A two-layer ReLU MLP with H hidden units trained by SGD on random samples from f. Watch the kinks move as both layers learn.