optimization-based meta-learning offers a promising direction for few-shot
learning that is essential for many real-world computer vision applications.
However, learning from few samples introduces uncertainty, and quantifying
model confidence for few-shot predictions is essential for