Adaptive Data Dropout: Towards Self-Regulated Learning in Deep Neural Networks