Thanks, I saw the article on that a few days ago here and it looks really interesting.
Is it really right that I could just use that to automatically recreate the entire backprop algorithm? Is it likely to be reasonable in its performance?
Pass. I've read the papers, I've seen the demo, but I haven't implemented it myself. I also read it here, but I've seen it before. It looks like it should work, but YMMV.
As always, you might need to spend some time understanding what it does and how it works before relying on it.
Is it really right that I could just use that to automatically recreate the entire backprop algorithm? Is it likely to be reasonable in its performance?