ALPINE: a 35K-parameter few-shot architecture that beats MAML with 27-53% fewer parameters

Neeraj Yadav · hf · 2026-09-23

ALPINE presents an ultra-lightweight (22,249–34,917 parameter) spatial-relational architecture for few-shot image classification, combining fixed Gabor edge-energy guidance with a windowed, content-adaptive patch locator.

Setup and results (strictly matched budget: 250 meta-training episodes, 5 seeds, 600 evaluation episodes per seed):

Honest ablations: falsification tests (zeroing relational tokens at inference, retraining without them) show the pairwise relational computation exists but is not the main performance driver — the content-adaptive patch locator is. Authors report a genuine accuracy plateau near 22-35k parameters and release full seed-level results and checkpoint hashes for reproducibility.

Original post →

More from Research

Research channel →