In general, modelling the statistical structure of language does not give you a solid grasp on its meaning. If humans communicated in an efficient, compressed way, our messages would have no statistical structure at all, yet they would not be any less meaningful.https://twitter.com/fchollet/status/1106604524678995969 …
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Understanding of language, or anything close to it, will come only with learning it's use - teaching agents to communicate with each other in a meaningful way via human language.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Speculation: maybe we should build agents that use lang to serve itself rather than as service to human. I.e. use it comm with us, to achieve whatever objective we implicitly give it. It must be integrated with its other abilities.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I'm not sure I agree, but it's interesting. No stat structure = no repeated patterns. As all words/phrases can't be unique (hapax legomenon = no info transfer), it means no context sharing by speech => context is fully shared between parties. Ergo: it only works for monologues.
-
In other words, it seems that no statistical structure implies no language. One can have uncorrelated stream of data if context is fully predefined (eg. instrumental measurements), but not for language. Unless you literally mean obfuscating text by ZIPing it before saying it.
End of conversation
New conversation -
-
-
Poetry is a kingdom thou shall not pass the taste Internet is a virtual DOM you can throw some tests
#OkJeSorsThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
A reductionist can say that all a neural network is a nonlinear mapping learnt using Gradt descent, weight sharing & some regularization and lots and lots of examples; but when you see a machine translation learn to aptly use a word in unseen contexts, sthg else happening;
-
Generalization with sparse data is a kind of understanding even through you could underestimate it as just interpolation or extrapolation.
End of conversation
New conversation -
-
-
How else do you define understanding except in two parts: parsing a situation as composed of object parts and inferring hidden properties of those parts?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Conversely, in the absence of any statistical regularities, babies would never learn to speak.
-
reinfocement learning
- Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.