Gaussian Process Regression: RBF Length-Scale, Noise and Posterior Uncertainty
Fit a Gaussian process to a handful of noisy points and see exactly what it knows and does not know. Move the kernel length-scale, signal amplitude, assumed noise and number of observations, and watch the posterior mean, the two-sigma uncertainty band, random functions drawn from the posterior, the marginal likelihood and the exact leave-one-out error change. A Blender-built plot easel with data pegs, plus numerical checks of the Cholesky, determinant, leave-one-out and sampling identities.
Supervised learning splits into two jobs: classification (predict a category) and regression (predict a number). This simulator focuses on the regression side of k-nearest neighbors — instead of voting on a class, it averages the target values of the k closest labeled points in a 3D feature space to predict a continuous value at any query location. A translucent wireframe traces the true underlying function that generated the noisy training data, so you can watch the k-NN prediction surface hug it tightly at low k and flatten out as k grows, with live readouts for the predicted value, the ground truth, the pointwise error, and the RMSE across the whole domain.
Fit a Gaussian process to a handful of noisy points and see exactly what it knows and does not know. Move the kernel length-scale, signal amplitude, assumed noise and number of observations, and watch the posterior mean, the two-sigma uncertainty band, random functions drawn from the posterior, the marginal likelihood and the exact leave-one-out error change. A Blender-built plot easel with data pegs, plus numerical checks of the Cholesky, determinant, leave-one-out and sampling identities.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install