Techniques that reduce noise in gradient estimates to improve optimization efficiency and convergence speed.