To train a neural network without a loss function.

To increase the number of target classes.

To remove all categorical features from a dataset.

To transform features to have approximately zero mean and unit variance.