Results Overview
Every result is measured on Galaxy10 DECaLS (17,736 images, 10 classes). Scores come from notebook outputs and are single runs; see Limitations before comparing two numbers.
83.9%
Best image model
DINOv2 fine-tuned · balanced accuracy
90.5%
Conformal coverage
α = 0.10 · fusion model
0.53
PC2 × Gini (ρ)
zero-shot DINOv2 features
−0.045
U-Net bulge fraction × g−r (ρ)
essentially uncorrelated
Provenance. Model scores in this section are copied from the output cells of the AstroVision notebooks (and, for DINOv2 per-class recall, read from an exported figure). The documentation data pack contains no metrics file for them, so they cannot be re-derived from the pack; the cell of origin is listed for each value in
_facts/astrovision/FACTS_INDEX.md §3 and shown when hovering over a number.| Section | Question | Main finding |
|---|---|---|
| Model Benchmarking | How do the five models compare? | Fine-tuned DINOv2 leads; fusion adds nothing measurable in the ablation |
| Uncertainty Quantification | Can confidence be trusted? | Conformal sets reach the target coverage; temperature scaling worsens ECE |
| Attention & Grad-CAM | Where does the model look? | Attention covers the galaxy after fine-tuning; attention masks separate classes better than Otsu masks |
| Physical Interpretation | Do features encode physics? | Moderate correlations with concentration and colour; no causal claim |
| Limitations | What do these results not show? | Pseudo-labels, split differences, possible feature leakage, ambiguous classes |
| Evaluation Protocol | How should results be reported? | Balanced accuracy as primary metric |
Related pages