There was once a neural network that used dreaming phases for regularisation. It would run in reverse on random data and whatever activated was down–weighted.