ALPINE beats ProtoNet, RelationNet and MAML on 5-shot CIFAR-FS and MiniImageNet with a 22,249-parameter adaptive locator, and its own ablations show relational tokens are not the main source of the gain

Under a strictly matched iso-episode budget, ALPINE's 5-shot accuracy exceeds all three baselines on both CIFAR-FS and MiniImageNet, with all five seeds agreeing in direction. Few-shot learning evaluation predominantly reports accuracy alone, with limited attention to the parameter and training-sample budgets needed to reach it; this work treats parameter budget and training-episode budget as first-class evaluation axes and compares against baselines under the same budget. Table 1 reports 5-seed mean plus or minus standard deviation: EXP-F3 (22,249 parameters) reaches 56.39 plus or minus 1.28 on CIFAR-FS 5-shot and 53.37 plus or minus 0.72 on MiniImageNet 5-shot, against ProtoNet at 53.77 plus or minus 0.34 and 48.82 plus or minus 0.62, RelationNet at 47.60 plus or minus 0.62 and 38.18 plus or minus 2.48, and MAML at 34.30 plus or minus 3.04 and 30.93 plus or minus 1.40; the authors report a Wilcoxon signed-rank p of 0.0625 and state this is the strongest attainable signal at n equals 5 seeds, below the conventional p less than 0.05 threshold.

The content-adaptive patch locator is the primary mechanism behind the architecture's efficiency and robustness, while pairwise relational tokens provide only a small auxiliary benefit. The architecture's original motivating hypothesis was explicit pairwise spatial relational computation; the authors test that hypothesis directly with falsification ablations rather than assuming it holds. Table 4 shows that zeroing all ten relational tokens at inference drops accuracy by only 0.19-1.09 percentage points, and retraining an architecturally relation-free variant from random initialization (21,689 parameters) drops 0.00-0.98 points, closely matching the first test and ruling out relational information being absorbed elsewhere during training; a separate control adding LayerNorm and AdamW to a vanilla ProtoNet degrades it instead (49.92 and 43.02), ruling out modernized training as an alternative explanation.

The locator's window constraint is the key design choice resolving a collapse failure mode: an unconstrained locator sends all five patches to one dominant energy peak at the cost of clean-image accuracy, while a moderate window restores spatial diversity while retaining most of the occlusion robustness. The authors report both the failure mode of the unconstrained version (window_frac equals 1.0) and its repair at a moderate setting (window_frac equals 0.5), with quantitative patch-center-distance measures. Under the unconstrained locator, pairwise patch-center distance collapses from 0.997 for the fixed grid to approximately 0.31, and clean accuracy falls 3.8-5.7 percentage points relative to the fixed-grid baseline; at window_frac equals 0.5 the distance returns to 0.68-0.74 with center drift of about 0.61, recovering clean accuracy to within 0.4-1.6 points of the fixed-grid baseline; random-shuffle and random-shift control conditions verify the gains come from genuine content-adaptivity.

The architecture transfers better to an unseen domain and to common perturbations, and converges faster, but its 1-shot advantage over ProtoNet is not stable. Cross-domain zero-retraining transfer, occlusion and translation robustness, and training-episode-versus-accuracy curves are all evaluated under the same matched-budget protocol, which the authors note is not jointly applied to few-shot architectures at this parameter scale. Table 3 shows EXP-F3 achieving the highest absolute accuracy on the unseen CUB-200-2011 domain at 42.09 plus or minus 0.94 versus ProtoNet at 36.86 plus or minus 1.04, consistent across 5 of 5 seeds with p equals 0.0625; Table 2 shows 46.54 plus or minus 1.33 (CIFAR) and 44.87 plus or minus 1.27 (Mini) under 50% masking, and 42.65 plus or minus 2.08 and 44.81 plus or minus 1.16 under 25% translation, with CIFAR-FS translation a statistical tie with ProtoNet at 3 of 5 seeds; Section 4.5 reports, over 3 seeds, that MiniImageNet accuracy exceeds 45% within the first 50 episodes while ProtoNet needs 100-150, which the authors explicitly flag as directional rather than confirmatory.

This work is aimed at practitioners under compute and data constraints: researchers working from a single consumer laptop, teams deploying on edge devices, and anyone for whom a multi-million-parameter backbone and tens of thousands of meta-training episodes are not an option. The setting it applies to is 5-way few-shot image classification at native 84x84 resolution, with small-budget from-scratch meta-training of 250 episodes, on CIFAR-FS, MiniImageNet, and CUB-200-2011 as an unseen fine-grained domain. The authors release all source code, the full 5-seed checkpoint set with SHA-256 hashes, reproduction manifests, figure-generation scripts and a checkpoint verification script, so follow-up work can reproduce or extend the protocol under the same iso-episode budget, for example by adding more benchmarks, more cross-domain targets, or porting the content-adaptive locator into other low-parameter architectures.

Several scope boundaries are worth watching. In the 1-shot setting the architecture shows no consistent, statistically significant advantage over ProtoNet and is effectively tied on CIFAR-FS; a capacity sweep up to 196,035 parameters did not resolve the gap, and the authors report that three diagnostic hypotheses (embedding noise, embedding-space compression, and softmax-confidence flatness) were each tested and rejected, leaving this an open problem. The MAML baseline is trained under the same 250-episode iso-budget, far short of the tens of thousands of episodes typical of published convergence, and the authors explicitly ask that MAML comparisons be read as budget-matched rather than literature-converged. The sample-efficiency curves rest on 3 seeds and are flagged by the authors as directional. All headline comparisons use n equals 5 seeds, at which the strongest attainable Wilcoxon signal is p equals 0.0625, reported honestly rather than implying conventional significance. In addition, although the full paper text was loaded, Figures 1 through 4 appear only as captions, so the visual content itself cannot be verified and the specific shapes of the patch-placement visualizations and the capacity curve rest on the prose descriptions.