From reading the abstract it seems that they are claiming that introducing some randomness into your gradient of weight changes allows for the quicker convergence of solution - I did not read the paper. I also don't exactly understand why it works - it sounds like they are claiming traditional back prop has room for improvement.