Technical question I haven't dug into: How much of this predictive power has to do with facial recognition, and how much has to do with compression artifacts,

Technical question I haven't dug into: How much of this predictive power has to do with facial recognition, and how much has to do with compression artifacts, pixel value abberations, etc? Edit: the paper addresses this question: > One question is whether the network is using “high-level” cues, such as shape and symmetry, or simply relying on “low-level” cues, such as resampling artifacts. > although the network may be taking advantage of certain low-level cues, such as resampling, some performance is retained even when those are averaged away Seems that it's both... but I'm not sophisticated enough to read this paper deeply.