Gary Marcus

You can't criticize Deep Learning for its lack of generalization, and at the same time argue that there are other more narrow AI methods that are more effective in special instances. You can't have your cake and eat it too!

Right. Nearly all the evidence we have for DL's superiority is based on a few, head-to-head comparisons on some narrow, toy supervised classification problems. The substantial diversity that used to exist in computer vision research has been replaced by a stifling DL monoculture.

Here is a nice point I will certainly use when asked about "biases in ML/AI/data"

or grabbing data that was annotated for a different purpose and not taking that into account at all

Correct. But this also biases researchers in considering annotation a straightforward and easy task, while it is often more tricky than they suspect. Discordance happens. https://arxiv.org/abs/1706.06838 

Really good point. With #Imagesnippets and #lio we can make very interesting semantic distinctions, but those terms are not the only vocabulary we can use, while all the data is stored with the image as subject of named graphs in triples with an end point.