This PR introduces the approximate computation of inverse-hessian-vector-products, useful for calculating influence vectors and values. By using forward-over-backward autodiff combined with a conjugate gradients solver, we are able to compute these quantities without the need for storing or inverting the hessian matrix.
I had to create this second version of the PR because of issues in the git history...
This PR introduces the approximate computation of inverse-hessian-vector-products, useful for calculating influence vectors and values. By using forward-over-backward autodiff combined with a conjugate gradients solver, we are able to compute these quantities without the need for storing or inverting the hessian matrix.
I had to create this second version of the PR because of issues in the git history...