ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning
ALPINE’s reported edge is a 22k-35k parameter few-shot vision model whose gains come mainly from adaptive patch localization.
The paper says ALPINE beat Prototypical Networks, Relation Networks, and MAML in 5-shot tests on CIFAR-FS and MiniImageNet under a matched episode budget. It used 27-53% fewer parameters than the baselines and converged in fewer training episodes. The author also reports stronger zero-retraining transfer to CUB-200-2011 and better robustness under occlusion and spatial translation. Ablations found the relational tokens were not the main driver; the content-adaptive patch locator was. HF Daily Papers' note
The paper says ALPINE beat Prototypical Networks, Relation Networks, and MAML in 5-shot tests on CIFAR-FS and MiniImageNet under a matched episode budget. It used 27-53% fewer parameters than the baselines and converged in fewer training episodes. The author also reports stronger zero-retraining transfer to CUB-200-2011 and better robustness under occlusion and spatial translation. Ablations found the relational tokens were not the main driver; the content-adaptive patch locator was. HF Daily Papers' note
score 4