[11] Jacob Andreas - Learning from Language
The Thesis Review
00:00
The Meta Learning of Language Learning
If you can learn to associate tasks with their descriptions then that gives you a much nicer hypothesis space for figuring out what to do next. You now have literally a language for describing decision rules that's sort of structured and that you can put priors on in easy ways without actually requiring you to pre-commit to some more highly structured hypothesis class like logicals or decision treesor something like that. And so the I mean the real thesis in that part of the thesis is that language learning is a kind of meta learning.
Transcript
Play full episode
