Survey: Exploiting Data Redundancy for Optimization of Deep Learning