On transfer learning of neural networks using bi-fidelity data for uncertainty propagation