Koorosh and others},
        journal={IEEE Software},
        year={2024},
        publisher={IEEE}
      }