The optimization techniques required to actually fit models are almost all powered by some form of gradient descent, and integration is usually required in truly probabilistic models to go from a density function to predictions.