MAML理解

line6，each $\theta_i$θi​ is the one-step updated parameter for each task, which would not be used as parameters for testing. Instead, it is used to update $\theta$θ, which is the final initilization that we wish to get.

This way, it is asctually minimizing the average loss across different tasks after one-step update. In other word, MAML could find the initilization parameters that is optimal for training with only one step update.

Line 4-8 could be expressed equivalently as follow:

$\theta=\theta-\beta\nabla_{\theta}L(\theta) \\L(\theta)=\sum_{i=1}^{N}l_i(\theta_i') \\\theta_i'=\theta-\alpha\nabla_{\theta}l(\theta)$θ=θ−β∇θ​L(θ)L(θ)=i=1∑N​li​(θi′​)θi′​=θ−α∇θ​l(θ)