But then I tried a very bad logo ("Kate's Florist" from http://www.paulmurraydesign.com/graphic-design/how-to-spot-a...) and it was ranked almost as good.
[looks at about page]
Okay, no, it's nothing like that at all. The logo gave me the wrong first impression of the company.
There's more about how they use deep learning here:
but for me it seems the constraints they've chosen are ill-conceived (things like matching logo strength to font weight) and unlikely to come up with something inspiring or even decent a lot of the time or rank things well which humans would have no difficulty recognising. This is a difficult problem, and I wouldn't expect rules based solutions to work very well, though some kind of permutation engine starting from known good designs could work better.
I guess it is a fair assumption that the neural network is trained using a corpus of logos that already look "well made" to the human eye.
It is possible there may not have been any/many samples that would be considered bad.
It would be interesting to see how this neural net evaluates a bad logo, using a tool such as LIME.