Central methodological contribution
Cluster stability
PathMNIST (histopathology)
Stability-vs-utility frontier
Stability-vs-utility scatter plot. PCA on pixels: stability 0.76, utility 0.16; Handcrafted color/texture: stability 0.91, utility 0.18; HOG: stability 0.42, utility 0.08; ResNet50 (ImageNet-supervised): stability 0.68, utility 0.24; DINO ViT-S/16: stability 0.72, utility 0.25; DINOv2 ViT-S/14: stability 0.73, utility 0.43.
Points in the upper-right are both stable and externally useful. A representation in the lower-right (stable but not useful) or upper-left (useful but unstable) is exactly the dissociation this project's central hypothesis (H3, see Methods) predicts may occur.
Seed-to-seed stability by algorithm
| Representation | Algorithm | k | Seeds | Mean pairwise ARI | Mean pairwise NMI |
|---|---|---|---|---|---|
| PCA on pixels | kmeans | 9 | 5 | 0.767 | 0.851 |
| PCA on pixels | gmm | 9 | 5 | 0.761 | 0.816 |
| Handcrafted color/texture | kmeans | 9 | 5 | 0.996 | 0.993 |
| Handcrafted color/texture | gmm | 9 | 5 | 0.817 | 0.874 |
| HOG | kmeans | 9 | 5 | 0.393 | 0.506 |
| HOG | gmm | 9 | 5 | 0.449 | 0.581 |
| ResNet50 (ImageNet-supervised) | kmeans | 9 | 5 | 0.747 | 0.812 |
| ResNet50 (ImageNet-supervised) | gmm | 9 | 5 | 0.606 | 0.714 |
| DINO ViT-S/16 | kmeans | 9 | 5 | 0.720 | 0.823 |
| DINO ViT-S/16 | gmm | 9 | 5 | 0.724 | 0.829 |
| DINOv2 ViT-S/14 | kmeans | 9 | 5 | 0.797 | 0.864 |
| DINOv2 ViT-S/14 | gmm | 9 | 5 | 0.659 | 0.791 |
Perturbation robustness (cluster agreement)
| Representation | brightness | contrast | crop | noise | blur | color_jitter |
|---|---|---|---|---|---|---|
| PCA on pixels | 0.39 | 1.00 | 0.99 | 1.00 | 1.00 | 1.00 |
| Handcrafted color/texture | 0.26 | 0.15 | 0.17 | 0.58 | 0.14 | 0.93 |
| HOG | 1.00 | 1.00 | 0.99 | 0.93 | 1.00 | 1.00 |
| ResNet50 (ImageNet-supervised) | 0.97 | 1.00 | 0.98 | 0.77 | 0.86 | 1.00 |
| DINO ViT-S/16 | 1.00 | 1.00 | 0.98 | 0.96 | 0.65 | 1.00 |
| DINOv2 ViT-S/14 | 0.99 | 1.00 | 0.98 | 0.97 | 0.89 | 1.00 |
Fraction of perturbed images assigned to the same cluster as their unperturbed original (1-NN proxy to the fitted k-means partition). Perturbations are modest, acquisition-plausible nuisance transforms only -- see Methods.
BloodMNIST (blood cell microscopy)
Stability-vs-utility frontier
Stability-vs-utility scatter plot. PCA on pixels: stability 0.83, utility 0.24; Handcrafted color/texture: stability 0.87, utility 0.10; HOG: stability 0.60, utility 0.20; ResNet50 (ImageNet-supervised): stability 0.64, utility 0.15; DINO ViT-S/16: stability 0.70, utility 0.30; DINOv2 ViT-S/14: stability 0.63, utility 0.18.
Points in the upper-right are both stable and externally useful. A representation in the lower-right (stable but not useful) or upper-left (useful but unstable) is exactly the dissociation this project's central hypothesis (H3, see Methods) predicts may occur.
Seed-to-seed stability by algorithm
| Representation | Algorithm | k | Seeds | Mean pairwise ARI | Mean pairwise NMI |
|---|---|---|---|---|---|
| PCA on pixels | kmeans | 8 | 5 | 0.927 | 0.926 |
| PCA on pixels | gmm | 8 | 5 | 0.735 | 0.810 |
| Handcrafted color/texture | kmeans | 8 | 5 | 0.841 | 0.890 |
| Handcrafted color/texture | gmm | 8 | 5 | 0.905 | 0.905 |
| HOG | kmeans | 8 | 5 | 0.573 | 0.674 |
| HOG | gmm | 8 | 5 | 0.620 | 0.686 |
| ResNet50 (ImageNet-supervised) | kmeans | 8 | 5 | 0.672 | 0.758 |
| ResNet50 (ImageNet-supervised) | gmm | 8 | 5 | 0.600 | 0.709 |
| DINO ViT-S/16 | kmeans | 8 | 5 | 0.684 | 0.750 |
| DINO ViT-S/16 | gmm | 8 | 5 | 0.712 | 0.786 |
| DINOv2 ViT-S/14 | kmeans | 8 | 5 | 0.640 | 0.708 |
| DINOv2 ViT-S/14 | gmm | 8 | 5 | 0.628 | 0.701 |
Perturbation robustness (cluster agreement)
| Representation | brightness | contrast | crop | noise | blur | color_jitter |
|---|---|---|---|---|---|---|
| PCA on pixels | 0.58 | 1.00 | 0.72 | 1.00 | 1.00 | 1.00 |
| Handcrafted color/texture | 0.15 | 0.03 | 0.06 | 0.36 | 0.04 | 0.99 |
| HOG | 1.00 | 1.00 | 0.99 | 1.00 | 1.00 | 1.00 |
| ResNet50 (ImageNet-supervised) | 0.94 | 1.00 | 0.94 | 0.76 | 0.42 | 1.00 |
| DINO ViT-S/16 | 0.99 | 1.00 | 0.98 | 0.93 | 0.69 | 1.00 |
| DINOv2 ViT-S/14 | 0.97 | 1.00 | 0.94 | 0.94 | 0.59 | 1.00 |
Fraction of perturbed images assigned to the same cluster as their unperturbed original (1-NN proxy to the fitted k-means partition). Perturbations are modest, acquisition-plausible nuisance transforms only -- see Methods.