Mah-NCM from six images through shared feature geometry to one prediction
For this illustration, 2 labeled real and 3 labeled generated support vectors determine two class means and one pooled covariance. The resulting tied ellipses and linear boundary classify the selected query vector as Q → Generated.
The Same Six Images
Labeled Support · Fit
Ground Truth:
Real
Ground Truth:
Generated
Unlabeled Query · Classify
Q
Only the selected image is classified
1
Frozen Encoder
Weights Fixed
φ
Feature vector
z ∈ ℝᵈ
1
Fit geometry from five supports
Feature Dimension 1
Feature Dimension 2
+ Real Mean μ₀
+ Generated Mean μ₁
Same Ellipse Shape = One Pooled Covariance Σₚ
Q
Q → Generated
2
Place the query, compare Mahalanobis scores
2
Prediction
Q → Generated
Images
Shared Feature Space
Decisions