Table 2 Algorithm for updating model parameters
Algorithm 1 Meta-Learning Regression Model |
|---|
Input: MAML with parameters \(\theta\), Base-Learner with step size hyperparameter \(\alpha\), Meta-Learner with step size hyperparameter \(\beta\) \(\theta_{0}\)\(\leftarrow\) random initialization #Randomly initialize model parameters |
while not done do \(D_{train}\), \(D_{test}\)\(\leftarrow\) random datasets for all \({\text{T}}_{{\text{i}}}\) do \(x_{j}\), \(y_{j}\)\(\leftarrow\) random batch from support set \(L \leftarrow L_{{T_{i} }} \left( {f_{\theta } } \right)\)#Get loss using Eq. (3) \(g \leftarrow \nabla_{\theta } L_{{T_{i} }} \left( {f_{\theta } } \right)\)#Evaluate gradient with respect to parameters \(\theta\) \(\theta_{i}^{{\prime}} \leftarrow \theta - \alpha \nabla_{\theta } L_{{T_{i} }} \left( {f_{\theta } } \right)\)#Update Meta-Learner parameters end for \(X\), \(Y\)← random batch from quest set Update \(\theta \leftarrow \theta - \theta_{i}^{{\prime}} \leftarrow \beta \nabla_{\theta } \sum\limits_{{T_{i} \sim P(T)}} {L_{{T_{i} }} \left( {f_{{\theta_{i}^{{\prime}} }} } \right)}\) #Update Base-Learner parameters end while |