Self-supervised Meta-Prompt Learning with Meta-Gradient Regularization for Few-shot Generalization