We visualized the result of evaluating various transfer functions on the same (non-cherry-picked) YouTube video. Below you can select any target task and any source task (or a set of source tasks in higher order transfers) to see how well the transfer works. You can compare that to transfering from ImageNet based FC7 features, a fully supervised network, or not employing any transfer learning. The results are insightful about what tasks transfer to each other and generally the space of vision tasks in terms of transfer learning.