PaperlayerView exact revision on arXiv

Source-linked AI summary

Query-Conditioned Prototype Adaptation for Cross-Domain Few-Shot Learning: Single-Query Inference, Controlled Comparisons, and Failure Modes

Rushab Rasik Karania, Tomas Maul

arXiv:2609.30769v1cs.CVcs.LG

TL;DR

Cross-domain few-shot learning asks how to recognize new target-domain classes from very few labelled examples. WIPT jointly transforms a query and its supports under a fixed representation to form query-specific prototypes. Its results are conditional: it helps in selected settings, while fixed prototypes remain strongest in others.

  • Problem

    Cross-domain few-shot learning must recognize new target-domain classes from very few labelled examples, and this study asks what joint query-support adaptation adds under a fixed representation.

    Source: §1 · p. 2

  • Method

    WIPT jointly transforms one unlabelled query and labelled supports, then averages transformed supports class-wise to form query-specific prototypes while keeping the encoder fixed.

    Source: §1 · p. 2

  • Results

    WIPT is useful in selected regimes rather than uniformly superior to ProtoNet; ProtoNet has the highest mean in all 5-shot targets, and query participation improves the matched support-only Transformer on 5-shot ISIC.

    Source: §1 · p. 2, §7 · p. 18

  • Takeaways & Limitations

    The study offers a controlled account of when query-conditioned prototypes help difficult cross-domain few-shot decisions and when fixed prototypes remain preferable.

    Source: §7 · p. 18

    The study covers one source dataset, three target domains, and one frozen ViT-S/16 backbone, rather than leaderboard comparisons with larger fine-tuned or transductive systems.

    Source: §6.5 · p. 17

Abstract

from arXiv · show

Cross-domain few-shot learning requires adapting a classifier to a new visual domain from very few labelled examples without target-time parameter updates. We isolate one question: under a fixed global representation, what does joint query-support adaptation contribute to prototype construction? The Within-Instance Prototypical Transformer (WIPT) implements single-query test-time prototype adaptation by jointly transforming one unlabelled query and the labelled support embeddings, then forming query-specific class means. Using a shared frozen ViT-S/16 encoder, miniImageNet source training, and CUB, EuroSAT and ISIC targets, we replicate the key comparisons across five independent training seeds. In 1-shot evaluation, WIPT improves frozen ProtoNet in every run on CUB (+0.21 percentage points) and EuroSAT (+2.07), but decreases ISIC (-0.22). In 5-shot evaluation, ProtoNet remains strongest overall, while WIPT consistently improves a capacity-matched support-only Transformer on ISIC (+0.99). Joint processing of up to five queries yields no reliable accuracy gain; in a head-only 5-shot benchmark, g = 5 reduces analytical attention-token pairs by 73% and peak allocated memory by 29% relative to g = 1, although latency is non-monotonic. Across all target/shot conditions, WIPT changes uncertain ProtoNet decisions far more than confident ones, and rescue/break decomposition accounts for the observed gains and losses. Source-shift and scorer controls further show that the benefit is not universal. Overall, WIPT provides a streaming-compatible form of test-time prototype adaptation that can improve difficult low-shot cross-domain decisions without target-time optimization.

1. Introduction

The study asks what joint query-support processing adds to prototype construction under a fixed image representation. Its findings are conditional: query participation helps in some settings but not uniformly.

1. Introduction

  • WIPT jointly transforms one unlabelled query and the labelled supports, then averages transformed supports class-wise to form query-specific prototypes.The encoder stays frozen, with no target-time gradient update or pseudo-labelling.Source: §1 · p. 2
  • The study compares WIPT with frozen ProtoNet and a capacity-matched support-only Transformer across five training seeds, three target datasets, and two shot regimes.Its questions cover joint query participation, multi-query processing and cost, decision changes, and source-only pseudo-domain training.Source: §1 · p. 2
  • WIPT improves ProtoNet on 1-shot CUB and EuroSAT but reduces 1-shot ISIC accuracy; ProtoNet has the highest mean in all 5-shot targets.Additional query context provides no reliable accuracy gain, while rescue/break and boundary analyses characterize decision changes.Source: §1 · pp. 2–3

2. Related work

Prior few-shot methods vary in their use of support examples, similarity rules, and parameter adaptation. WIPT is positioned as a controlled study of joint query participation under a shared frozen representation.

2. Related work

  • Matching Networks, ProtoNet, and Relation Networks classify from support sets using fixed or learned similarity rules, whereas MAML learns an initialization for rapid parameter adaptation.FEAT adapts support-derived class representations, while PrototypeFormer uses prototype tokens and within-class supports.Source: §2 · p. 3
  • CrossTransformers, QPN, and HCPNet already use query information in single-query support matching or prototype formation [9].Other methods combine query sets with mechanisms such as rectification, self-training, or representation adaptation.Source: §2 · p. 3
  • WIPT excludes those additional mechanisms to study joint query participation under a shared frozen representation, and its support-only control is not a reproduction of FEAT or PrototypeFormer.The paper also tests one source-only appearance curriculum while freezing the encoder.Source: §2 · p. 3

3. Method

The method keeps the image encoder fixed and adapts prototypes by letting the current query interact with support embeddings. Matched controls distinguish this joint processing from support-only transformation.

3.1. Problem setting and test-time prototype adaptation

  • WIPT changes the episode-specific class representative as a function of the current query without target-time gradient updates, backbone tuning, or pseudo-label self-training.The task uses labelled supports and unlabelled queries in a closed set where every query belongs to a support class.Source: §3.1 · pp. 3–4

3.2. Shared frozen encoder

  • All controlled methods use the same frozen ViT-Small/16 encoder with 384-dimensional global embeddings, so comparisons differ only after embedding production.The implementation uses its default global feature output without an additional projection before the episodic head.Source: §3.2 · p. 4

3.3. Single-query WIPT

  • WIPT passes one query and all support embeddings through Transformer layers, allowing transformed supports and the transformed query to depend on each other.Class prototypes average the transformed supports by class, and classification uses negative Euclidean distance.Source: §3.3 · p. 4

3.4. Matched support-only control

  • The support-only Transformer matches WIPT's head configuration but transforms only supports, making its class means episode-specific rather than query-specific.It scores the raw frozen query against those means; this comparison also changes the query representation used for scoring.Source: §3.4 · p. 5

3.5. Query context and computational cost

  • For multi-query context, WIPT processes groups of up to five queries with the same supports, using the same 3,549,696-parameter head.The study measures attention-token pairs, head-only latency, and peak allocated memory; lower attention-pair counts do not imply proportional latency reductions.Source: §3.5 · p. 5

4. Experimental design

The experiments use shared representations and controlled episodic comparisons across three target domains. Training-run variation, decision changes, and secondary studies are assessed with explicitly bounded protocols.

4.1. Datasets and cross-domain protocol

  • Evaluation samples 5-way episodes from CUB, EuroSAT, and ISIC without target-domain parameter updates, with disjoint support and query images.The shared ImageNet-pretrained encoder means held-out miniImageNet classes are not guaranteed to be unseen during encoder pretraining.Source: §4.1 · p. 7

4.2. Controlled baselines and exact pipeline

  • The primary comparison fixes Euclidean class-mean scoring across frozen ProtoNet, support-only Transformer, and WIPT using the exact same encoder state.The scorer sensitivity study is separate, and the controls are not full reproductions of Matching Networks, Relation Networks, FEAT, or PrototypeFormer.Source: §4.2 · p. 7

4.3. Training and statistical evaluation

  • Trainable-model results report means and sample standard deviations across five independent training runs, with paired 95% Student's t intervals across five seed differences.Intervals are conditional on a fixed episode bank, unadjusted for multiple contrasts, and do not establish equivalence when they contain zero.Source: §4.3 · p. 8

4.4. Decision-change analysis and secondary experiments

  • Decision-change analysis counts ProtoNet errors corrected by WIPT and correct predictions broken by WIPT, then locates transitions relative to the ProtoNet margin.Query labels enter only post-prediction analysis, and the diagnostics describe patterns rather than establish causal pathways.Source: §4.4 · pp. 8–9
  • Secondary studies test EuroSAT corruption, source-only shared-shift training, and a five-seed 5-shot adaptation-scorer grid.Selected-checkpoint ablations and diagnostics are explicitly exploratory.Source: §4.4 · p. 9

5. Results

Across the replicated cross-domain comparisons, WIPT’s effect depends on target domain and shot count, and added query context yields no reliable accuracy gain. The results also characterize decision changes and show that corruption, source-shift, and scorer controls do not establish a universally stronger configuration.

5.1. In-domain reference

  • Held-out miniImageNet is near saturation: support-only reaches 98.37%, frozen ProtoNet 98.34%, WIPT 98.33%, and cosine matching 98.47%.The paper treats this as an in-domain reference because its questions concern non-saturated cross-domain targets.Source: §5.1 · p. 9

5.2. RQ1: replicated cross-domain performance

  • In 1-shot, WIPT beats ProtoNet on CUB (+0.210 points) and EuroSAT (+2.068) in all five runs, but falls below it on ISIC (−0.218); in 5-shot, ProtoNet has the highest mean on all three targets.Against support-only, WIPT’s clearest replicated 5-shot gain is +0.994 points on ISIC, positive in all five runs.Source: §5.2 · p. 9, Table 1 · p. 10, Figure 2 · p. 11

5.3. RQ2: additional query context and cost

  • Across tested multi-query settings, accuracy changes relative to g = 1 are not reliable; in 5-shot, g = 5 cuts token pairs by 73% and peak allocated memory by 29%, while latency is non-monotonic.The 5-shot EuroSAT g = 5 interval is [−1.536, +0.536] points, so the result does not establish equivalence.Source: §5.3 · p. 11, Figure 3 · p. 12

5.4. RQ3: shot, domain and decision changes

  • Rescues and breaks account for net accuracy changes, while transitions concentrate near ProtoNet’s decision boundary; the observed boundary localization does not explain why individual transformations help or harm.Support-to-query centroid error and query-conditioned prototype variability decrease from 1-shot to 5-shot, but these are descriptive associations.Source: §5.4 · pp. 12–13, Figure 4 · p. 14

5.5. Controlled corruption response

  • Under EuroSAT corruption, WIPT’s mean effect versus ProtoNet shifts from −0.26 points at severity 3 to +0.75 and +1.02 at severities 4 and 5; WIPT averages above support-only in every condition.Brightness/contrast alone mostly favours ProtoNet or ties it, while Gaussian-noise effects remain small.Source: §5.5 · p. 14

5.6. RQ4: source-only shared-shift training

  • The shared pseudo-domain curriculum provides no evidence of improvement over ordinary WIPT: all six domain/shot paired intervals include zero.WIPT still exceeds shift-trained support-only by +0.82 points on 5-shot EuroSAT and +1.47 on ISIC, positive in all five runs.Source: §5.6 · p. 15

5.7. Sensitivity to adaptation and scoring

  • The replicated scorer grid finds no stronger absolute alternative to raw Euclidean scoring; relation-scored WIPT improves over support-only relation on CUB by +8.515±2.382 points but remains well below Euclidean accuracy.The reported configurations combine scorer and training choices rather than isolating a metric substitution.Source: §5.7 · p. 15

6. Discussion

The discussion frames WIPT as a conditional intervention, not a replacement for class averaging: its value varies with domain and support size. Additional query context shows no reliable accuracy gain, and the experiments characterize decision changes and limits without establishing a general mechanism or universally stronger configuration.

6.1. A conditional benefit, not a universal replacement

  • WIPT improves ProtoNet in every 1-shot run on CUB and EuroSAT but is consistently harmful on ISIC; in 5-shot, ProtoNet remains strongest overall, with WIPT’s clearest gain over support-only on ISIC.The study does not separate query-side adaptation from support-side conditioning.Source: §6.1 · p. 15

6.2. No detected accuracy benefit from additional query context

  • Jointly processing up to five queries yields no reliable cross-domain accuracy gain; larger groups reduce analytical attention work per query, but head-only latency is non-monotonic.The authors do not establish that grouped-query accuracies are equivalent; g = 1 supports independent, streaming-compatible inference.Source: §6.2 · p. 16

6.3. Localization and the balance of decision changes

  • WIPT changes uncertain ProtoNet decisions more often than confident ones, and weighting transition rates by their opportunity pools accounts exactly for the aggregate accuracy difference.More support coincides with lower centroid error and prototype variability, but a causal test would need to separate query and support adaptation pathways.Source: §6.3 · p. 16

6.4. What the negative extensions tell us

  • Corruption and source-shift controls limit the method’s reach, while scorer results show that query participation is not invariant to the classifier: Euclidean remains strongest overall.The tested cosine configuration performs poorly, but the authors caution against generalizing this to all cosine-prototype methods.Source: §6.4 · p. 17

6.5. Limitations and future work

  • The study covers one source dataset, three target domains, and one frozen ViT-S/16 backbone, rather than leaderboard comparisons with larger, fine-tuned, spatial, or transductive systems.The multi-query study uses at most five queries; the 3 × 3 scorer grid is fully replicated only in 5-shot, and the mixed-shift follow-up has one seed.Source: §6.5 · p. 17

7. Conclusion

WIPT is a fixed-encoder, single-query form of test-time prototype adaptation that helps in selected regimes, not uniformly. Broader tests show no reliable gain from multiple-query context and identify boundary-, source-shift-, and scorer-sensitive behavior.

7. Conclusion

  • WIPT improves ProtoNet in selected regimes, with its clearest gain on 1-shot EuroSAT, and consistently improves the matched support-only Transformer on 5-shot ISIC.The encoder and model parameters remain fixed while each target query interacts with the labelled supports in global embedding space.Source: §7 · p. 18
  • None of the tested variants processing more than one query provides a reliable accuracy gain over independent query processing, though the intervals do not establish equivalence.Source: §7 · p. 18
  • Decision changes concentrate near ProtoNet’s boundary, while outcomes depend on rescue-versus-break balance, source-shift training, and the downstream scorer.Source-only shift training shows no improvement over standard WIPT, and the scorer grid shows sensitivity to the downstream metric.Source: §7 · p. 18

Appendix A. WIPT design ablations

Design ablations favor the two-layer, class-mean, Euclidean WIPT-2, while scorer-grid controls show that comparisons depend on how cosine scoring is trained and aggregated.

Appendix A. WIPT design ablations

  • The ablations test depth, residual refinement, support dropout, attention temperature, query-weighted pooling, multi-prototype scoring, and margin training against WIPT-2.Trained variants reuse the frozen ViT-S/16 encoder and episodic protocol; inference-only variants reuse the WIPT-2 checkpoint.Source: Appendix A · p. 19
  • Four-layer WIPT reduces EuroSAT by 3.12 points, six-layer WIPT remains 0.57 points below WIPT-2, and no modification improves all three domains.Query-weighted pooling is almost neutral, inference-only cosine trades a small CUB gain for losses elsewhere, and multi-prototype scoring is consistently weaker.Source: Appendix A · pp. 19–20, Table A.1 · p. 19
  • The inference-only EuroSAT cosine result of 81.963% uses a Euclidean-trained WIPT-2 checkpoint and is not comparable to separately cosine-trained heads.The two evaluations also use different cosine aggregation rules.Source: Appendix A · p. 20

Appendix B. Five-seed adaptation–scorer factorial grid

  • The five-shot scorer grid crosses raw, support-only, and joint/WIPT adaptation with Euclidean, cosine, and learned relation scoring across five training seeds.Its factorial cosine averages similarities of normalized query and support vectors within each class; the relation scorer is a trained MLP, so the grid compares full scorer/training configurations.Source: Appendix B · pp. 20–21

Appendix B.1. Full condition results

The five-seed 5-shot grid finds Euclidean scoring strongest in every target domain, while WIPT’s effect varies substantially with the scorer. Cosine contrasts are consistently negative, whereas relation scoring has a replicated CUB benefit but less stable effects elsewhere.

Appendix B.1. Full condition results

  • The 5-shot grid evaluates raw, support-only, and joint/WIPT adaptation with three scorers on the same 1,500 matched target episodes per domain.Trainable conditions use five independent seeds; Table B.1 reports condition means and Table B.2 paired WIPT contrasts.Source: Appendix B.1 · p. 21, Table B.1 · p. 21, Table B.2 · p. 22
  • Euclidean scoring remains the strongest absolute formulation in every target domain, while WIPT’s effect depends strongly on the scorer.The grid is a secondary sensitivity analysis rather than an isolated metric substitution.Source: Appendix B.1 · p. 22
  • With cosine scoring, WIPT is lower than both raw and support-only variants in all five training runs on all three domains.The EuroSAT drop is especially large.Source: Appendix B.1 · p. 22
  • On CUB, WIPT exceeds support-only relation by +8.515 ± 2.382 points, with five of five runs positive.Its +2.051 ± 2.209 point mean advantage over raw relation is less stable, with four of five runs positive and an interval including zero.Source: Appendix B.1 · p. 22
  • EuroSAT relation contrasts are highly variable, and the positive-average ISIC WIPT–support relation contrast has an interval that includes zero.Source: Appendix B.1 · p. 22

Appendix C. Exploratory mixed pseudo-domain screening

These appendices extend the controlled comparison with exploratory source-shift and support-noise diagnostics, computational benchmarks, related-work distinctions, and a multi-query protocol. The additional analyses clarify the scope and trade-offs of the evaluation without establishing a general advantage for query-conditioned adaptation.

Appendix C. Exploratory mixed pseudo-domain screening

  • The single-seed mixed pseudo-domain screen shows no coherent large improvement: 1-shot changes switch sign, while 5-shot changes are positive but at most +0.233 points.The 5-shot results use one initialization and do not alter the five-seed RQ4 conclusion.Source: Appendix C · p. 23, Table C.1 · p. 23

Appendix D.1. Normalized prototype geometry

  • WIPT and ProtoNet have nearly identical mean normalized margins across CUB, EuroSAT, and ISIC, with matched episode margins highly correlated.WIPT does not systematically enlarge normalized margins beyond ProtoNet.Source: Appendix D.1 · p. 23, Figure D.1 · p. 24

Appendix D.2. Support-noise pathway diagnostic

  • The support-noise diagnostic holds WIPT’s clean-pass query representation fixed while recomputing prototypes from noisy supports, so it isolates support-to-prototype propagation rather than normal inference.At severity 5, relative margins are approximately 0.027 for ProtoNet and 0.008 for WIPT, with accuracies of 55.69% and 47.53%; the selected-checkpoint comparison does not establish the cause of the gap or variation across independently trained heads.Source: Appendix D.2 · pp. 24–25, Figure D.2 · p. 25

Appendix E. Additional computational-cost results

  • The head-only benchmark measures synthetic-embedding head operations, not image encoding, and its CUDA memory and latency results describe specific implementations rather than hardware-independent performance.Checkpoint-selection times indicate when the best validation model was obtained, not total training time or convergence speed.Source: Appendix E · p. 26, Table E.1 · p. 27, Table E.2 · p. 27

Appendix F.1. Metric and embedding adaptation in few-shot learning

  • The related-work comparison distinguishes support-centric prototype adaptation from WIPT’s joint query-support context, while noting that prior query-conditioned methods also use spatial or richer representations and additional mechanisms.The support-only Transformer is a custom control, not a reproduction of FEAT or PrototypeFormer.Source: Appendix F.1 · p. 28, Appendix F.2 · p. 28, Table F.1 · p. 29

Appendix F.3. Cross-domain few-shot learning and source-domain variation

  • The source-variation extension motivates exposing episodic heads to pseudo-domains while keeping the encoder frozen and using no target-domain images.The controlled pipeline trains support-only and WIPT heads episodically on miniImageNet, then evaluates them on matched target episodes without target updates.Source: Appendix F.3 · pp. 29–30, Appendix G · p. 30

Appendix H.1. Multi-query context comparison

  • The multi-query comparison groups unlabelled queries with the same supports, shuffles queries before grouping, and trains group-specific heads across five seeds.The 1-shot g = 3 and g = 5 follow-up was selected after inspecting the 5-shot sweep and was not a preregistered equivalence test.Source: Appendix H.1 · pp. 30–31

Appendix H.2. Computational cost

  • For 5-way 5-shot WIPT, each single-query Transformer sequence has 26 tokens; grouping queries amortizes support tokens, while the analytical token-pair count excludes projections and feed-forward layers.The head-only benchmark excludes the shared ViT cost.Source: Appendix H.2 · pp. 31–32

Appendix H.3. RQ2: does additional query context improve accuracy?

Jointly processing additional unlabelled queries yields no detected accuracy advantage over single-query WIPT across the tested shot regimes. The appendix also examines computational trade-offs and how decision changes relate to ProtoNet margins.

Appendix H.3. RQ2: does additional query context improve accuracy?

  • Across both shot regimes, additional query context provides no reliable accuracy advantage over independently processing one query.All five-seed intervals include zero; 5-shot changes for g = 2–5 range from −0.012 to −0.017 points on CUB, −0.459 to −0.502 on EuroSAT, and −0.130 to −0.141 on ISIC.Source: Appendix H.3 · p. 32, Table H.1 · p. 33
  • In 5-shot, g = 5 reduces analytical attention work from 101,400 to 27,000 token-pairs per episode and peak allocated memory from 75.8 to 53.6 MiB, but latency is non-monotonic.The 5-shot EuroSAT g = 5 interval remains compatible with changes from −1.536 to +0.536 points, so intervals containing zero do not establish equivalence.Source: Appendix H.3 · p. 32

Appendix I.1. Mechanism and boundary analysis

  • The boundary analysis sorts ProtoNet errors and successes separately by raw margin into equal-count bins, then measures rescue and break probabilities within each bin.Queries are also summarized by the fraction of rescues and breaks among the 20% with the smallest absolute ProtoNet margin; this is descriptive, not a tuned prediction rule.Source: Appendix I.1 · pp. 33–34
  • Equation 5 decomposes the aggregate accuracy difference exactly by weighting the rescue rate among ProtoNet errors and break rate among its successes by their respective opportunity pools.The pooled rates are computed from query counts and differ from an unweighted average of episode-wise conditional rates.Source: Appendix I.1 · p. 34

Appendix I.2. RQ3: where do the shot- and domain-dependent effects arise?

The analyses describe how support count and domain relate to WIPT’s decision changes, while distinguishing observed patterns from causal explanations. Controlled corruption produces regime-dependent comparisons, and the source-shift extension uses shared pseudo-domain transforms with a one-seed mixed variant reported only descriptively.

Appendix I.2. RQ3: where do the shot- and domain-dependent effects arise?

  • From 1-shot to 5-shot, normalized support-to-query centroid error and query-conditioned prototype variability both fall across CUB, EuroSAT, and ISIC.The changes are descriptive associations with shot count, not evidence that either statistic alone causes an accuracy improvement.Source: Appendix I.2 · p. 34
  • Rescue/break contributions explain the contrasting EuroSAT outcomes: 1-shot yields +2.068 points, while 5-shot yields −0.721 points as the ProtoNet error pool shrinks from 37.91% to 17.24%.On 5-shot CUB, 2,385 of 112,500 queries are ProtoNet errors, leaving a small negative net change despite rescuing 8.38% of baseline errors.Source: Appendix I.2 · pp. 34–35
  • Rescue and break probabilities are largest near zero ProtoNet margin and decline away from the boundary; the nearest 20% of queries contain most transitions across the six conditions.The boundary concentration does not guarantee a positive accuracy change, and the analysis excludes switches between two incorrect predictions.Source: Appendix I.2 · p. 35, Figure I.1 · p. 36
  • Episode-level associations between individual geometry metrics and WIPT gain are weak, with the largest absolute Pearson or Spearman coefficient only 0.146.The authors treat geometry measures as exploratory diagnostics and boundary concentration as a reproduced pattern, not an explanation of individual rescues and breaks.Source: Appendix I.2 · p. 35

Appendix J.1. Controlled corruption protocol

  • The EuroSAT corruption study applies brightness/contrast, Gaussian noise, or their combination to support and query images, using 1,000 episodes per category-severity condition.The replicated analysis focuses on EuroSAT because the initial diagnostic showed the strongest corruption interaction there.Source: Appendix J.1 · pp. 36–37, Table J.1 · p. 37

Appendix J.2. Controlled corruption response

  • Under combined EuroSAT corruption, WIPT’s mean difference from frozen ProtoNet crosses from −0.26 points at severity 3 to +0.75 at severity 4 and +1.02 at severity 5; all five runs are positive at severities 4 and 5.WIPT is positive on average against the matched support-only Transformer in every corruption condition, but does not uniformly outperform ProtoNet.Source: Appendix J.2 · p. 37, Figure J.1 · p. 38

Appendix K.1. Source-only shared-shift training extension

  • Shared-shift training applies one pseudo-domain transform to support and query in half of source episodes, with shift strength scheduled over 50 epochs and the encoder frozen.The mixed shared/cross pseudo-domain follow-up has only one training seed and is reported descriptively, not as a replicated claim.Source: Appendix K.1 · pp. 38–39

Appendix K.2. RQ4: source-only shared-shift training

The shared source pseudo-domain curriculum provides no evidence of improvement over standard WIPT, although WIPT still exceeds a shift-trained support-only control in several comparisons. The manuscript reports no specific funding and provides no public code repository with this version.

Appendix K.2. RQ4: source-only shared-shift training

  • All six shared-shift versus standard WIPT paired intervals include zero, so the curriculum provides no evidence of improvement on natural targets.The reported mean changes range from −0.399 to −0.055 points in 1-shot and from −0.419 to +0.018 in 5-shot.Source: Appendix K.2 · p. 39, Table K.1 · p. 39
  • At 5-shot, WIPT exceeds shift-trained support-only by +0.82 points on EuroSAT and +1.47 on ISIC, with all five runs positive.This comparison concerns query participation relative to a matched transformed-support head, not improvement over ordinary WIPT.Source: Appendix K.2 · p. 39

Funding

  • The research received no specific grant from public, commercial, or not-for-profit funding agencies.Source: Funding · p. 40

Data and code availability

  • The authors retain implementation, trained-head checkpoints, and experimental records, but provide no public code repository with this manuscript version.Source: Data and code availability · p. 40

Related papers

Loading 2609.30769v1…