To train a model to the random variation in the sample data. Overfit models contain too many parameters (weights), and they do not generalize well.