ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning
QUESTION — How can we build a computer vision architecture that achieves high few-shot accuracy under extremely constrained parameter and sample budgets?
The authors present ALPINE, an ultra-lightweight (22,249-34,917 parameter) spatial-relational architecture for few-shot image classification. It combines fixed Gabor edge-energy guidance with a windowed, content-adaptive patch locator. Under a strictly matched, iso-episode-budget protocol, the architecture achieves 5-shot accuracy gains over Prototypical Networks, Relation Networks, and MAML on both CIFAR-FS and MiniImageNet, while using 27-53% fewer parameters than any baseline.
The architecture contains 22,249-34,917 parameters.
It uses 27-53% fewer parameters than any baseline.
It is more robust to 50% occlusion and 25% spatial translation than all three baselines.
NJ50 · 16 Sept 2026
read the original ↗