A training approach that penalizes the model for assigning high probability to incorrect or undesired outputs.