On the capabilities of neural networks using limited precision weights