ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning
Neeraj Yadav
Independent Researcher
Published September 16, 2026 · on the feed September 23, 2026 · ▲ 10 · arXiv 2609.22323
The abstract
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. We present an ultra-lightweight (22,249-34,917 parameter) spatial-relational architecture for few-shot image classification that combines fixed Gabor edge-energy guidance with a windowed, content-adaptive patch locator. Under a strictly matched, iso-episode-budget protocol (250 meta-training episodes, 5 canonical seeds, 600 evaluation episodes per seed), our architecture achieves 5-shot accuracy gains, consistent across all five seeds, over Prototypical Networks, Relation Networks, and MAML on both CIFAR-FS and MiniImageNet, while using 27-53% fewer parameters than any baseline. It also converges in fewer training episodes, generalizes better to an unseen fine-grained domain (CUB-200-2011 birds, zero retraining), and is more robust to 50% occlusion and 25% spatial translation than all three baselines. A series of falsification ablations - zeroing relational tokens at inference and retraining without them entirely - shows that the architecture's pairwise relational computation, while present, is not the primary driver of its performance; the content-adaptive patch locator is. We report this honestly, together with a capacity sweep showing a genuine accuracy plateau near 22-35k parameters, and release full seed-level results and checkpoint hashes for reproducibility.
The Brief
2:34A short podcast summary of the paper. Written and voiced by AI.
Transcript
ALPINE. Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning. By Neeraj Yadav, working independently. Published on the sixteenth of September. Its code repository is public and real: seventeen source files.
Few-shot learning research is mostly judged on accuracy alone. Much less attention goes to the budgets needed to reach that accuracy: how many parameters, and how many training samples. For practitioners without large-scale compute, those budgets are a real constraint, and a method that wins on accuracy by spending far more of them may not be the one they can use.
ALPINE is an ultra-lightweight spatial-relational architecture for few-shot image classification. Its size ranges from twenty-two thousand two hundred and forty-nine parameters, at the low end, to thirty-four thousand nine hundred and seventeen, at the high end. It combines two components. One is fixed Gabor edge-energy guidance. The other is a windowed, content-adaptive patch locator.
The comparison is run under a strictly matched protocol with the same episode budget for every method: two hundred and fifty meta-training episodes, five canonical seeds, and six hundred evaluation episodes per seed. The baselines are Prototypical Networks, Relation Networks, and MAML, on CIFAR-FS and MiniImageNet. Matching the episode budget is the point. Every method gets the same amount of training, so any difference in accuracy cannot be put down to one method simply having trained for longer.
ALPINE achieves gains in five-shot accuracy over all three baselines on both datasets, and the gains are consistent across all five seeds. It does so with twenty-seven to fifty-three percent fewer parameters than any baseline. It also converges in fewer training episodes. It generalizes better to an unseen fine-grained domain, the CUB-200-2011 birds, with no retraining. And it is more robust than all three baselines to fifty percent occlusion and to twenty-five percent spatial translation.
Then comes a set of falsification ablations. The relational tokens were zeroed at inference, and the model was retrained without them entirely. The result: the pairwise relational computation is present but is not the main driver of performance. The content-adaptive patch locator is. The author reports this plainly, alongside a capacity sweep showing a genuine accuracy plateau at roughly twenty-two to thirty-five thousand parameters.
The paper reports a finding against its own headline component. The relational part named in the architecture turns out to matter less than the patch locator beside it, and the author states that directly, in the abstract itself. Full seed-level results and checkpoint hashes are released for reproducibility.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a laptop, a handful of examples, and a grid that cuts faces in half.
You are building a few-shot image classifier on a single laptop: no big backbone, no tens of thousands of training episodes. It must learn new classes from a handful of examples.
The road ahead
- 1 · The problem ~7 min A tiny few-shot classifier on a laptop cuts images into a fixed grid, and that grid splits the very structures it needs to recognise.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Let a fixed edge-energy map pull each patch toward informative regions, while a window around its home spot keeps the patches spread out.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Filter the image into an edge-energy map, place five patches with the windowed locator, then relate every pair of patches.
- 6 · The experiment ~6 min You’ll guess: Falsification ablations zeroed and removed ALPINE's relational tokens. Which component turned out to be the main driver of its performance?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.