You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
There might be, but it will be by definition less optimal when applied to just English. You won't find it by optimizing for English.
The more general point is that generalization requires you to store seemingly useless information at training time (info that doesn't help your training objective), that will become useful in the future (when generalization actually happens). That's the opposite of compression.
Now, obviously, generalization requires abstraction, which requires erasing irrelevant details, so your high-generalization system will be doing *some* amount of compression. But it will be storing lots of seemingly useless info as well.
Is this why you argue in the paper that it is a necessity for a generalizable system to have a cost. Like it can't get away from storing useless info?
Here I see what the "cost" part means. The point I am struggling with is it being completely the opposite. I think I might be influenced by my intuition on compression as a tool for understanding. For me compression of information feels useful and transferable.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
