Bridging Few-Shot Learning and Adaptation: New Challenges of Support-Query Shift

PGADA: Perturbation-Guided Adversarial Alignment for Few-shot Learning Under the Support-Query Shift

“…The conventional few-shot learning methods assume the support set and the query set lie in the same distribution. A more realistic setting is that the support set S and the query set Q follow different distributions, i.e., the support-query shift [2]. While these two sets are sampled from different distributions µ s and µ q , the embeddings for the support set S (i.e., φ(x s )) and the query set Q (i.e., φ(x q )) are likely to lie in different embedding spaces.…”

“…While these two sets are sampled from different distributions µ s and µ q , the embeddings for the support set S (i.e., φ(x s )) and the query set Q (i.e., φ(x q )) are likely to lie in different embedding spaces. Thus, it would lead to a wrong classification result via the comparison module M (φ(x s ), φ(x q )) [2].…”

“…Datasets. Following [2], we validate our framework on three benchmark datasets in few-shot learning. 1) CIFAR100 [20]…”

“…Moreover, the labels of training data and testing data are usually disjoint, i.e., the labels in the testing phase are unseen in the training phase. In contrast, few-shot learning aims to learn the model parameter by a handful of training data (support sets) and adapt to the testing data (query sets) [30,11,2], effectively. In the testing phase, the model classifies the query set into the support sets according to the distance of their embeddings.…”

“…While these approaches extract rich contextual features from the images, the embeddings from the support set and the query set usually exhibit certain distribution shifts, i.e., the Support-Query Shift [2]. For example, the images are captured by various devices, e.g., smartphones and single-lens reflex cameras in different environments, e.g., foggy and high-luminance.…”