Initializing with our Models Genesis, the annotation cost can be reduced by 30%, 50%, 57%, 84%, and 44% for target tasks NCC, NCS, ECC, LCS, and BMS, respectively. With decreasing amounts of labeled data, Models Genesis (red) retain a much higher performance on all five target tasks, whereas learning from scratch (grey) fails to generalize. Note that the horizontal red and gray lines refer to the performances that can eventually be achieved by Models Genesis and learning from scratch, respectively, when using the entire dataset.