that's a post about how people think, not about how to optimize a reward function
-
-
Replying to @bobpoekert @alicemazzy
We don’t know how to craft a reward function other than in terms of a conceptual vocabulary.
2 replies 0 retweets 0 likes -
Replying to @Meaningness @alicemazzy
nope, that's the whole point of deep nets. a lower-dimensional manifold does same job as a "conceptual vocabulary"
2 replies 0 retweets 0 likes -
Replying to @bobpoekert @alicemazzy
That’s the hope… I find it dubious. https://meaningness.com/probability-and-logic … points out one reason, as Sarah Constantin explained.
1 reply 0 retweets 0 likes -
-
Replying to @bobpoekert @alicemazzy
My prior for things like this is “nope, nope, nope, bullshit”—based on a lot of experience.
1 reply 0 retweets 0 likes -
Which makes me reluctant to look at that specific case; but that means I might miss something important.
1 reply 0 retweets 0 likes -
Replying to @Meaningness @alicemazzy
that's a good prior but word2vec does work for real practical NLP tasks
2 replies 0 retweets 0 likes -
Replying to @bobpoekert @alicemazzy
My prior is “if it really worked, I would have heard about it before”… Impress me with evidence?
1 reply 0 retweets 0 likes -
Replying to @Meaningness @alicemazzy
here's a practical application of it for medical documents https://arxiv.org/pdf/1502.03682
2 replies 0 retweets 0 likes
“maximum accuracy of 49.28%” Nope, sorry, not impressed at all
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.