关键突破：从Transformer到”任何复杂神经网络” 论文的结论部分有个惊人的陈述： “Our results remain valid if the self-attention layer is switched by other forms of contextual layers, like that of a RNN, or any layer that can take an in