The thing that’s wrong about “hooray, let’s eliminate the value function and replace it with rollouts” is that the value function is often way more interesting than the policy

Show Post

Have little trouble believing such a result but it does point to something quite weird. What would have to be going on in a field such that a proliferation of improvements wouldn’t turn into gradient ascent on architectures. Like, there’s so much more work now than when LSTMs came out

Show Post

I'm sorry, are you saying that "the other finbarr" was lighting up a lot of timber?

Show Post

Awful suspicious

Show Post

Drunken uni students:)

Show Post