There are several instances that I know about of machine learning / signal processing methods ( for example empirical mode decomposition) that purposely inject noise into the algorithm to improve accuracy / fidelity / independence. I'm sure that others here can provide better examples than I.