3 ms·
I'm trying to figure why the person making the post failed. There are existing Haskell libraries for automatic differentiation [1]. As opposed to ordinary symb
by joe_the_user 8y ago
I'm trying to figure why the person making the post failed. There are existing Haskell libraries for automatic differentiation [1].
As opposed to ordinary symbolic differentiation or numerical differentiation, automatic differentiation is a method that takes the code of a function and a value and returns the derivative of the function at run time.
Of course, the author is working with a vector library and so his version will take N-dimension vectors/tensors and a functions of these and return N+something dimension vectors/tensors. That part is kind of head spinning and I assume that's where things fell apart.
[1] http://hackage.haskell.org/package/ad http://hackage.haskell.org/package/ad
- eastWestMath 8y agoWell if you look at the categorical semantics of differentiation, it takes in a function A -> B and spits out AxA -> B. It’s a fairly well developed area of logic/type theory (even at the level of term rewriting systems), it seems the author of the blog post missed it. I probably would have emailed him some references if I’d seen the initial announcement...
- nmca 8y agoBecause this was intended to support `accelerate` and uses the variable-mode ad approach presented in 'the simple essence of automatic differentiation' which afaik has not been implemented in conjunction with `accelerate` . (Your comment reads quite uncharitable/mean btw, hard to tell if it was intentional)
- tree_of_item 8y ago> (Your comment reads quite uncharitable/mean btw, hard to tell if it was intentional) Really? I didn't get that at all. Was it edited?