Adam parameter update
Updates each parameter using a smoothed gradient divided by an adaptive scale derived from recent squared gradients.
Theta at step t equals theta at the previous step minus alpha times bias-corrected first moment m hat t divided by the square root of corrected second moment v hat t plus epsilon.
Move in the smoothed gradient direction, but take smaller coordinate steps where recent gradients have been large.
The equals sign defines an iterative update; the fraction is the adaptive coordinate-wise preconditioner.
This line is only the final Adam update; the paper's algorithm also defines how m, v, and their bias corrections evolve.