to randomize the trajectory of the neural network parameters during training

to smooth the gradient of the loss function in order to avoid getting trapped in small local minimas

to scale the gradient descent step in proportion to the gradient magnitude

to compute the gradient of the loss function with respect to the neural network parameters