ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning
ALPINE is an ultra-lightweight few-shot image classifier that beats Prototypical Networks, Relation Networks, and MAML with 27-53% fewer parameters.
ALPINE combines fixed Gabor edge-energy guidance with a windowed, content-adaptive patch locator in only 22,249-34,917 parameters. Under a matched 250-episode budget, it achieves consistent 5-shot accuracy gains over Prototypical Networks, Relation Networks, and MAML on CIFAR-FS and MiniImageNet. Ablations show the patch locator, not the pairwise relational computation, drives performance; the paper releases seed-level results and checkpoint hashes for reproducibility.
- Ultra-lightweight model at 22k-35k parameters for few-shot classification
- Outperforms ProtoNet, RelationNet, and MAML with 27-53% fewer parameters
- Falsification ablations show patch locator is the primary performance driver
- Generalizes zero-shot to CUB-200-2011 and resists 50% occlusion
From the original article
1 min read at huggingface.co
Few-shot learning research is predominantly evaluated on accuracy alone, with limited attention to the parameter and training-sample budgets required to reach that accuracy - a real constraint for practitioners without large-scale compute.
By Neeraj Yadav · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.