You're a model on a loss landscape. Gravity is the gradient — you roll downhill.
Tune the learning rate (↑↓): too high overshoots off the
overfitting cliffs; too low stalls in a local minimum. ←→
nudge to escape, then settle in the deepest valley — the global minimum — to converge.