ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning
Neeraj Yadav
cs.CV, cs.LG
2026-09-16
ALPINE (22,249 params) uses a Gabor-guided patch locator. Matched 250-episode 5-shot: 56.39% on CIFAR-FS vs ProtoNet 53.77%; dropping relation tokens costs at most 1 point.
Few-shot classification papers mostly report accuracy. For anyone without a GPU cluster, two other axes are equally hard: how many parameters the model needs, and how many training episodes it takes to get there. Typical comparisons use million-parameter backbones and tens of thousands of meta-training episodes. This paper restricts the contest to from-scratch training, 250 episodes, and a 22k-50k parameter band.
The original hunch was perceptual, not computational. A fixed rectangular patch grid can split a face or a limb, or dump unrelated pixels into one cell, and the mismatch gets worse under occlusion and translation. The bet was that patches which drift toward informative regions would be more robust, without extra parameters. A second bet, pairwise spatial relations such as "a head belongs with a body", was the project's original story. The later falsification tests knock that second bet down.
The system is named ALPINE; the experimental configs are EXP-F3. An image goes through four stages.
A frozen 4-orientation Gabor bank (0°, 45°, 90°, 135°, 5×5, not learned) produces an edge-energy map and extra input channels. Five patch slots sit at the four corners plus the center. Each slot takes a spatial softmax over the energy map, plus a Gaussian window nailed to its own anchor, with windowfrac=0.5. The patch center is the expected (y, x) under that distribution, then bilinear sampling cuts the patch. A fully open window collapses all five patches onto one global energy peak. A tight window barely moves. 0.5 is the compromise that recovered clean accuracy while keeping most of the occlusion gain.
A tiny CNN encodes each patch at two scales into five vectors. Pairwise differences and Hadamard products yield ten relation tokens; four fixed inner-crop boundary tokens join them. Nineteen tokens enter a one-layer 4-head self-attention block. A learned query summarizes the sequence, then a gated residual fuses that summary with mean-pooled patch vectors. Classification is vanilla Prototypical Networks: class prototypes are mean support embeddings, queries use negative squared Euclidean distance.
The canonical embedding width of 16 has 22,249 trainable parameters; width 32 has 34,917. Baselines: ProtoNet 47,630, RelationNet 49,617, MAML 49,481.
The protocol is locked: 5-way CIFAR-FS and MiniImageNet, exactly 250 meta-training episodes, five seeds, 600 evaluation episodes per seed. Published MAML usually trains for tens of thousands of episodes; numbers here are iso-budget only and are starred.
| Method | Params | CIFAR-FS 5-shot | MiniImageNet 5-shot |
| EXP-F3-35k | 34,917 | 58.67% | 53.81% |
| EXP-F3 | 22,249 | 56.39% | 53.37% |
| ProtoNet | 47,630 | 53.77% | 48.82% |
| RelationNet | 49,617 | 47.60% | 38.18% |
| MAML | 49,481 | 34.30% | 30.93% |
The 5-shot lead holds on all five seeds. 1-shot is a tie with ProtoNet on CIFAR-FS (40.36 vs 40.45) and a small lead on MiniImageNet (35.88 vs 35.25). With n=5, Wilcoxon can get no stronger than p=0.0625, short of the usual 0.05.
Under 50% center masking, MiniImageNet 5-shot is 44.87 vs ProtoNet 40.36. Under 25% translation on CIFAR-FS the comparison is a statistical tie (42.65 vs 43.61), the one perturbation where EXP-F3 does not beat ProtoNet. MiniImageNet checkpoints transferred to CUB-200-2011 with no retraining: 42.09 vs ProtoNet 36.86, with similar drops (11.28 vs 11.96). RelationNet drops only 1.77 points because its in-domain baseline is 38.18.
Sample-efficiency curves use three seeds. On MiniImageNet, EXP-F3 crosses 45% within the first 50 episodes; ProtoNet needs 100 to 150.
The falsification tests matter more than the leaderboard. Zeroing the ten relation tokens at inference costs 0.84 to 1.09 points at 5-shot. Retraining a relation-free variant from scratch costs 0.00 to 0.98. A capacity sweep from 22k to 196k peaks near 35k on 1-shot and declines a little beyond that. Bolting LayerNorm and AdamW onto ProtoNet makes it worse (CIFAR-FS 5-shot 49.92).
For edge devices and single-machine researchers, this is a clean reference point: inside this parameter and episode budget, adaptive localization is worth more than another relational module. Five-seed checkpoints with SHA-256 hashes and a reproduction manifest are public.
Do not compare it with the current MiniImageNet 5-shot ceiling. That literature uses larger, often pretrained backbones and sits around 70-81%. The paper says it is not in that race.
1-shot has no stable edge over ProtoNet, and scaling to 196k parameters does not close it. Three diagnostic hypotheses (embedding noise, space compression, softmax-confidence flattening) were tested and pointed the wrong way. The MAML numbers are not a win over a fully converged MAML. Benchmarks are CIFAR-FS, MiniImageNet, and CUB at 84×84. Sample-efficiency curves have only three seeds. Pairwise relations were the original motivation; the author's own tests show they are not the main driver. The ten tokens stay in the final model because they add a small, consistent gain. This is a single independent researcher; there is no multi-lab reproduction.