Matrix momentum for practical natural gradient learning