Magnitude Potential · NeurIPS 2026 · interactive figure
Three overlapping Gaussian classes (1,050 training points, the paper's toy set). Move the scale t and the MP ratio is recomputed in your browser with leave-one-out weights, exactly as in the paper. Small t sees global structure; large t sees local detail.
Click any point to change its class and watch the ratio flag it.
Potential field
MP ratio of class 0 evaluated across the plane
Per-point MP ratio
Each training point, leave-one-out. Large and solid means atypical.
Distribution of MP ratios
All 1,050 training points at the current t
Magnitude at this scale
|X|t = Σi wi is the effective number of points. It runs from 1 (t → 0) to n (t → ∞).
Each point gets a potential from the magnitude weights of its own class and of the whole dataset. A ratio near 1 marks a typical class member. A ratio near 0 marks a point that sits in a dense region of the data but is sparse within its own class, such as boundary, overlapping or mislabelled points.
Similarities decay as e−t·d. At small t every point sees most of the dataset, so the ratio reflects coarse class layout. At large t each point sees only its nearest neighbours, so the ratio picks out local outliers and the potential field shrinks to small islands around the data.
The page factors S + 10⁻⁵I (Cholesky, float64) for the full set and for each class. Leave-one-out potential at a training point uses the identity p₋ᵢ(xᵢ) = 1 − wᵢ / (S⁻¹)ᵢᵢ, which matches the reference implementation without refitting n times.
Data: make_gaussians(k=3, d=2, n_per_class=500, spread=2, std=1, seed=0), 70% stratified training split. From “Generalization in Neural Networks Through the Lens of Magnitude Potential”, NeurIPS 2026.