This is possible because the Hessian of a deep net has a matrix polynomial structure that factorizes nicely. The Hessian-inverse-product algorithm that takes advantage of this is similar to running backprop on a dual version of the deep net. It echoes an old idea of Pearlmutter's for computing Hessian-vector products.

Maybe this idea is useful as a preconditioner for stochastic gradient descent?