- 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
- 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.
- 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.
- The experimental results validate the effectiveness of the proposed method.
Critical weaknesses
- 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.
- 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
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 construction | Setting | ChestX | ISIC | EuroSAT | CropDisease | Average |
|---|---|---|---|---|---|---|
| Task-conditioned spectral kernel | 1-shot | 22.76 | 34.62 | 72.74 | 81.07 | 52.79 |
| Class-conditioned spectral kernel (Ours) | 1-shot | 22.78 | 34.69 | 73.02 | 82.35 | 53.21 |
| Task-conditioned spectral kernel | 5-shot | 23.77 | 49.42 | 87.73 | 94.76 | 63.92 |
| Class-conditioned spectral kernel (Ours) | 5-shot | 26.42 | 51.89 | 89.12 | 96.07 | 65.88 |
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.
-
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.
-
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.
| Generator | Setting | ChestX | ISIC | EuroSAT | CropDisease | Average |
|---|---|---|---|---|---|---|
| Mean-only encoder-decoder | 1-shot | 21.48 | 28.03 | 56.53 | 63.39 | 42.36 |
| CACFM (Ours) | 1-shot | 22.78 | 34.69 | 73.02 | 82.35 | 53.21 |
| Mean-only encoder-decoder | 5-shot | 24.50 | 41.71 | 81.62 | 91.24 | 59.77 |
| CACFM (Ours) | 5-shot | 26.42 | 51.89 | 89.12 | 96.07 | 65.88 |
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.
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.
| Method | ChestX | ISIC | EuroSAT | CropDisease | Average |
|---|---|---|---|---|---|
| Euclidean ProtoNet | 26.89 | 47.79 | 81.03 | 87.79 | 60.88 |
| CACFM (Ours) | 25.92 | 48.03 | 82.30 | 91.82 | 62.02 |
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.
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:
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:
As shown, while Mahalanobis and MMD provide competitive results compared to the Euclidean baseline, our Class-adaptive CF achieves the best overall average performance.