Investigate how to make gradient wrt input layer efficient in regularization

[facusapienza21]

Both finite differences and AD are slow at computing gradients with respect to the weights on the NN, which is strange since the regularization step is currently more time expensive that the actual adjoint calculation.