Comment by RSchaeffer

Go watch Yatin Gal's talk on dropout in neural networks. He shows pretty convincingly that the belief that dropout reduces network overfitting by introducing noise is wrong.