{"id":"e753888b-4eda-442d-81ca-6b2a23c6d9b6","arxiv_id":"2506.03461","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A soft K-means prototype classifier with adaptive receptive fields is claimed to keep near-perfect few-shot accuracy under 60% label noise, but the paper omits code, hyperparameters, and pretraining disclosure.","lead":"This paper proposes RoNFA, a few-shot image classifier that builds class prototypes with soft clustering and then classifies queries using adaptive neural receptive fields. It reports near-perfect accuracy on MiniImageNet and TieredImageNet even when 60% of support-set labels are wrong.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frozen-ViT test-class leakage: MiniImageNet/TieredImageNet are ImageNet subsets, so the near-perfect clean and noisy accuracies likely reflect pretraining overlap, not the proposed robust classifier.","rationale":"The reader's weakest_assumption centers on the unsupported claim that soft K-means is insensitive to label noise. I agree that Section 4.1 gives no proof or sensitivity analysis, and at 60% symmetric noise each class has about two correct labels. However, that concern alone does not explain the magnitude of the reported result: even an ideal noise-robust prototype cannot produce 99.17% clean MiniImageNet accuracy under a genuinely disjoint few-shot protocol with a frozen generic ImageNet ViT. The most load-bearing threat is that the frozen ViT was pretrained on data containing the MiniImageNet and TieredImageNet test classes, since both benchmarks are ImageNet subsets and the paper never discloses the pretraining corpus or excludes overlap. This would make 5-way 5-shot classification near-trivial for the backbone and would make soft K-means appear robust regardless of its label-noise handling. The reader did flag this overlap as a strong hint in the rationale, which is why I mark agreement as partial, but it was not listed as the weakest assumption. My concrete test is designed to settle this directly: if a class-disjoint or self-supervised ViT collapses the accuracy, the central claim is invalidated; if the accuracy persists, the robustness claim would be credible and worth investigating further. No code, no hyperparameters, and the internal VCIM/RoNFA naming inconsistency further impede verification, but the leakage check is the decisive one.","tokens_in":11297,"tokens_out":2616,"duration_ms":30918,"concrete_test":"Obtain the exact ViT variant and pretraining checkpoint (e.g., ViT-B/16 pretrained on ImageNet-21k) and determine whether its training set contains the MiniImageNet/TieredImageNet test classes. Then rerun the same RoNFA pipeline using a frozen ViT pretrained on a class-disjoint ImageNet split (or a self-supervised ViT trained without labels on data from which the benchmark classes are withheld), and compare clean and noisy 5-way 5-shot accuracy. If accuracy drops from ~99% to typical few-shot levels (~60-75%), or if it requires fine-tuning on non-overlapping base classes to recover, the reported robustness is largely leakage-driven. As a supplementary check, report per-class accuracy on the 20/160 test classes to expose leakage artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on Tables 1-4: RoNFA achieves 99.17% clean and 98.33% at 60% symmetric noise on MiniImageNet, and 95.88%/90.57% on TieredImageNet, using a frozen, not fine-tuned ViT (Section 5.1.3). MiniImageNet is explicitly an ImageNet subset [25] and TieredImageNet is also built from ImageNet [18]. Standard pretrained ViT checkpoints [2] are trained on ImageNet-1k or ImageNet-21k, so the encoder has almost certainly seen the benchmark test classes. Under this leakage, the 5-way 5-shot task is no longer few-shot generalization for this backbone: features are already class-discriminative for the test classes, and soft K-means trivially appears noise-robust. The paper never identifies the exact ViT checkpoint, its pretraining corpus, or any exclusion of overlapping classes, and no code is provided. The comparison against Conv4-based methods in Tables 1-4, and the claim to surpass clean-trained SOTA, therefore do not establish that the proposed neural-field mechanism is what drives the results; they may only confirm pretraining leakage. This is a threat to the evidence base, not merely to experimental hygiene.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RoNFA (also called VCIM), a two-neural-field classifier for few-shot image classification with noisy labels. It uses a frozen ViT feature extractor, soft K-means clustering initialized with per-class support means to obtain class representatives, and a Mexican-hat receptive-field activation with scale adaptation during prediction. Experiments on MiniImageNet and TieredImageNet under symmetric, paired, and outlier label noise report clean accuracies above 99% and 95%, with only small drops at 60% noise, and the authors claim that RoNFA outperforms state-of-the-art FSL methods and even surpasses clean-trained SOTA. The central claim is that the proposed neural-field mechanism provides strong robustness to label noise in the few-shot setting.","tokens_in":11554,"tokens_out":10360,"duration_ms":109841,"significance":"If the reported results were valid, RoNFA would be a notable contribution: a non-backpropagation classifier that maintains roughly 99% accuracy on MiniImageNet at 60% support-label noise would be practically relevant and scientifically interesting. The paper has several strengths: it examines three distinct noise models, includes ablations for soft K-means and scale adaptation, and reports confidence intervals over 600 episodes. However, the evidence base is seriously compromised by the likely overlap between the frozen ViT's pretraining data and the evaluation classes; the paper provides no code, no hyperparameter values, and no analysis of the K-means robustness claim. These issues currently outweigh the strengths, and the manuscript does not establish its central claims.","major_comments":[{"comment":"The frozen ViT is almost certainly pretrained on ImageNet, while MiniImageNet [25] and TieredImageNet [18] are ImageNet subsets. The paper does not identify the exact checkpoint, its pretraining corpus, or whether overlapping classes were excluded. The near-perfect accuracies in Tables 1-4 may therefore reflect test-class leakage rather than the proposed classifier's few-shot or noise-robust behavior. Please re-run all experiments with a feature extractor whose pretraining data excludes the benchmark classes (for example, trained only on each benchmark's base split), disclose the checkpoint identity, and report a class-overlap analysis.","section":"Section 5.1.3 / Tables 1-4"},{"comment":"The initial cluster centers μ_c^0 are the means of the (noisy) labeled support samples, so the procedure is not label-insensitive as claimed. The sentence 'The K-means clustering is an unsupervised process insensitive to label noises' is unsupported; at 60% symmetric noise each 5-way 5-shot class has on average only two correct samples, and the initial centers are strongly biased. Please provide a formal condition or an empirical sensitivity analysis (for example, cluster purity versus noise rate, recovery rate of true centers, or comparison with oracle prototypes) to justify the claim that the representatives ̄x_c are correct.","section":"Section 4.1, Eq. (1)"},{"comment":"The claim that noisy-label accuracy surpasses clean-trained SOTA is contradicted by the paper's own tables for symmetric noise. Table 1 reports 98.33% at 60% symmetric noise versus 98.46% for clean ProtoNet VIT and 98.57% for clean RNNP VIT; Table 2 reports 90.57% versus 94.67% and 94.42%. Please restrict the claim to the noise types and settings where it actually holds and revise the abstract accordingly.","section":"Abstract / Section 5.2"},{"comment":"The method is not reproducible as reported. No values are given for σ_0, λ, σ_max, σ_min, ε, k_up, h_u, A, B, or the specific ViT variant. The scale-adaptation procedure is under-specified: step 1 sets σ_max = σ_min = 0 but step 2 uses these values, and there is no explicit stopping rule for the case where exactly one neuron is activated. Please provide the hyperparameters, the exact checkpoint, and either working code or a precise pseudocode with convergence behavior.","section":"Section 4.3 / Section 5.1.3"}],"minor_comments":[{"comment":"The soft-assignment weight formula uses k both as an iteration index and as a category index; the denominator should sum over categories c' = 1, ..., m with distinct indices. Please rewrite with separate indices.","section":"Section 4.1"},{"comment":"The sentence defining the constants says 'A = 1/(3√(2πσ))' where the second constant should be B; the text later says A = 1.5 and B = 0.5, but the derivation should be corrected.","section":"Section 4.2"},{"comment":"The dataset description contains a typo: '20 for testing classes for training' should read '20 for testing'.","section":"Section 5.1.1"},{"comment":"The caption 'The results in Tables, 1 by [10]' is ungrammatical; it should read 'The results in Tables 1 and 2 are from [10]' or similar.","section":"Table 1 caption"},{"comment":"The text says 'accuracy advantage of 1.09% and 3.15% over Vinilla ProtoNet(VIT) with 60% symmetric label swap noise' in the paragraph on outlier noise; it should say 'outlier noise' and 'Vanilla'.","section":"Section 5.2"},{"comment":"The sentence 'we may not relabel some support samples because of their low weights. In this case, we have to abandon them' is unclear: the method does not relabel samples, and it is not explained what abandoning a sample means for the prototype computation.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about frozen-ViT leakage is valid and is the central reason I cannot recommend acceptance. The authors must disclose the exact checkpoint and pretraining data, and provide a non-overlapping evaluation. If the non-overlapping evaluation cannot be produced, the paper should be rejected. The lack of code and hyperparameters is a separate, serious reproducibility issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know: the headline result is not credible as evidence. The paper reports 99.17% clean and 98.33% at 60% symmetric noise on MiniImageNet with a frozen ViT. MiniImageNet and TieredImageNet are ImageNet subsets, and the ViT checkpoint is almost certainly pretrained on ImageNet. The paper never identifies the exact checkpoint, never excludes overlapping classes, and ships no code. That combination of near-perfect numbers and undisclosed pretraining overlap is a leakage red flag, not a robustness result. The stress-test note is right: under 60% symmetric noise each 5-way 5-shot class has only two correct labels on average, so the only way soft K-means recovers true clusters is if the feature space is already class-discriminative for those exact classes, which is exactly what pretraining on ImageNet gives you.\n\nCredit where it is due. The method is a legitimate extension: soft K-means prototypes initialized by per-class means, plus a Mexican-hat receptive field with per-query scale adaptation. The two-neural-field framing is decorative, but the components fit together, and the ablations show that soft clustering and scale adaptation each help. The paper is clearly written enough for the mechanism to be understood, and it does not fit parameters to query labels, so the circularity burden is low. This is not a crackpot paper; it is an experimental evaluation that does not support its claims.\n\nThe other soft spots are real but secondary. The robustness mechanism rests on an unproven assertion that soft K-means is insensitive to label noise; no sensitivity analysis is given. Hyperparameters for the scale adaptation, the stopping threshold, and the feature extractor are all missing. There are also internal inconsistencies—the model is sometimes called VCIM, sometimes RoNFA—and at least one likely equation typo in the receptive-field definition. None of these matter if the evaluation is invalid, but they would matter in a revision.\n\nBottom line: the paper is for readers who want a training-free classifier with a mild robustness story, and the method idea might survive contact with a properly designed experiment using a backbone with no overlap on the target classes. As submitted, the evidence base is broken. A serious editor could send this to referees because the problem is interesting and the method is nonstandard, but the revision would need a complete redo of the experimental protocol, including explicit disclosure of the pretraining corpus and, ideally, a non-leaking backbone. I would not cite it in its current form.","headline":"The method is a reasonable extension of prototype/kernel ideas, but the near-perfect accuracy almost certainly reflects frozen-ViT pretraining overlap with MiniImageNet/TieredImageNet, so the central empirical claim does not hold as presented.","tokens_in":789,"tokens_out":1104,"would_cite":false,"duration_ms":35862,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a two-neural-field few-shot classifier, RoNFA, keeps accuracy near 99% on MiniImageNet even when 60% of support labels are randomly wrong, by building category prototypes through soft K-means and classifying with…","keywords":["Few-shot learning","Noisy labels","Neural field model","Soft K-means clustering","Receptive field","Scale adaptation","Vision Transformer"],"falsifier":"Measure, on the same MiniImageNet 5-way 5-shot tasks with 60% symmetric noise, the purity of the soft K-means assignments relative to the true clean labels; if the assignments disagree with the true classes for more than a few percent of support samples, the reported 98.33% accuracy cannot be produced by the clustering mechanism as described. A second check is to rerun the method with a frozen convolutional encoder that yields less separated features and see whether accuracy at 60% noise collapses.","tokens_in":11038,"feed_emoji":"🎯","tokens_out":6626,"duration_ms":64642,"temperature":0.7,"pith_summary":"Few-shot classifiers learn new categories from a handful of labeled images, so even a small number of wrong labels can destroy accuracy. This paper claims that a classifier which constructs category prototypes from noisy support features can stay almost unaffected: with 60% of support-set labels randomly swapped, its 5-way 5-shot accuracy remains 98.33% on MiniImageNet and 90.57% on TieredImageNet, close to its clean-support results and above what strong baselines achieve on clean labels. The mechanism is to place each category neuron in a neural field at a soft K-means center of that category's noisy support features, then classify a query by which category neuron's adaptive receptive field activates. If the claim holds, label noise need not be corrected or relabeled at all in few-shot settings; robust prototypes can absorb it.","feed_headline":"Few-shot classifier keeps ~99% accuracy at 60% label noise","feed_subtitle":"Soft K-means prototypes and adaptive receptive fields absorb the noise, beating clean-support baselines.","key_machinery":"The central object is the Mexican-hat receptive field defined by Eq. (3): $\\phi_\\sigma(\\mathbf{x},\\bar{\\mathbf{x}}_c)=A e^{-\\frac{1}{2}\\|\\mathbf{x}-\\bar{\\mathbf{x}}_c\\|_2^2/\\sigma^2}-B e^{-\\frac{1}{2}\\|\\mathbf{x}-\\bar{\\mathbf{x}}_c\\|_2^2/(3\\sigma)^2}$ with $A=1.5$, $B=0.5$. Each category neuron in the category field has one such field centered at the soft K-means prototype $\\bar{\\mathbf{x}}_c$; a query feature activates neurons only when their field response exceeds a resting level. The scale-adaptation loop shrinks or grows $\\sigma$ until exactly one neuron fires, turning the kernel width into an adaptive decision boundary. The soft K-means update in Eq. (1) supplies the prototype centers that the fields lock onto, which is what makes the pipeline insensitive to noisy labels.","core_discovery":"On the paper's own terms, the central discovery is that a two-neural-field classifier—one field holding support features, one holding category neurons with Mexican-hat receptive fields—is enough to make few-shot classification nearly immune to label noise. The model, called RoNFA in the title and VCIM in the body, computes per-class representatives as fixed points of soft K-means initialized at the noisy per-class means, places each category neuron's receptive-field center at that representative, and during prediction adjusts the field scale until exactly one category neuron activates. With a frozen Vision Transformer encoder and no backpropagation, the reported 5-way 5-shot accuracy is 99.11% on MiniImageNet and 94.85% on TieredImageNet at 40% symmetric noise, and 98.33% and 90.57% at 60% symmetric noise, versus clean-support baselines that drop far below those levels. The paper states this as an empirical demonstration of robustness rather than as a theoretical guarantee.","pith_inferences":["Editorial inference: the method's success appears contingent on the frozen encoder's feature separability; a natural stress test is to repeat the protocol with a convolutional encoder or with visually confusable classes, where the cluster-initialization premise would degrade.","Editorial inference: the reported 'even surpasses clean-trained baselines' result suggests an upper-bound experiment: measure the clean-label accuracy ceiling of the same two-field classifier; if it is already near 99%, the noise robustness is less surprising than if it is substantially lower.","Editorial inference: the scale-adaptation loop can be read as a parameter-free nearest-neighbor rule with an adaptive kernel width; this connection suggests studying the method under distribution shift where no single scale exists, for example mixed-resolution or out-of-distribution queries.","Editorial inference: since the method never corrects labels, it may also apply to unsupervised few-shot adaptation where support identities are unknown except for the number of clusters; the paper does not test this."],"forward_implications":["If the claim is right, few-shot systems can tolerate support-set labels that are mostly wrong without any relabeling, loss reweighting, or fine-tuning.","The reported numbers imply a noise tolerance ceiling: accuracy drops less than 1% on MiniImageNet and about 5% on TieredImageNet as symmetric noise goes from 0% to 60%.","At 60% symmetric noise, the method's accuracy is stated to exceed the clean-support accuracy of the ViT-based ProtoNet and RNNP baselines, which would make noise robustness a property of prototype construction rather than of extra model capacity.","Because prediction requires only responses of fixed field neurons rather than iterative training, inference stays parameter-free and cheap, suggesting the method could transfer to settings where per-task fine-tuning is impractical."],"supporting_citations":[{"why":"Supplies the frozen Vision Transformer feature encoder used for all ViT results, including the reported 99%-level accuracies.","marker":"[2]"},{"why":"Defines the few-shot-learning-with-noisy-labels setting, provides the TraNFS baseline, and is the source of the comparison numbers reproduced in Tables 1-4.","marker":"[10]"},{"why":"Provides the RNNP robust-prototype baseline whose clean and noisy accuracies are compared against.","marker":"[14]"},{"why":"Defines symmetric label swap noise, the main noise type used in the experiments.","marker":"[19]"},{"why":"Defines paired label swap noise, the harder confusion-based noise condition.","marker":"[4]"},{"why":"Defines prototypical networks, the prototype-based baseline whose clean-support accuracy the paper claims to surpass.","marker":"[22]"},{"why":"Provides the MiniImageNet dataset used for the main evaluations.","marker":"[25]"},{"why":"Provides the TieredImageNet dataset used for the main evaluations.","marker":"[18]"}],"fun_headline_variants":["Neural fields keep few-shot accuracy near 99% under 60% label noise","RoNFA: Adaptive receptive fields resist noisy labels in few-shot tasks","Soft K-means prototypes give 99% few-shot accuracy despite 60% noise","Neural-field classifier outperforms clean-support baselines under noise","Two neural fields absorb label noise for few-shot learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on soft K-means, initialized at the noisy per-class means, actually finding representatives that match the true categories; in the frozen feature space this requires the classes to be so well separated that cluster recovery succeeds even when, at 60% symmetric noise in a 5-way 5-shot task, each class has on average only two correct labels.","fun_headline_variants_meta":{"raw":{"variants":["Neural fields keep few-shot accuracy near 99% under 60% label noise","RoNFA: Adaptive receptive fields resist noisy labels in few-shot tasks","Soft K-means prototypes give 99% few-shot accuracy despite 60% noise","Neural-field classifier outperforms clean-support baselines under noise","Two neural fields absorb label noise for few-shot learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000789,"raw_usage":{"total_tokens":3497,"prompt_tokens":979,"completion_tokens":2518,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":2422}},"tokens_in":595,"tokens_out":2518,"duration_ms":18118,"temperature":1.0,"reasoning_tokens":2422,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:03:00.271702+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, on the same MiniImageNet 5-way 5-shot tasks with 60% symmetric noise, the purity of the soft K-means assignments relative to the true clean labels; if the assignments disagree with the true classes for more than a few percent of support samples, the reported 98.33% accuracy cannot be produced by the clustering mechanism as described. A second check is to rerun the method with a frozen convolutional encoder that yields less separated features and see whether accuracy at 60% noise collapses.","supporting_citations":[{"cited_title":"Meta-learning for semi- supervisedfew-shotclassification,in:ProceedingsoftheInternational Conference on Learning Representations (ICLR)","cited_arxiv_id":null,"evidence_quote":"Provides the TieredImageNet dataset used for the main evaluations."},{"cited_title":"An image is worth 16x16 words: Transformers for image recognition at scale, in: Proceedings of the International Conference on Learning Representations (ICLR)","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen Vision Transformer feature encoder used for all ViT results, including the reported 99%-level accuracies."},{"cited_title":"Few- shot learning with noisy labels, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp","cited_arxiv_id":null,"evidence_quote":"Defines the few-shot-learning-with-noisy-labels setting, provides the TraNFS baseline, and is the source of the comparison numbers reproduced in Tables 1-4."},{"cited_title":"Rnnp: A robust few-shot learning approach, in: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pp","cited_arxiv_id":null,"evidence_quote":"Provides the RNNP robust-prototype baseline whose clean and noisy accuracies are compared against."},{"cited_title":"Learning with symmetric label noise: the importance of being unhinged, in: Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp","cited_arxiv_id":null,"evidence_quote":"Defines symmetric label swap noise, the main noise type used in the experiments."},{"cited_title":"Co-teaching:robusttrainingofdeepneuralnet- workswithextremelynoisylabels,in:ProceedingsoftheConference onNeuralInformationProcessingSystems(NeurIPS),p.8536–8546","cited_arxiv_id":null,"evidence_quote":"Defines paired label swap noise, the harder confusion-based noise condition."},{"cited_title":"Prototypical networks for few-shot learning, in: Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp","cited_arxiv_id":null,"evidence_quote":"Defines prototypical networks, the prototype-based baseline whose clean-support accuracy the paper claims to surpass."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MiniImageNet dataset used for the main evaluations."}],"review_version":1}