How can a machine-learned model make predictions on data that was not part of the training data?

In general, a machine-learned model will exploit similarities between the training data and the out-of-training-set data to make predictions. As an example, suppose you are given a training set of height measurements for ten men and ten women. Given another height measurement for which you do not know the associated gender, you can probably still make a good guess at the gender of the person with that height. This is because the measurement will either be closer to those of either the men or the women in your original training set.