Higher order derivatives #25
Description
Greetings!
For my application, I need higher order (>= 2) derivatives of the equilibrium point. I realize that doing this with the current implementation is quite inefficient, as it requires to backprop through the custom backward hook, i.e., backprop through the vjp solver. I believe I know how to compute the second derivative, but I am not sure how to do it using backward hooks (only using the functional approach).
It was mentioned in the NeurIPS'20 tutorial that "the function above will not work with double backprop, though again this can be adressed with some additional effect if needed". So is there some solution that I'm not seeing?
Also, I couldn't find any implementations of such higher order derivatives/double backprop of DEQs, so it would be very helpful if you could share some references (if any).