Ridge Regression: Shrinkage, Effective Degrees of Freedom and Leave-One-Out Cross-Validation for Correlated Features
Eight correlated features, three truly active. Slide the L2 penalty to watch coefficients shrink smoothly along the ridge path, read off the effective degrees of freedom from the eigenvalues of the Gram matrix, and compare leave-one-out error, generalised cross-validation and the exact test error to see where the best penalty sits. A Blender-built plot table with a hat-matrix trace check and an explicit leave-one-out refit check built in.
GDPR's "right to erasure" is easy to grant for a database row and hard to grant for a trained model — a naive fix means retraining from scratch every time someone asks to be forgotten. This simulator builds a real SISA-sharded ensemble of logistic-regression classifiers over a synthetic 2D dataset, each shard trained by genuine gradient descent on cross-entropy loss, and renders the ensemble's live decision surface as a colored 3D landscape. Click any training point to file an erasure request, then process it and watch only that point's shard retrain while every other shard — and its portion of the model — stays untouched, with a running counter comparing the true SISA retraining cost against what a naive full retrain would have cost.
Eight correlated features, three truly active. Slide the L2 penalty to watch coefficients shrink smoothly along the ridge path, read off the effective degrees of freedom from the eigenvalues of the Gram matrix, and compare leave-one-out error, generalised cross-validation and the exact test error to see where the best penalty sits. A Blender-built plot table with a hat-matrix trace check and an explicit leave-one-out refit check built in.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install