设 和 分别是门控网络和第 个 Expert 的输出，那么对于在当前的输入x，输出就是所有 experts 的加权和： Targeting cookies These cookies are grup on our şehir by our advertising partners. They may be used by relevant companies to build your interests profile and show relevant ads on other sites. 对比一下可以看出，在计算每个 expert 的损失之后，先把它给指数化了再进行加权求和，最后取了log。这也... https://www.blogger.com/u/4/profile/02778284436385530111