Gary Marcus

completely agree: remarkable ≠ semantically coherent. system has no idea that a unicorn has one horn, blithely asserts it has four, etc. simultaneously close and not close to Searle’s room, not at all close to genuine understanding.https://twitter.com/Grady_Booch/status/1096841495519232000 …

Seems like you're moving the yardstick here. It's coherent wrt entities often over whole paragraphs. Certainly that falls into the camp of things "deep learning shouldn't do". Obviously it's not solving the common sense problem, but it's hard to argue that it's not a step forward

if i did same with n-grams where n > 50, yielding similar local coherence, would you insist that I count that as progress? we may not have had the data or compute in 1960s, but we have long known that local quasi-coherence can be captured that way - without real understanding

N-gram models of the size needed to capture this range of structure would overfit to an absurd degree. At what point does "local conference" make the jump to global?

What makes you think there isn’t overfitting to a high degree? The pastiche-y nature of the output seems like that to me. Some of the coherence violations are actually local (unicorns can’t have four horns), and I don’t see this approach as on a track to genuine coherence.

They report the n-gram overlap with the training set, so it's at least not overfitting in the trivial sense. To be clear, I'm happy with just syntactic coherence, and agree it won't learn "real" semantics. But I view that as a problem with next step prediction rather than DL.

but you can get syntactic coherence with n-grams, if you are indifferent to semantics.

As a non-linguist, my jargon was probably wrong. I meant that these examples show tracking of novel entities over long time horizons, but are bad at inferring their real-world properties. N-grams can't even do the former, and many argued that neither could neural networks.