A technique to reduce noise and instability in gradient estimates, especially during early training phases.