The potential for simulation in deep learning is vast, however it is currently limited when it comes to dealing with novel perturbations in the real world. With increased transfer capabilities, simulation could become a much more useful tool for real world applications.