Bias-Variance Decomposition of Polynomial Regression: Underfitting, Overfitting and Test Error (2D)
Polynomial fits to repeated noisy training sets, with the exact bias and variance curves and Blender-rendered desk props. The same shared model as the 3D board.
Fit polynomials of any degree to noisy samples of a sine and see the expected test error split exactly into squared bias, variance and irreducible noise. Grey curves are earlier training sets, the yellow curve is the mean fit. A Blender-built data board with six independent numerical checks (Monte Carlo, hat-matrix trace, orthogonality, interpolation).
Polynomial fits to repeated noisy training sets, with the exact bias and variance curves and Blender-rendered desk props. The same shared model as the 3D board.
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install