A learning problem where the training data has very few error examples compared to correct ones, making optimization difficult.