Loss function for implementation of BMAML about few_shot_meta_learning HOT 2 CLOSED

Ok. Thanks for letting me know. My understanding is that the chaser loss encourages the MAML models to be diverse which leads the predicted uncertainty to be meaningful. Basically, modifying the loss function to encourage diverse MAML models is important. This might not have been fair to Bayesian MAMl.

Still, great work and thanks for sharing the code.
Deep

from few_shot_meta_learning.

Hi Deep,

You are right. I only implemented the simple cross-entropy loss for BMAML. I cannot recall exactly, but I think the chaser loss can be implemented for other meta-learning methods. Hence, I decided to use cross-entropy loss to fairly compare BMAML to other meta-learning algorithms.

from few_shot_meta_learning.