Neural networks-based backward scheme for fully nonlinear PDEs