1. If more data means more noise, then less quality data is better
2. Overfitting with the noisy pattern
3. Good feature selection can make less data better
4. The task or data distribution is simple enough to learn from less data

References