Less is More: An Exploration of Data Redundancy with Active Dataset Subsampling