Transfer learning for chemically accurate interatomic neural network potentials