Hacker News Reader

Yeah, two nearly identical versions (probably only resolution differences) of our logo resulted in 21% & 94% similarity scores

it's definitely not perfect. I think the embedding-vector approach of image search isn't scale invariant.

Forgive me if this is a naive question, as I am not well versed in ML.

Would it be possible to train using Mechanical Turk so that you get real human feedback and perception of quality?

Considering all the bad logos in the world that humans make, use and like - I wouldn't bet on it.