We sample some batch of tasks from the task distributions, learn their concepts via the concept generator, perform meta learning on those concepts, and then we compute the meta learning loss:
Our meta learning loss varies depending upon what meta learner we use, such as MAML or Reptile.
Our final loss function is a combination of both of these, concept discrimination and meta learning loss:
In the previous equation, lambda is a hyperparameter balancing between meta learning and concept discrimination loss. So, our objective becomes finding the optimal parameter that minimizes this loss:
We minimize the loss by calculating gradients and update our model parameters: