Skip to main content
. 2020 Feb 27;11(3):1633–1661. doi: 10.1364/BOE.386361

Fig. 4.

Fig. 4.

Key concepts of training a NN. A) Loss. The loss function (here MSE) is a metric for the deviation of the prediction of the NN from the (known) ground truth. B) Gradient descent. The loss is reduced with respect to variation of all trainable parameters of the NN. C) Backpropagation. Each trainable parameter is iteratively adjusted backwards through the NN.