How do you update the base_leaner's parameters

[xugy16]

Thank you for the code.

I have a question about the base_learner update.

1. The base-learner is fast updated using 100 steps.
2. then we return qry_logits and calculate the cross entropy loss for qry_set
3. using sel.optimizer to update.

But what gradient is stored in base_learner? Because you use fast-model to calcualte qry loss

[Yaoyao Liu]

Thanks for your interest in our work.

The fast model can be regarded as a function of the base learner. Thus, we can calculate the derivative of the query loss with respect to the base model. We update the base learner using Eq. 5 in our paper.

If you have any further questions, please feel free to contact me.

Best,

Yaoyao

[xugy16]

Really appreciate for the response.

So you are using 1st-Order MAML to update the classifier-head (base learner)?

[Yaoyao Liu]

Yes. We use the first-order approximation MAML to update the FC classifier.

If you have any further questions, please do not hesitate to contact me.