CVPR 2021
간단하게 생각하면 얼마나 비슷하게 복원 가능한지가 그 클래스로 분류될 가능성이다.
is n-way, k-shot episode set of support images.
Objective: Predict class label of query image
is a feature map of image , with being height times width and being the number of channels.
is a pool of feature maps of support images for class
Objective -> Find weight such that
Optimal , or can be expressed as a least-squares problem with ridge regression(penalty ):
Closed form solution using matrix calculus:
To calculate the probability of class membership Euclidean distance is used:
where is temperature(learned)
Then cross-entrophy loss is calculated.
,