Source-linked AI summary
ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning
Neeraj Yadav
TL;DR
Few-shot research often emphasizes accuracy while giving less attention to model size and training episodes, both important under limited compute. ALPINE combines fixed Gabor guidance with adaptive patch localization; under matched budgets, it reports seed-consistent 5-shot gains with fewer parameters, alongside faster convergence, transfer, and robustness results.
Problem
Few-shot evaluations focus predominantly on accuracy, with comparatively little attention to parameter counts or the training episodes needed to reach it, despite their relevance to practitioners with limited compute.
Source: §1 · p. 1Method
ALPINE combines fixed Gabor edge-energy guidance with a windowed, content-adaptive patch locator in a lightweight few-shot image-classification architecture.
Source: §1 · pp. 1–2Results
27–53% fewer parameters than any baseline, with seed-consistent 5-shot accuracy gains over Prototypical Networks, Relation Networks, and MAML on CIFAR-FS and MiniImageNet.
Source: §1 · p. 2, §6 · p. 10Takeaways & Limitations
Falsification tests identify adaptive spatial localization, rather than relational computation, as the primary driver of the architecture’s efficiency gains.
Source: §1 · p. 2, §6 · p. 10The architecture has no consistent, statistically significant advantage over Prototypical Networks in 1-shot performance, and the paper reports no mechanistic explanation for this gap.
Source: §5 · p. 9
Abstract
from arXiv · showhide
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.
1. Introduction
Few-shot evaluation often emphasizes accuracy while giving less attention to parameter and training-episode budgets, which matter to practitioners with limited compute. ALPINE replaces fixed patch placement with a windowed adaptive locator and reports gains across accuracy, convergence, transfer, and robustness under matched resource conditions.
1. Introduction
- The motivating hypothesis is that content-adaptive patches can target informative regions despite fixed-grid misalignment and improve robustness without additional parameters.A systematic ablation-driven process led to a windowed locator that resolves a collapse failure in the unconstrained design.Source: §1 · pp. 1–2
- On CIFAR-FS and MiniImageNet, 5-shot accuracy advantages over Prototypical Networks, Relation Networks, and MAML were consistent across all five seeds, with 27–53% fewer parameters than any of them.Source: §1 · p. 2
- It reaches comparable or higher accuracy within the first 50 of 250 meta-training episodes.Source: §1 · p. 2
- MiniImageNet-trained checkpoints generalize to unseen CUB-200-2011 bird species without retraining.Source: §1 · p. 2
- The architecture is more robust than all three baselines to 50% occlusion and 25% spatial translation, except that it ties ProtoNet on CIFAR-FS translation.Source: §1 · p. 2
- Falsification tests find that relational computation contributes only 0.3–0.9 percentage points; the content-adaptive spatial locator is the primary driver of efficiency.Removing relational computation and retraining reproduces the small performance gap, ruling out relational information being absorbed elsewhere during training.Source: §1 · p. 2
2. Related Work
The paper combines established techniques rather than claiming a new mathematical primitive. Its contribution is their parameter-efficient integration and evaluation under matched-budget, multi-seed, robustness, transfer, and falsification conditions.
Metric-based few-shot learning
- Prototypical Networks classify queries by distance to class-mean embeddings, and this work retains their prototype-distance classification head.The stated contribution is entirely in how the embedding is computed.Source: §2 · p. 2
Relation-based comparison
- Relation Networks compare query and support embeddings, whereas this architecture computes pairwise features among patches within one image.The paper says this specific relational mechanism is not its primary performance source.Source: §2 · p. 2
Optimization-based meta-learning
- MAML learns an initialization for rapid task-specific fine-tuning; the paper compares it under the same 250-episode budget and does not claim a win over converged MAML.Source: §2 · p. 2
Spatial attention and soft localization
- The locator uses differentiable sampling like spatial-attention methods, but its signal is fixed Gabor energy and sampling is constrained by a Gaussian prior to prevent collapse.Source: §2 · p. 2
Positioning
- The contribution is an extremely parameter-efficient integration of content-adaptive localization with established components, evaluated across matched budgets, robustness perturbations, cross-domain transfer, and falsification testing.The matched-budget, multi-seed protocol follows the evaluation-rigor concern raised by.Source: §2 · p. 3
Gabor and hand-crafted filter banks
- The architecture uses fixed, non-learned Gabor filters alongside established methods; its claimed novelty is their efficient integration and joint evaluation, not a new primitive.Source: §2 · p. 3
3. Method
The method combines fixed Gabor edge-energy guidance with a windowed, content-adaptive locator, patch and relational tokens, attention refinement, and prototype-distance classification. A moderate locator window balances spatial diversity with content adaptivity.
3.1 Overview
- The image pipeline applies fixed Gabor pre-filtering, selects five patch centers adaptively, encodes patches and relational and boundary tokens, then refines tokens with attention before gated fusion and prototype-distance classification.Source: §3.1 · p. 3
3.2 Gabor Pre-Filtering and the Adaptive Locator
- The locator uses fixed Gabor energy and a Gaussian penalty around each canonical anchor; window_frac = 0.5 balances spatial diversity with adaptivity, unlike unconstrained collapse or barely moving at 0.25.The patch center is the softmax expectation, and bilinear grid sampling extracts the patch there.Source: §3.2 · p. 3
3.3 Patch, Relation, and Boundary Tokens
- Each patch embedding combines coarse whole-patch and fused four-sub-patch views; ten pairwise relation tokens and four fixed boundary tokens join the five patch tokens.The 19-token sequence enters a single-layer, four-head self-attention block, followed by a learned-query summary and gated residual fusion.Source: §3.3 · pp. 3–4
3.4 Classification
- Classification uses class-mean support embeddings as prototypes and assigns queries by negative squared Euclidean distance, trained with cross-entropy.Source: §3.4 · p. 4
3.5 Parameter Budget
- The canonical configuration has 22,249 trainable parameters; the wider configuration has 34,917, versus 47,630–49,617 for each baseline.The wider configuration was found in an exploratory capacity sweep and improves 5-shot accuracy without materially changing the parameter-efficiency comparison.Source: §3.5 · p. 4
4. Experiments
Experiments use a matched 250-episode budget and evaluate standard-condition accuracy across five seeds. Both ALPINE configurations lead the baselines on 5-shot accuracy on both benchmark datasets, while 1-shot results are closer.
4.1 Setup
- All models receive 250 meta-training episodes and are evaluated over 600 episodes per seed across five canonical seeds under the iso-episode-budget protocol.The paper cautions that this is far below typical published MAML convergence budgets, so MAML results are budget-matched rather than literature-converged.Source: §4.1 · p. 4
4.2 Main Results
- Both ALPINE configurations outperform all three baselines on 5-shot accuracy on CIFAR-FS and MiniImageNet, with gains in the same direction across all five seeds.Against ProtoNet and RelationNet, the Wilcoxon signed-rank test gives p = 0.0625, the strongest attainable signal at five seeds but above the conventional p<0.05 threshold.Source: §4.2 · p. 4, Table 1 · p. 4
- The 1-shot comparison is closer: ALPINE is near a tie with ProtoNet on CIFAR-FS and has a modest, seed-consistent lead on MiniImageNet.Source: §4.2 · p. 4
4.3 Robustness
Under 50% masking and 25% spatial translation, EXP-F3 generally outperforms the baselines, with a CIFAR-FS translation tie against ProtoNet. On zero-retraining transfer to CUB-200-2011, EXP-F3 has the highest absolute accuracy among the four models.
4.3 Robustness
- EXP-F3’s masking advantage over ProtoNet is large and consistent across all five seeds on both datasets, while CIFAR-FS translation is a statistical tie.On MiniImageNet translation, EXP-F3 has a moderate, seed-consistent advantage over ProtoNet (4/5 seeds); it leads RelationNet and MAML by large margins on both perturbations and datasets.Source: §4.3 · pp. 4–5, Table 2 · p. 5
4.4 Cross-Domain Generalization
- EXP-F3 achieves the highest absolute accuracy among all four models on zero-retraining transfer from MiniImageNet to CUB-200-2011.The comparison is consistent across all five seeds against ProtoNet in both shot settings, with p = 0.0625, the strongest attainable signal at this sample size rather than conventional p<0.05 significance.Source: §4.4 · p. 5, Table 3 · p. 5
- RelationNet’s smaller raw accuracy drop on CUB-200-2011 reflects its low in-domain baseline, not superior transfer or higher task accuracy.Source: §4.4 · p. 5
4.5 Sample Efficiency
EXP-F3 reaches accuracy thresholds in fewer meta-training episodes than the baselines in the reported three-seed analysis. Ablations show that constraining the adaptive locator avoids patch collapse, while relational tokens add only a small auxiliary benefit.
4.5 Sample Efficiency
- On MiniImageNet, EXP-F3 exceeds 45% accuracy within 50 episodes across three seeds, while ProtoNet takes 100–150 and RelationNet and MAML do not reach it within 250.On CIFAR-FS, the pattern is less pronounced at 45% and clearer at 50%; the finding is directional and not yet verified across five seeds.Source: §4.5 · p. 5, Figure 3 · p. 6
4.6.1 Resolving Locator Collapse
- Setting window_frac to 0.5 restores patch spatial diversity and recovers clean accuracy to within 0.4–1.6 points of the fixed-grid baseline while retaining most occlusion-robustness gains.The unconstrained locator collapsed all five patches onto one energy peak and reduced clean accuracy by 3.8–5.7 points; random-shuffle and random-shift controls supported genuine content-adaptivity as the source of the gains.Source: §4.6.1 · p. 6
4.6.2 Boundary/Outline Tokens
- On the resolved locator, adaptive boundary rings show no consistent control-matched advantage over fixed-grid boundary tokens and are excluded from the final architecture.The initial test with the collapsed locator was confounded because the rings inherited the patches’ spatial collapse.Source: §4.6.2 · p. 6
4.6.3 Are Relational Tokens Load-Bearing?
- Zeroing relational tokens at inference drops accuracy by 0.19–1.09 points, and retraining without them drops it by 0.00–0.98 points, indicating a small auxiliary benefit rather than a primary driver.The retraining test addresses whether patch embeddings had absorbed relational structure; the authors identify the content-adaptive locator as the primary source of accuracy and efficiency. Table 4 summarizes the falsification ablations.Source: §4.6.3 · p. 6, Table 4 · p. 7
- Modernizing vanilla ProtoNet with layer normalization and AdamW degrades its performance rather than closing the gap.Source: §4.6.3 · p. 7
4.7 Capacity Analysis
The capacity sweep finds that 1-shot accuracy peaks near 34,917 parameters and then plateaus or mildly declines. A five-seed test at 122,897 parameters does not resolve the 1-shot gap against ProtoNet.
4.7 Capacity Analysis
- Accuracy rises from 22,249 parameters to a peak near 34,917, then plateaus and mildly declines at higher capacities on both datasets.The pattern is consistent with overfitting under the fixed 250-episode training budget; Figure 2 plots 1-shot accuracy against parameter count on a log scale.Source: §4.7 · p. 7, Figure 2 · p. 7
- At 122,897 parameters, the five-seed run does not resolve the 1-shot gap against ProtoNet significantly on CIFAR-FS or MiniImageNet (Wilcoxon p = 0.44 and p = 0.63, respectively).This contrasts with modest single-seed suggestions in earlier exploratory testing.Source: §4.7 · p. 7
- The paper reports the 1-shot gap as an open, apparently capacity-independent limitation.Source: §4.7 · p. 7
4.8 Mechanism Visualization
The visualizations show EXP-F3's adaptive patches shifting toward unmasked regions and tracking translated content, with Gabor energy maps shown as the locator's signal. They connect these placements to the finding that adaptive localization, rather than relational computation, is the primary mechanistic driver.
4.8 Mechanism Visualization
- Figure 1 compares fixed-grid and EXP-F3 adaptive patch placement under standard, 50%-masked, and 25%-shifted conditions; adaptive patches move toward unmasked regions and track shifted content.The accompanying interpretation identifies the locator, rather than relational computation, as the primary mechanistic driver.Source: §4.8 · pp. 7–8, Figure 1 · p. 8
- Figure 4 pairs three input images with their Gabor energy maps and resulting energy-guided patch placements.Source: Figure 4 · p. 9
Code and Data Availability
The paper releases code, checkpoints, reproduction manifests, and figure-generation scripts, along with a script for verifying checkpoint hashes.
Code and Data Availability
- The release includes source code, five-seed checkpoints with SHA-256 hashes, reproduction manifests, figure-generation scripts, and a checkpoint verification script.The verification script reproduces the paper's hash-integrity audit.Source: Code and Data Availability · p. 9
5. Limitations
The paper's limitations concern 1-shot results, absolute-accuracy and benchmark scope, budget-matched MAML comparisons, relational-token framing, and statistical scope. These boundaries qualify the claims and comparisons the study supports.
5. Limitations
- The 1-shot advantage over Prototypical Networks is not consistent or statistically significant, and the paper reports this gap as an open problem.A capacity sweep up to 196,035 parameters did not resolve it significantly; three diagnostic hypotheses were tested and rejected.Source: §5 · p. 9
- The work targets parameter and sample efficiency under a small from-scratch budget, not state-of-the-art absolute accuracy.MiniImageNet 5-shot results of approximately 53–59% are below approximately 70–81% results typically reached with larger, often pretrained backbones; comparisons are restricted to similar small parameter scales.Source: §5 · pp. 9–10
- The MAML baseline received the same 250-episode budget as other models, far short of typical published convergence, so comparisons are budget-matched rather than literature-converged.Source: §5 · p. 10
- Falsification tests do not support explicit pairwise spatial relational computation as the key mechanism, though relational tokens remain for their small, consistent, non-zero benefit.Source: §5 · p. 10
- Evaluation covers CIFAR-FS, MiniImageNet, and one cross-domain target, CUB-200-2011; broader benchmarks, more transfer targets, higher resolutions, and non-vision domains remain future work.Source: §5 · p. 10
- Headline comparisons use five seeds, for which the maximum attainable Wilcoxon significance is p=0.0625, while sample-efficiency curves use three seeds and are directional rather than confirmatory.Source: §5 · p. 10
6. Conclusion
Under a constrained, matched-budget protocol, the paper reports efficiency and robustness gains over three standard baselines. Its falsification testing finds that adaptive localization, not relational computation, is the primary source of the gains.
6. Conclusion
- Under a constrained matched-budget protocol, the architecture outperforms three standard baselines on 5-shot accuracy using 27–53% fewer parameters, converges faster, transfers better to an unseen domain, and is more robust to common corruptions.Falsification testing identifies adaptive spatial localization, rather than relational computation, as the primary source of these gains; checkpoints, seed-level results, and reproduction manifests are released.Source: §6 · p. 10