acceptodds
Under review as a conference paper at ICLR 2027

Class-Adaptive Characteristic Function Matching for Cross-Domain Few-Shot Learning

Abstract

Under domain shift, successful few-shot recognition requires not only transferable source-trained features but also a comparison rule appropriate for novel target classes. Most metric-based approaches to cross-domain few-shot learning (CD-FSL) construct class prototypes from target support examples, yet evaluate all queries using a shared Euclidean or cosine metric. Target supervision therefore adapts the class representatives but leaves the comparison function unchanged across classes. We address this asymmetry with Class-Adaptive Characteristic Function Matching, a plug-and-play metric module that uses target support examples to adapt both class representation and query-class comparison. Our key observation is that the frequency arguments of an empirical characteristic function can parameterize the comparison rule: each frequency induces a nonlinear response to a feature projection, while a finite set of frequencies forms a compact representation of the support class. For each candidate class, a lightweight KL-regularized stochastic generator conditions on its support set and produces a compact set of class-specific frequency arguments. Given a query, we tentatively include it in each candidate support set and measure the resulting change between the original and query-augmented empirical characteristic-function representations. This change serves as the membership score, allowing the same query to be evaluated through a distinct support-conditioned comparison for every novel target class. The proposed module can replace conventional point-based similarities without modifying the feature backbone or the target-domain adaptation procedure of the host method. Experiments across four CD-FSL benchmarks, three base learners, and multiple pretrained backbones demonstrate consistent improvements with only a small parameter overhead. Code and models will be publicly released.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Characteristic Function Matching, Meta-learning, Cross-Domain Few-Shot Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TBpoOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper targets the asymmetry in CD-FSL where support examples shape class prototypes but leave the shared comparison rule unchanged. Each candidate class is represented by finite evaluations of its empirical characteristic function (CF) at class-specific frequency arguments, generated by a KL-regularized encoder-decoder from the class mean feature. A query is scored by the change in the CF representation when tentatively added to the support set (smaller change = higher membership).

Critical strengths
  1. The motivation is reasonable. The method of conducting class-adaptive comparison is a sensible conceptual contribution, distinct from task-level metric adaptation (TADAM) and feature adaptation.
  2. Eq. 23–24 shows the query-induced change score reduces to support-size-weighted direct Fourier matching up to a class-independent constant under the standard balanced protocol.
  3. The experimental results validate the effectiveness of the proposed method.
Critical weaknesses
  1. The motivation is asserted rather than demonstrated, and the presentation makes the paper hard to follow. The core claim, i.e., different target classes need different comparison rules, has no motivating experiment, failure analysis, or visualization. The leap to characteristic functions is abrupt, with no explanation of why Fourier frequencies beat per-class linear/Mahalanobis projections or task-conditioned metrics, and no connection to the CF-distance/MMD two-sample-testing literature.
  2. The core algorithmic choices is not well-justified (i) Why frequencies at all? The score is a learned random-Fourier-feature kernel distance, yet the paper never states which kernel it induces. How the kernel selection influence the performance? (ii) Why stochastic generation? Inference uses a single sample; the obvious baseline (deterministic decoder on μ_c) is not ablated, and the ablated deterministic variant uses a different parameterization, confounding stochasticity with parameterization — its collapse (44.87% vs. 52.14% for even fixed random frequencies) suggests an optimization pathology, not a benefit of stochasticity.
Questions for authors

It is better to visualize the learned per-class frequency distributions what does query inclusion add beyond narrative given Eq. 24; how were L and λ selected and at what label cost; is the unbalanced-case temperature bias intended and how does the method perform on unbalanced episodes; what kernel/divergence does the score correspond to.

14 comments
AuthorsOpenReview ↗

Response to Reviewer TBpo (Part-1)

W1a. Regarding motivation.

Response:

1. Why characteristic functions (CF)?
Under severe cross-domain shifts, feature distributions across target classes are highly abstract and non-linear. A single spatial metric (e.g., Euclidean or cosine distance) struggles to capture these distributional differences. Characteristic functions (via Fourier frequencies) allow us to compare distributions spectrally, naturally capturing the multi-scale and global discrepancies that are critical for robust cross-domain comparison.

2. Why class-adaptive rather than global/task-conditioned?
Different target classes often exhibit distinct distributional shifts relative to the source domain. A globally shared or task-conditioned rule implicitly assumes that one single metric is optimal for all classes within a task. Our ablation study (Table 3) directly tests this assumption:

MethodSettingChestXISICEuroSATCropDiseaseAverage
Baseline1-shot22.5733.2972.4281.0252.33
Fixed Random Frequency1-shot22.7334.0671.4680.2152.14
Ours w/o Class-Adaptive CF Matching1-shot22.7634.6272.7481.0752.79
Ours (Full)1-shot22.7834.6973.0282.3553.21
Baseline5-shot25.7648.6388.8894.7564.51
Fixed Random Frequency5-shot25.9748.9088.6294.0464.38
Ours w/o Class-Adaptive CF Matching5-shot23.7749.4287.7394.7663.92
Ours (Full)5-shot26.4251.8989.1296.0765.88

As shown, a single fixed frequency set (Fixed Random Frequency) yields 52.14/64.38, failing to improve over the baseline. Even when frequencies are learnable but shared across all classes (Ours w/o Class-Adaptive CF Matching), the average remains at 52.79/63.92. Only when the frequency rule is conditioned on individual support classes (Ours) do we observe the best performance (53.21/65.88). This confirms that different classes indeed require distinct comparison rules.


W1b. Why Fourier comparisons instead of per-class linear/Mahalanobis projections or task-conditioned metrics?

Response:

1. What Fourier mapping adds. If we simply apply a linear projection and average the features, the comparison relies solely on the support mean. CF matching is different: it averages the periodic nonlinear responses of individual features. Because it incorporates both direction and scale sensitivity, this mechanism can capture distributional information beyond just the mean in multi-shot settings.

2. Experimental comparison. To directly answer this, we conduct experiments to compare our method with two alternative metrics on the ProtoNet baseline:

  • Mahalanobis: We compute the mean and covariance of each support class, and predict the query based on the minimum Mahalanobis distance.

  • MMD: We compute the Maximum Mean Discrepancy between the query and each support class, and predict the closest one.

  • Task-conditioned: This corresponds to the Ours w/o Class-Adaptive CF Matching row in our ablation study, representing a shared rule for all classes.

    The results are summarized below:

    MethodSettingChestXISICEuroSATCropDiseaseAverage
    ProtoNet (Baseline, Euclidean)5-shot25.7648.6388.8894.7564.51
    Mahalanobis Distance5-shot26.5848.8689.1195.5065.11
    MMD Distance5-shot26.5048.5689.0094.8464.73
    Task-conditioned (Ours w/o Class-Adaptive)5-shot23.7749.4287.7394.7663.92
    Ours (Class-adaptive CF)5-shot26.4251.8989.1296.0765.88

As shown, while Mahalanobis and MMD provide competitive results compared to the Euclidean baseline, our Class-adaptive CF achieves the best overall average performance.

AuthorsOpenReview ↗

Response to Reviewer TBpo (Part-2)

W1c / W2(i) / question on kernel and divergence.

Response:

1. The induced kernel and MMD connection.
For a fixed support set, our method essentially induces a shift-invariant kernel. It compares two samples by computing the average cosine similarity of their feature differences at various frequencies.

This score is mathematically equivalent to computing the MMD between the query sample and the empirical distribution of the support class. The only subtle difference is that our formulation is a kernel mean distance (which includes the support self-similarity term on the diagonal), rather than an unbiased two-sample-test statistic.

2. Connection to CF distances.
The MMD induced by our kernel can be proven to be equivalent to integrating the differences of characteristic functions (CFs) over the frequency space.

We will explicitly add this classical identity to the manuscript. We also note that while a finite bank of sampled frequencies does not guarantee unique identification of the true distribution, our empirical results demonstrate that it serves as a highly robust approximation for cross-domain comparison.

3. Effect of kernel selection.
To address the impact of kernel choice, we directly compare our class-conditioned spectral kernel with a task-conditioned variant (where the spectrum is shared across all classes).

Kernel / spectral constructionSettingChestXISICEuroSATCropDiseaseAverage
Task-conditioned spectral kernel1-shot22.7634.6272.7481.0752.79
Class-conditioned spectral kernel (Ours)1-shot22.7834.6973.0282.3553.21
Task-conditioned spectral kernel5-shot23.7749.4287.7394.7663.92
Class-conditioned spectral kernel (Ours)5-shot26.4251.8989.1296.0765.88

As shown, conditioning the kernel on individual classes yields consistent improvements over the task-conditioned variant.


W2(ii). Deterministic control.

Response:

Yes, the original low-rank ablation changes the parameterization and does not isolate stochasticity.

  1. Following your suggestion, we added a mean-only control. Each class prototype is passed through the encoder to obtain its latent mean, which is then fed into the MLP decoder to generate frequencies. The variance branch and Gaussian sampling are removed. This directly tests the requested deterministic path. We will replace the stronger interpretation of the original low-rank ablation with this comparison.

  2. The mean-only control achieves 42.36/59.77 average accuracy, compared with 53.21/65.88 for CACFM. This supports sampled frequency generation over this mean-only construction. It does not establish that every deterministic generator must fail, nor does it by itself identify the specific cause of the degradation.

GeneratorSettingChestXISICEuroSATCropDiseaseAverage
Mean-only encoder-decoder1-shot21.4828.0356.5363.3942.36
CACFM (Ours)1-shot22.7834.6973.0282.3553.21
Mean-only encoder-decoder5-shot24.5041.7181.6291.2459.77
CACFM (Ours)5-shot26.4251.8989.1296.0765.88

W2(ii)b. What does stochasticity provide if inference uses a single sample?

Response:

The stochasticity fundamentally arises from how the frequency bank is constructed. As shown in Eqs. 11–15, the generator does not directly predict the frequencies. Instead, it takes the support features, outputs the parameters (mean and variance) of a Gaussian distribution, and samples a latent vector from this distribution via reparameterization.

Even in the 1-shot setting, the single support image provides the conditions for this distribution. The stochasticity provides a mechanism to model the uncertainty/variability of the spectral space around that single support prototype. At inference, we sample one bank of L=1024 independent frequency instances from this class-conditioned distribution. This finite bank acts as a Monte Carlo approximation of the spectral expectation, and it is reused across all queries for that class. This avoids deterministically collapsing the high-dimensional frequency matrix into a single point, which our ablation (Table 4) shows is highly prone to optimization pathology.

AuthorsOpenReview ↗

Response to Reviewer TBpo (Part-3)

Q1. Visualize the learned per-class frequency distributions.

Response:

Thank you for your suggestion. We are currently running this experiment, and once it is completed, we will add the results to the revised manuscript (Note that since OpenReview does not support image uploads in this text box.).


Q2. What does query inclusion add beyond the narrative, given Eq. 24?

Response:

Mathematically, query inclusion adds no independent classification mechanism beyond Eq. 24 (or Eq. 23) under the balanced shot setting. As shown in the derivation, both formulations yield exactly the same class ranking (up to a constant factor).

The query inclusion serves as a conceptual set-update interpretation: it frames the comparison as evaluating how much the query perturbs the empirical characteristic function of the class. The actual comparison remains entirely determined by the learned class-conditioned frequencies.


Q3. How were L and lambda selected, and at what label cost?

Response:

This hyperparameter tuning relies solely on the target support set. The selection process does not access any target query labels. The reported target evaluation results are strictly produced after the parameters are fixed, following standard CD-FSL protocol. The label cost is therefore identical to common settings (only target support examples are involved).

Q4. Unbalanced support sets.

Response:

In our paper, to ensure a fair and consistent comparison with existing state-of-the-art methods, we primarily adopted the standard balanced settings (e.g., 1-shot and 5-shot) in our main experiments. Following your suggestion, we further evaluated our method under unbalanced support settings. We applied the same unbalanced setting to both the baseline and our method for a fair comparison. The results are summarized below.

MethodChestXISICEuroSATCropDiseaseAverage
Euclidean ProtoNet26.8947.7981.0387.7960.88
CACFM (Ours)25.9248.0382.3091.8262.02

Overall, our method outperforms the baseline under unbalanced settings, demonstrating its robustness.


W1d. Presentation

Response:

Thank you for this suggestion. We will revise our writing to make it easier to follow. The most critical changes will be clarifying the preliminaries and presenting our motivation more intuitively from a mathematical perspective. We will update our paper accordingly. Thank you very much.

Reviewer_xv3POpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper considers cross-domain few-shot learning (CD-FSL) and argues that target support examples should adapt both the class prototypes and the query–class comparison rule. The paper proposes Class-Adaptive Characteristic Function Matching, a plug-and-play module that generates class-specific frequency arguments from each support set using a KL-regularized stochastic generator. These frequencies define a class-specific characteristic-function representation, and query-class compatibility is measured by the representation change caused by adding the query to each candidate support set. Experiments on ChestX, ISIC, EuroSAT, and CropDisease show generally consistent improvements across ProtoNet, PMF, and CD-CLS, as well as across DINOv1, DINOv3, and CLIP backbones, although many of the improvements are small.

Critical strengths
  1. The paper is clearly motivated and proposes a plug-in module that adapts the query-class comparison rule without modifying the feature backbone or the target-domain adaptation procedure of the host method.
  2. The proposed module consistently improves average performance across a broad evaluation covering three host methods (ProtoNet, PMF, and CD-CLS), four target domains (ChestX, ISIC, EuroSAT, and CropDisease), multiple backbones (DINOv1, DINOv3, and CLIP), and different support-set sizes.
Critical weaknesses

The novelty and empirical advantage over related methods are not sufficiently established. MetaVRF and MetaKernel similarly learn spectral distributions for few-shot classification, while the proposed method generates frequencies separately for each class rather than for each task. However, direct experimental comparisons are missing, making the benefit of this distinction unclear. In addition, the module is evaluated mainly with ProtoNet, PMF, and CD-CLS; PMF is from 2022 and CD-CLS is from 2024. Evaluation with at least one more recent CD-FSL method would better demonstrate the relevance of the proposed plug-in. Finally, although the proposed module generally improves average performance, many of the individual improvements over the corresponding baselines are small.

Questions for authors
  1. How does the proposed method compare with related spectral methods such as MetaVRF and MetaKernel? These methods similarly learn spectral distributions for few-shot classification, while the proposed method generates frequencies separately for each class rather than for each task. A direct comparison would help clarify whether class-specific frequency generation provides a meaningful advantage.
  2. Could the authors evaluate the proposed module with a more recent CD-FSL method? The current experiments use ProtoNet, PMF, and CD-CLS. Adding the module to a recent method would better demonstrate whether it remains beneficial with stronger CD-FSL approaches.
3 comments
AuthorsOpenReview ↗

Response to Reviewer xv3P (Part-1)

W1 / Q1. Related spectral methods.

Response:

1. Methodological distinction. MetaVRF and MetaKernel learn task-specific spectral kernels. In contrast, our method generates frequencies for each support class and classifies by the resulting kernel mean distance. Our core contribution is this class-conditioned construction for CD-FSL, rather than spectral learning itself.

2. Experimental comparison. The first two rows in the table below compare our method against the complete MetaVRF and MetaKernel frameworks. Since these methods were originally proposed in different domains and settings, their spectral inference and classification mechanisms were designed for their specific contexts. To ensure a fair comparison, we adapted and applied their original mechanisms to the same cross-domain few-shot learning (CD-FSL) setting under a matched backbone.

MethodSettingConditioningClassification ruleChestXISICEuroSATCropDiseaseAverage
MetaVRF1-shotTaskOriginal kernel learner22.4734.6971.9782.5952.93
MetaKernel1-shotTaskOriginal kernel learner22.3734.8271.7882.4252.85
Ours w/ task-conditioned control1-shotTaskFourier mean matching22.7634.6272.7481.0752.79
Ours1-shotClassFourier mean matching22.7834.6973.0282.3553.21
MetaVRF5-shotTaskOriginal kernel learner26.7850.2888.5095.0065.14
MetaKernel, matched-backbone reproduction5-shotTaskOriginal kernel learner26.7150.2688.4495.0165.10
Ours w/ task-conditioned control5-shotTaskFourier mean matching23.7749.4287.7394.7663.92
Ours5-shotClassFourier mean matching26.4251.8989.1296.0765.88
AuthorsOpenReview ↗

Response to Reviewer xv3P (Part-2)

W2 / Q2. Integration with a more recent CD-FSL method.

Response:

Following your suggestion, we integrated our proposed module into a recent state-of-the-art CD-FSL baseline in a plug-and-play manner. We just checked the run and it is still in the training phase. We will update the results here immediately once the training is complete.

AuthorsOpenReview ↗

Response to Reviewer xv3P (Part-3)

W3. Improvements over the baselines.

Response:

1. Magnitude and Consistency.

First, we would like to emphasize that the improvements are highly consistent. Across the three primary baselines, the domain-averaged gains range from 0.69 to 2.74 percentage points, and our method improves 23 out of 24 host/domain/shot combinations.

Host / MethodBackbone1-shotGain5-shotGain
ProtoNet (Baseline)DINOv1 / ViT-Small52.33-64.51-
ProtoNet + OursDINOv1 / ViT-Small53.21+0.8865.88+1.37
PMF (Baseline)DINOv1 / ViT-Small50.91-64.08-
PMF + OursDINOv1 / ViT-Small53.31+2.4066.82+2.74
CD-CLS (Baseline)DINOv1 / ViT-Small54.62-67.54-
CD-CLS + OursDINOv1 / ViT-Small55.31+0.6968.42+0.88

Second, we also tested our method on stronger backbones. As shown below, we get consistent gains across all domains and shots on DINOv3 and CLIP, showing that our approach works well even with stronger pretrained features.

Host / MethodBackbone1-shotGain5-shotGain
ProtoNet (Baseline)DINOv3 / ViT-Base54.30-64.47-
+ OursDINOv3 / ViT-Base55.32+1.0266.86+2.39
ProtoNet (Baseline)CLIP / ViT-Base48.64-59.55-
+ OursCLIP / ViT-Base51.35+2.7163.96+4.41

Third,  we accept the reviewer's suggestion and will explicitly state this limitation in the revised manuscript.

2. Statistical Evidence.

To check robustness, we report the mean accuracy with 95% confidence intervals over 600 episodes. For the ProtoNet backbone (DINOv1), our method consistently improves the mean accuracy across all domains and both shot settings without widening the confidence intervals. This shows the gains are reliable and do not come with higher variance.

MethodSettingChestXISICEuroSATCropDiseaseAverage
Baseline (ProtoNet)1-shot22.57 ± 0.4133.29 ± 0.5672.42 ± 0.5081.02 ± 0.5152.33
+ Ours1-shot22.78 ± 0.4234.69 ± 0.5873.02 ± 0.4982.35 ± 0.4753.21
Baseline (ProtoNet)5-shot25.76 ± 0.4548.63 ± 0.6088.88 ± 0.4594.75 ± 0.4164.51
+ Ours5-shot26.42 ± 0.4451.89 ± 0.5189.12 ± 0.4196.07 ± 0.3465.88
Reviewer_dgPnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Class-Adaptive Characteristic Function Matching (CACFM), a plug-and-play metric module for cross-domain few-shot learning (CD-FSL). The proposed method aims to adapt the comparison function to each target class instead of adapting class representations, by using empirical characteristic function. Through extensive experiments on four target-domain benchmarks—ChestX, ISIC, EuroSAT, and CropDisease—using three host methods (ProtoNet, PMF, and CD-CLS), the results show consistent improvements in average accuracy across the three host methods.

Critical strengths
  1. The proposed CACFM module consistently improves the performance of three mainstream CD-FSL methods (ProtoNet, PMF, and CD-CLS) across four benchmarks, demonstrating its effectiveness and compatibility with diverse existing approaches.
Critical weaknesses
  1. The distinction between class representation learning and comparison-function adaptation remains unclear. In my view, the proposed method still primarily focuses on class representation learning, with the comparison function derived from the empirical characteristic-function representation. The authors should clarify what fundamentally distinguishes their approach from existing representation-based methods. Furthermore, the rationale for choosing characteristic functions over alternative representations is insufficiently justified; the discussion in Lines 83–90 does not convincingly establish their unique advantages for cross-domain few-shot learning.
Questions for authors

The proposed method has not yet fully convinced me of its core methodological motivation. While I acknowledge the reported experimental improvements. A satisfactory answer to the following two questions would substantially influence my overall assessment.

  1. What is the fundamental advantage of characteristic-function-based representations for cross-domain few-shot learning (CD-FSL)? Specifically, why are characteristic functions better suited to capturing class-discriminative information under domain shifts than alternative representations or conventional similarity metrics? Can the authors provide further theoretical or empirical evidence to justify this design choice?

  2. In the era of large pretrained models, what role should large models play in few-shot learning? Given the increasingly strong generalization capabilities of pretrained vision and vision-language models, what specific limitations of these models does CACFM address?

5 comments
AuthorsOpenReview ↗

Response to Reviewer dgPn (Part-1)

W1a. Representation and comparison

Response:

Yes, our work (CACFM) adapts the class representation. It also conditions the mapping of the query on the candidate class.

  1. What is adapted. CACFM (ours) adapts both the class representation and the comparison rule. The key is that the mapping applied to the query depends on the candidate support class, rather than being shared across classes.
  2. How the comparison changes. Each class generates its own frequencies, inducing the kernel kc(x,y)=1L∑ℓ=1Lcos⁡((ωℓc)T(x−y))k_c(x,y)=\frac{1}{L}\sum_{\ell=1}^{L}\cos\left((\omega_{\ell}^{c})^T(x-y)\right). The same query is therefore compared with different classes under different support-conditioned kernels. Representation and comparison adaptation need not be mathematically disjoint.
  3. Existing evidence. As validated in our ablation study (reproduced below), the "globally shared learnable frequencies" corresponds to the variant "Ours w/o Class-Adaptive CF Matching". Our method consistently outperforms this variant across all settings. This supports class conditioning over a global rule. | Method | Setting | ChestX | ISIC | EuroSAT | CropDisease | Average | | --- | --- | --- | --- | --- | --- | --- | | Baseline | 1-shot | 22.57 | 33.29 | 72.42 | 81.02 | 52.33 | | Ours w/ globally shared learnable frequencies | 1-shot | 22.76 | 34.62 | 72.74 | 81.07 | 52.79 | | Ours | 1-shot | 22.78 | 34.69 | 73.02 | 82.35 | 53.21 | | Baseline | 5-shot | 25.76 | 48.63 | 88.88 | 94.75 | 64.51 | | Ours w/ globally shared learnable frequencies | 5-shot | 23.77 | 49.42 | 87.73 | 94.76 | 63.92 | | Ours | 5-shot | 26.42 | 51.89 | 89.12 | 96.07 | 65.88 |

W1b / Q1. Why characteristic functions?

Response:

  1. Why this representation. Empirical characteristic functions (CFs) average bounded nonlinear responses over support examples without fitting a density model. Their frequencies control projection direction and variation scale, providing a direct way to condition the comparison on each target class.
  2. Mathematical interpretation. Define ϕc(x)=L−1/2[cos⁡(ωℓcTx),sin⁡(ωℓcTx)]ℓ=1L\phi_c(x) = L^{-1/2}[\cos(\omega_\ell^{cT} x), \sin(\omega_\ell^{cT} x)]_{\ell=1}^{L} and Pc=Kc−1∑iδ(fic)P_c = K_c^{-1} \sum_i \delta(f_i^c). Then ∥ϕc(fq)−1Kc∑iϕc(fic)∥2=MMD2(δ(fq),Pc)\Vert \phi_c(f_q) - \frac{1}{K_c} \sum_i \phi_c(f_i^c) \Vert^2 = \text{MMD}^2(\delta(f_q), P_c). CACFM (ours) thus learns a support-conditioned comparison kernel. This connection will be added to Lines 83-90.
  3. Existing evidence. Fixed random frequencies alone achieve 52.14/64.38, which merely reflects the performance of applying static Fourier features without class adaptation. In contrast, our Class-Adaptive CF Matching reaches 53.21/65.88. The consistent improvements across all datasets demonstrate that the gains are not explained by fixed Fourier features alone, but fundamentally stem from our class-adaptive mechanism. | Method | Setting | ChestX | ISIC | EuroSAT | CropDisease | Average | | --- | --- | --- | --- | --- | --- | --- | | Baseline | 1-shot | 22.57 | 33.29 | 72.42 | 81.02 | 52.33 | | Fixed Random Frequencies | 1-shot | 22.73 | 34.06 | 71.46 | 80.21 | 52.14 | | Ours | 1-shot | 22.78 | 34.69 | 73.02 | 82.35 | 53.21 | | Baseline | 5-shot | 25.76 | 48.63 | 88.88 | 94.75 | 64.51 | | Fixed Random Frequencies | 5-shot | 25.97 | 48.90 | 88.62 | 94.04 | 64.38 | | Ours | 5-shot | 26.42 | 51.89 | 89.12 | 96.07 | 65.88 |
AuthorsOpenReview ↗

Response to Reviewer dgPn (Part-2)

Q2. What limitation remains with large pretrained models?

Response:

1. Role of pretrained models. Large pretrained models (e.g., VLMs) excel at providing universal visual representations and offer excellent feature initialization for downstream tasks. However, their focus is primarily on general vision. Limitations remain: (1) A strong general feature space does not guarantee that a single fixed metric (e.g., standard Euclidean or cosine distance) is optimal for every novel class, especially in specialized domains like remote sensing, agriculture, or medical imaging; and (2) with only extremely limited visual support samples, fully fine-tuning large models (e.g., 4B or 8B VLMs) is highly impractical.

2. How CACFM complements. Our method is complementary to these large models. Specifically, CACFM explores how to data-efficiently adapt these pre-trained features to challenging cross-domain scenarios. By generating class-adaptive frequencies, it learns a tailored comparison rule directly from the target support supervision, compensating for the limitations of a fixed metric.

3. Choice of backbones and existing evidence. Because of the strict sample constraints, we evaluate our method on foundational models of practical scale (CLIP and DINO series) rather than massive 4B+ VLMs. As shown in Submitted Table 4, our method yields consistent gains across every domain/shot combination on DINOv3 and CLIP. Notably, the gains are even larger on CLIP (+4.41 on 5-shot), establishing strong compatibility and demonstrating the necessity of adapting the comparison rule for downstream cross-domain tasks.

Pretraining / backboneBaseline 1-shotOurs 1-shotGainBaseline 5-shotOurs 5-shotGain
DINOv1 / ViT-Small52.3353.21+0.8864.5165.88+1.37
DINOv3 / ViT-Base54.3055.32+1.0264.4766.86+2.39
CLIP / ViT-Base48.6451.35+2.7159.5563.96+4.41

Subject: Regarding the ongoing discussion and any remaining questions

Dear Reviewer,

As the discussion deadline approaches, we would appreciate hearing whether you have any remaining questions or concerns about our paper or rebuttal. We would be happy to provide further clarification or additional information.

Thank you again for your time and thoughtful feedback!

Best regards,

The Authors

Comments on the paper

10 comments

Dear Reviewer, as the discussion deadline approaches, we would appreciate hearing whether you have any remaining questions or concerns about our paper or rebuttal. We would be happy to provide further clarification or additional information. Thank you again for your time and thoughtful feedback!

Dear Reviewer, as the discussion deadline approaches, we would appreciate hearing whether you have any remaining questions or concerns about our paper or rebuttal. We would be happy to provide further clarification or additional information. Thank you again for your time and thoughtful feedback!

Thank you for the additional experiments and analyses. The comparisons with MetaVRF, MetaKernel, and the task-conditioned control substantially address my concern. The additional results also show consistent improvements across host methods and backbones, although some individual gains remain small. It appears that the evaluation with a recent CD-FSL method is still in progress. I will maintain my current score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.