Active learning in open experimental environments: selecting the right information channel(s) based on predictability in deep kernel learning.