An analysis of noise in recurrent neural networks: convergence and generalization

Theoretical results show that applying a controlled amount of noise during training may improve convergence and generalization performance, and it is predicted that best overall performance can be achieved by injecting additive noise at each time step.