The relevance to this list is sketchy, it's true, but here is a pointer to the work of Stanford researchers on distinguishing gay people using their facial images with a deep neural net and a large image file - actually noticeable is that it is not THAT huge; 35000 portraits of men & women, adverting to their sexual preferences.I notice the data were first compressed X8 which is not necessarily in the spirit of deep learning in general.This application lifts recognition from 57% to 70 or 80 % which sounds even more surprising if it is expressed as the difference from 50/50.Unaided view 7/50 better than chance: AI 20/50 to 30/50 better than chance. https://psyarxiv.com/hv28a/