Robust Training of Neural Networks against Bias Field Perturbations