The blue dots (noise ε) never change. Moving mu or sigma slides and stretches the green samples z = mu + sigma·ε smoothly.
Each dot is one single-sample estimate of the gradient with respect to mu at mu = 0 for the loss E[(z − 3)²]. The dashed line is the true gradient.
Dark samples have high loss. Each grey arrow is one sample's push on mu, away from itself and harder if its loss is high. The green arrow is their average.