Improving accuracy of neural networks compressed using fixed structures via doping