The patent would apply to the model's architecture rather than the trained values so the black box nature of deep learning doesn't matter too much.