1) as I mentioned elsewhere, the explicit coaching happens when you submit a photo that Keegan think is bad (say score < 5).
2) Mostly true, except that it's surprisingly good at capturing the notion of good timing, which I did not expect from ML ^^
3) well, as for the Mick Jagger picture, I can see why Keegan dislikes it: it's simply not that good technically. What makes it great is that it's Mick Jagger and that he has a very deep sight (imho), which the model indeed does not capture. For the other picture, yeah, the score is not that good, but the comment is spot on (in my humble and very biased opinion) :p
Thanks for the feedback, and yeah, definitely needs more love before it's a perfect predictor! Nicely enough, classification tasks with tons of supervision still make a lot of errors too ^^ (http://pbs.twimg.com/media/CdOxQRbWAAEUZM6.jpg)
1) It utterly fails as a coaching tool. "Your framing is 71% correct". Great. Awesome. That's maybe technically correct (the best kind!), but it's utterly unhelpful.
2) It fails at images taken to capture a mood instead of a motif. (Exactly what I'd expect from ML :)
3) It has problems with things that we usually consider iconic photos. Annie Leibovitz's picture of Mick Jagger in an elevator? Meh, 5.2. Dorothea Lange's "Migrant Mother" gets at least a 7.6, so on par with my vacation selfies.
It's a great idea. It needs a lot of love.