Galaxy Morphology XAI
Systematic evaluation of post-hoc explainability across four CNN architectures
Role
Sole researcher
Timeline
Jan 2026 — May 2026
Year
2026
Category
Research · AI/ML
Survey notes
Deep learning classifiers are now standard tools for automated galaxy morphology analysis at the scale of next-generation astronomical surveys like Rubin LSST. However, black-box neural networks present a critical trust problem for scientific use — astronomers cannot verify whether predictions rely on genuine morphological features or spurious correlations. This project conducts a systematic evaluation of four post-hoc explainability methods (Grad-CAM, LIME, Integrated Gradients, GradientSHAP) across four CNN architectures on two galaxy morphology datasets, using quantitative faithfulness metrics to show that no single explanation method dominates universally.
Notable terrain
- 01Systematic evaluation of Grad-CAM, LIME, Integrated Gradients, and GradientSHAP across ResNet-18, VGG-16, EfficientNet-B0, and Custom CNN
- 02Quantitative faithfulness metrics: Deletion AUC, Insertion AUC, Consistency, and Sparsity
- 03Cross-dataset stress test on Galaxy Zoo Evo to validate explainability generalization
- 04Statistical significance via Wilcoxon signed-rank and McNemar tests
- 05Per-class analysis revealing Featured galaxies receive 33% lower Deletion AUC than Smooth galaxies
Next sheet
Lattice