“…) A recurring theme in unsupervised learning is the use of self-(or meta-) supervision (Pathak et al, 2016;Larsson et al, 2016;Zhang et al, 2016;Doersch et al, 2015;Gao et al, 2016;Misra et al, 2016;Wang & Gupta, 2015). This refers to a network trained for a pretext (or proxy) task, which is not of direct interest, but significantly relates to the final high-level task, e.g., object detection, classification, and action recognition (Girshick, 2015;Simonyan & Zisserman, 2014;Sun et al, 2017;Gkioxari et al, 2015). Automatic image colorization (Larsson et al, 2016;Zhang et al, 2016) is a typical example of a pretext task; naturally colorizing grey images requires prior knowledge of natural image appearance.…”