Efficient Approximations of the Fisher Matrix in Neural Networks using Kronecker Product Singular Value Decomposition