Simulate a single gradient descent update from the current parameter, gradient and learning rate, and watch how the loss moves toward the minimum.
L2 正则化使权重向零收缩。