Normally I tend to agree with both you gentlemen ... butt! ... it’s not really like our kids have spend 15 to 20 years learning to drive ... it typically takes 6 - 12 month. Imo driving may be one of the least impressive arguments to use to make a valid point? 
this crystallizes what’s wrong with AI today: we have no techniques whatsoever to transfer a general, if hard-won understanding of the world to new tasks. we instead do almost everything from scratch, because our representations of what is learned are too superficial.https://twitter.com/benedictevans/status/1050441082809413632 …
-
-
-
for sure many things are harder.
End of conversation
New conversation -
-
-
Meta-learning aims to solve this. It is still far from it, but I think this is the most promising aspect of deep learning!
-
an aspect of deep learning, or a more general idea that you could conceivably apply to deep learning, perhaps hampered by deep learning’s inherent opacity?
- 1 more reply
New conversation -
-
-
Also you are correct this is a pain point for DL, but let's not pretend that DL people wouldn't agree with this. Of course transfer is hard and unsolved in DL. Of course we shouldn't have to train a system both to book airline tickets and learn all of English at the same time.
-
Arguably the only success stories of representation transfer in DL are word embeddings, and maybe also re-using the lower layers of a trained conv-net, but I think most people in the field or cognisant of these being shallow successes. Or I live in a particularly good DL bubble…
- 2 more replies
New conversation -
-
-
Yes, there's no way to do abstract learning. This is a known problem and everybody is without a clue. Level 3 maturity has yet to be achieved:https://medium.com/intuitionmachine/an-advanced-capability-maturity-level-for-artificial-general-intelligence-b300dafaca3f …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Except the story of the last few years is one of increasing transfer learning, with ImageNet pretraining in vision, and glove/word2vec and then ELMo in NLP, figuring out how to transfer larger pieces of knowledge to new tasks. I expect this trend to continue.
- 12 more replies
New conversation -
-
-
Things are changing : https://arxiv.org/abs/1810.04805 and multiple other recent work in nlp
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
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.