Comparison

Representation comparison

External validity, label-free clustering geometry, stability, linear-probe accuracy, and retrieval quality for every representation, on every dataset. No composite score: read across columns, not down a single ranking.

PathMNIST (histopathology)

Representation	Feature dim	Selected k	External ARI [95% CI]	External NMI	Silhouette	Seed stability (mean ARI)	Linear probe acc.	Retrieval P@5
PCA on pixels baseline	50	4	0.159 [0.139, 0.184]	0.384	0.352	0.764	0.484	0.557
Handcrafted color/texture baseline	56	10	0.184 [0.159, 0.231]	0.288	0.573	0.907	0.668	0.288
HOG baseline	1296	5	0.085 [0.074, 0.105]	0.157	0.014	0.421	0.245	0.203
ResNet50 (ImageNet-supervised) generic-pretrained	2048	4	0.239 [0.211, 0.277]	0.419	0.129	0.676	0.722	0.780
DINO ViT-S/16 self-supervised	384	4	0.251 [0.223, 0.287]	0.456	0.160	0.722	0.824	0.892
DINOv2 ViT-S/14 self-supervised	384	6	0.435 [0.405, 0.473]	0.607	0.126	0.728	0.825	0.852

External ARI/NMI use the label-free-selected k (chosen via silhouette on the validation split, never on test labels). Seed stability is the mean pairwise ARI across k-means seeds at k = number of annotated classes (a labeled reference condition, used only to fix k for the stability axis).

BloodMNIST (blood cell microscopy)

Representation	Feature dim	Selected k	External ARI [95% CI]	External NMI	Silhouette	Seed stability (mean ARI)	Linear probe acc.	Retrieval P@5
PCA on pixels baseline	50	4	0.242 [0.220, 0.281]	0.390	0.137	0.831	0.705	0.520
Handcrafted color/texture baseline	56	8	0.103 [0.091, 0.125]	0.224	0.526	0.873	0.873	0.258
HOG baseline	1296	4	0.200 [0.177, 0.233]	0.315	0.033	0.596	0.597	0.369
ResNet50 (ImageNet-supervised) generic-pretrained	2048	4	0.146 [0.126, 0.177]	0.253	0.095	0.636	0.818	0.544
DINO ViT-S/16 self-supervised	384	6	0.299 [0.269, 0.342]	0.425	0.089	0.698	0.903	0.678
DINOv2 ViT-S/14 self-supervised	384	7	0.178 [0.156, 0.211]	0.245	0.061	0.634	0.893	0.566