Shows f(x) (true function) and multiple learned f_hat(x) curves from different training sets. At a fixed input x, it visualizes bias as the gap between f(x) and E[f_hat(x)], variance as the spread of f_hat(x) across datasets, and a stacked bar for noise + bias^2 + variance. Model complexity animates over time to demonstrate underfitting (high bias) to overfitting (high variance).
Deterministically simulates multiple training sets with noisy samples around a fixed true function. A complexity parameter blends a low-variance linear fit with a high-variance kernel smoother; predictions at a fixed x are aggregated to estimate bias and variance. Uses snapped drawing, green-on-black palette, and a 4.2s ping-pong animation cycle.