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):
- Beats Prototypical Networks, Relation Networks, and MAML on 5-shot CIFAR-FS and MiniImageNet while using 27-53% fewer parameters.
- Converges in fewer episodes; generalizes zero-shot to unseen fine-grained CUB-200-2011 birds.
- More robust to 50% occlusion and 25% spatial translation than all three baselines.
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.
More from Research
- TensorSharp's logit-reading approach beats LocalJev at structured decisions, 3.3x faster — fuzhongkai · 2026-09-23
- Naproche: a proof assistant that reads math proofs written in controlled natural language — zetalyrae · 2026-09-23
- Nanjing University used AI to design proteins 4x stronger than any natural protein — MikePFrank · 2026-09-23
- Open-source GUI agents top out at 8% task success on composite cross-device tasks — maier_ak · 2026-09-23
- JarvisGUI benchmark tests GUI agents across Android, Windows and Ubuntu in one workflow — maier_ak · 2026-09-23
- Cross-DURIAN: A New Benchmark for Multi-Device GUI Agents — maier_ak · 2026-09-23