{"id":"2683f617-7388-400f-99f9-bd67fa16a07e","arxiv_id":"2501.13418","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A contrastive pre-training method with separate losses for same-sample, same-class, and different-class relations achieves 1-shot state-of-the-art on miniImageNet, tieredImageNet, CIFAR-FS, and CUB.","lead":"This paper proposes a pre-training method for few-shot classification that models three kinds of sample relationships separately, and reports the best 1-shot accuracy on four standard benchmarks. A generalist might care because the method is simple, can be added to existing few-shot systems, and improves their accuracy when used as a feature extractor.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'state-of-the-art on all four benchmarks' claim is not secured: the reported 1-shot margins over the closest baselines are within overlapping single-run 95% confidence intervals, and the loss weights and temperatures are tuned on miniImageNet and then applied to the other three datasets.","rationale":"The paper's central claim is empirical: MGRCL is a simple pre-training recipe that achieves state-of-the-art 5-way 1-shot accuracy on four benchmarks and improves existing methods when used as a backbone. The ablations are genuinely informative—TCL and CCL each add non-trivial gains over the augmented baseline, and using MGRCL as pretraining consistently improves FEAT, Meta-Baseline, and STV AE over their re-implemented baselines. The t-SNE and grid-search sections are useful. However, the load-bearing part of the central claim is the SOTA statement, and that is exactly where the evidence is least secure. The reader's weakest assumption about TCL is a real interpretability risk, but even if the semantic-consistency premise is imperfect, MGRCL could still work well as a regularizer and the SOTA claim would survive. Conversely, if the 0.2-0.7% margins over PAL and ESPT are training-seed noise, the 'surpasses a majority of leading methods' claim does not survive as stated. Thus the statistical robustness of the comparisons is the most load-bearing concern. I would keep the reader's conditional verdict, with the condition being multi-seed, protocol-locked replication and either released code or a complete augmentation specification.","tokens_in":18532,"tokens_out":8825,"duration_ms":78945,"concrete_test":"Run MGRCL together with the closest comparators (PAL, ESPT, and a SupCon-equipped baseline) on miniImageNet and CUB with 5 independent training seeds each, using identical augmentation and evaluation code; fix α, β, τ1, and τ2 on a validation split of miniImageNet before evaluating CIFAR-FS and CUB. Report paired mean±std across seeds and paired differences. If the 95% confidence interval for the best pairwise difference includes 0, the SOTA claim should be downgraded to 'competitive' and the verdict adjusted accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Tables 1-4, the 5-way 1-shot advantage over the nearest competitor is small: 69.57±0.45 vs PAL's 69.37±0.64 on miniImageNet, 72.98±0.51 vs ESPT's 72.68±0.22 on tieredImageNet, and 86.14±0.38 vs ESPT's 85.45±0.18 on CUB. Each pairwise difference is well inside the quoted 95% intervals, so the word 'SOTA' is not supported by the reported statistics. Section 4.4 selects α=1.0 and β=0.1 by grid search on miniImageNet only, then applies those same values to all four datasets; this can inflate the headline result through selection on the first benchmark. The paper reports no variance across training seeds and provides no code or complete augmentation protocol, so the reader cannot determine whether the margins reflect the method or a favorable run. If these margins collapse under multi-seed replication, the strongest claim reduces to 'competitive pre-training recipe,' while the qualitative contribution of modeling three relation types is not disproven.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Multi-Grained Relation Contrastive Learning (MGRCL), a pre-training method for few-shot classification that models sample relations at three granularities: intra-sample relations (Transformation Consistency Learning, TCL, which aligns temperature-smoothed label distributions of a sample and its augmented versions via Jensen-Shannon divergence), and intra-class/inter-class relations (Class Contrastive Learning, CCL, a memory-bank-based contrastive loss that pulls a sample toward a homogenous sample and pushes it away from samples of all other classes). The base learner jointly optimizes a classification loss, a self-supervised transformation-prediction loss, TCL, and CCL. The method is evaluated on miniImageNet, tieredImageNet, CIFAR-FS, and CUB with a ResNet-12 backbone, reporting 5-way 1-shot accuracies of 69.57%, 72.98%, 78.54%, and 86.14%, respectively, and it is claimed to be state-of-the-art in 5-way 1-shot classification. The paper also shows that using MGRCL as a pre-trained backbone improves the two-stage methods FEAT, Meta-Baseline, and STV AE, and it provides component ablations, hyperparameter sweeps, comparisons to UnSupCon/SupCon, and t-SNE visualizations.","tokens_in":18814,"tokens_out":17899,"duration_ms":106524,"significance":"If the empirical claims hold, MGRCL is a simple and transferable pre-training recipe with a conceptually clean decomposition of sample relations, and the plug-and-play gains for FEAT, Meta-Baseline, and STV AE are practically useful. The ablations in Table 5 and the same-base-learner comparison in Fig. 5 give explicit credit to each designed component, and Sections 5.2-5.3 candidly acknowledge the computational overhead and memory-bank scaling cost. Balanced against these strengths, the headline state-of-the-art claim is not statistically secured on miniImageNet and tieredImageNet (overlapping confidence intervals), the hyperparameters are selected on the headline benchmark (Section 4.4), and no code or complete augmentation protocol is provided. These are fixable within the manuscript's scope, so I regard the contribution as potentially solid but in need of substantial additional evidence before it can be accepted.","major_comments":[{"comment":"The claim that MGRCL achieves state-of-the-art 5-way 1-shot accuracy is not supported by the reported statistics on two of the four benchmarks. On miniImageNet, Ours (69.57±0.45) is 0.20 points above PAL (69.37±0.64), and on tieredImageNet, Ours (72.98±0.51) is 0.30 points above ESPT (72.68±0.22); the quoted 95% confidence intervals overlap in both cases. Since Section 4.1.5 states that the intervals are computed over 2,000 sampled tasks, and no variance across training seeds is reported anywhere, the pairwise margins do not demonstrate a statistically meaningful advantage. Please report means and standard deviations over multiple training seeds with a significance test, or qualify the state-of-the-art wording.","section":"4.2, Tables 1-2"},{"comment":"The loss weights α and β are selected by grid search on miniImageNet (Table 6), and the temperature parameters τ1 and τ2 are tuned on miniImageNet (Fig. 4); the headline miniImageNet results in Tables 1 and 5 coincide exactly with the best grid point (α=1.0, β=0.1, 69.57). The reported miniImageNet number is thus the maximum over the searched grid, which is a selection-on-the-headline-benchmark effect. The paper should either verify the chosen hyperparameters on the other three benchmarks (e.g., by reporting the grid landscape on CUB), or report the results of a configuration fixed without reference to the test benchmark.","section":"4.4, Table 6"},{"comment":"The quantitative claims are not reproducible from the manuscript. Section 4.1.4 lists the augmentation types ('three scaling transformations, three rotation transformations, one random erasing, one graying, and one Sobel edge detection') without any parameters, the self-supervised MLP in Section 4.1.2 is underspecified (hidden dimension, activation), and no code or checkpoints for MGRCL are provided - the only URL in the paper (Section 4.5) points to the SupCon repository used for baseline comparisons. Please provide the full augmentation protocol and training details and release code, so that the central empirical claims can be reproduced and checked across training seeds.","section":"4.1.4 and code availability"},{"comment":"The TCL component rests on the premise stated in Section 3.4 that 'the label output can represent the semantic content of the sample.' This premise is load-bearing: for aggressive augmentations such as Sobel edge detection or rotation, forcing exact Jensen-Shannon consistency of label distributions could suppress useful invariant or equivariant features. The aggregate gains in Table 5 are consistent with the premise but do not isolate it. Please add an ablation that removes each augmentation type from the TCL objective one at a time, or that compares TCL against a feature-level consistency loss, to verify that the semantic-consistency assumption holds for every transformation used.","section":"3.4"},{"comment":"The 5-way 1-shot Baseline in Table 5 is 66.78±0.43 on miniImageNet, which exactly matches the FEAT result reported in Table 1 (66.78±0.20), even though Section 4.3 describes the baseline as 'the same as RFS' with added augmented samples, and RFS-distill is reported as 64.82±0.60 in Table 1. Please clarify the exact training recipe that produced the baseline and state explicitly whether it is an RFS-style cross-entropy model, a FEAT model, or another configuration; every incremental improvement in Table 5 is measured from this value.","section":"4.3, Table 5"}],"minor_comments":[{"comment":"The heading 'Limitions' in Section 5.1 should be 'Limitations', and 'effevtiveness' in Section 4.6 should be 'effectiveness'.","section":"5.1 and 4.6"},{"comment":"The dataset name is spelled 'CIF AR-FS' with a space in Sections 1, 4.1.1, and the Table 3 caption; use 'CIFAR-FS' consistently.","section":"1, 4.1.1, Table 3"},{"comment":"The phrase 'the second of Figure 6b' is unclear; it should be 'the second row of Figure 6b' or similar.","section":"4.6"},{"comment":"Reference [37] is missing a comma between the authors and has a spacing error in the venue name ('onInformation theory'); reference [14] lacks volume and page information.","section":"References"},{"comment":"The sentence 'outperforming the second-best results of 0.69% and 0.73%' should read 'outperforming the second-best results by 0.69 and 0.73 percentage points.'","section":"4.2"},{"comment":"The footnote about using the SupCon code for both the unsupervised and supervised baselines should state explicitly which loss function (SimCLR-style or SupCon-style) is applied as 'UnSupCon' and 'SupCon' in Fig. 5, since both are implemented with the same repository.","section":"4.5"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern raised by the reader lands: the SOTA claim on miniImageNet and tieredImageNet rests on margins within overlapping confidence intervals, and the hyperparameters are selected on miniImageNet. The exact equality between the Table 5 baseline (66.78) and the FEAT number in Table 1 deserves a written clarification from the authors; if the baseline is in fact a FEAT-style model, the ablation story changes. I see the contribution as a useful empirical recipe rather than a conceptual breakthrough; the novelty is a clean combination of prediction-consistency regularization and memory-bank supervised contrastive learning. The missing code and incomplete augmentation protocol should be treated as a requirement for acceptance at a journal that values reproducible empirical work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea—splitting sample relations into intra-sample, intra-class, and inter-class, and handling each with a different loss—is simple and sensible, and the ablation evidence says it helps. The headline \"SOTA on four benchmarks\" is not actually secured by the reported statistics. The margins over the closest baselines are within the quoted 95% confidence intervals, and those intervals only reflect task-sampling variance, not training-seed variance. Hyperparameters are tuned on miniImageNet and then applied to the other sets, and no code is released. So the right description is \"competitive and probably useful,\" not \"SOTA.\"\n\nThat said, the paper does several things well. The decomposition is a real framing difference from SimCLR/SupCon, and the authors back it with a fair comparison: on the same base learner, their MGRCL beats both UnSupCon and SupCon. The ablations on miniImageNet and CUB show each component adds a point or more, which is meaningful in this field. The plug-in results—e.g., Meta-Baseline jumping from 63.38 to 69.01 on miniImageNet 1-shot when using their pre-trained backbone—are a strong demonstration that the pretraining transfers well.\n\nThe soft spots are mostly about the strength of the claims, not the method. The TCL assumption that label outputs capture semantic content is plausible, but the paper doesn't test whether aggressive augmentations like Sobel edge detection actually preserve that consistency. The memory-bank supervised contrastive loss is similar to SupCon, so the novelty rests on the distinction between TCL and CCL, which is fine but modest. The SOTA margins are small enough that multi-seed runs could erase them entirely.\n\nThis paper deserves a serious referee. It is a solid empirical contribution to few-shot pretraining, and the core recipe is easy to adopt. The revision should add multi-seed runs, report full training variance, and ideally release code. The SOTA language should be softened to \"competitive\" unless the intervals hold up under proper replication. Whoever reviewers are, they should focus on whether the margins reproduce, not on whether the idea is novel.","headline":"A sensible decomposition of sample relations with a useful pretraining recipe, but the SOTA claim is not statistically secured; the method is competitive and deserves revision, not desk rejection.","tokens_in":19299,"tokens_out":2442,"would_cite":true,"duration_ms":22811,"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":"Modeling sample relations at three granularities — exact consistency within a sample, closeness within a class, separation between classes — yields state-of-the-art few-shot classification accuracy and a pre-trained backbone that improves…","keywords":["few-shot classification","contrastive learning","sample relations","transformation consistency","class contrastive learning","pre-training","feature learning"],"falsifier":"Train the same base learner with and without TCL on a dataset where one of the used augmentations (for instance, 90-degree rotation on digit images) flips the semantic label; if the TCL-trained model does not beat the no-TCL baseline on held-out 5-way 1-shot tasks, the premise that label outputs capture semantic content and that alignment preserves it is not supported.","tokens_in":18371,"feed_emoji":"🎯","tokens_out":5979,"duration_ms":47377,"temperature":0.7,"pith_summary":"The paper claims that existing few-shot contrastive pre-training fails because it models all sample relations with one recipe, ignoring that a transformed view of an image is semantically identical to it, while a same-class image is only similar and a different-class image is dissimilar. It proposes Multi-Grained Relation Contrastive Learning (MGRCL), which splits relations into intra-sample, intra-class, and inter-class, enforces the first with label-distribution alignment (TCL) and the latter two with a relative contrastive loss (CCL). A sympathetic reader would care because, if correct, a simple single-stage pre-training method matches or exceeds two-stage meta-learners and can be dropped into existing few-shot methods as a better frozen backbone.","feed_headline":"Three-way relation split tops few-shot accuracy on four benchmarks.","feed_subtitle":"One pre-training pass with TCL and CCL beats two-stage meta-learners and plugs into methods as a better backbone.","key_machinery":"The central objects are the two auxiliary losses added to a standard classification backbone. TCL operates on the classifier's softmax outputs, softened by temperature $\\tau_1 = 4.0$, and penalizes the Jensen-Shannon divergence between the prediction for the original image and the prediction for each of its transformed versions, enforcing identity rather than mere similarity at the label level. CCL operates on global-pooled feature embeddings, drawing a positive same-class feature from a memory bank, treating features of the other base classes as negatives, and optimizing a temperature-$\\tau_2=0.1$ softmax over cosine similarities; the memory bank is updated with momentum $r = 0.99$. A self-supervised branch that predicts which transformation was applied completes the base learner. Together the losses separate the three relation types that unsupervised and supervised contrastive learning conflate.","core_discovery":"MGRCL claims that the three kinds of sample relations — same sample under different transformations, different samples of the same class, and samples of different classes — have different strengths of semantic similarity and therefore need different modeling. Transformation Consistency Learning (TCL) enforces exact semantic consistency by aligning temperature-smoothed predicted label distributions of an image and its transformed versions with Jensen-Shannon divergence, while Class Contrastive Learning (CCL) uses a memory bank and a softmax over cosine similarities to keep each sample closer to a same-class sample than to samples of other classes. On this scheme the paper reports 5-way 1-shot accuracies of 69.57% on miniImageNet, 72.98% on tieredImageNet, 78.54% on CIFAR-FS, and 86.14% on CUB, and shows that using the resulting network as a frozen pre-trained backbone improves FEAT, Meta-Baseline, and STV AE on the same benchmarks.","pith_inferences":["Editorial inference: If label-output alignment is doing the work, TCL may also act as a strong regularizer against overfitting on small base datasets, which would predict larger gains when base-class data is scarce; this is not tested in the paper.","Editorial inference: The same three-way relation split could be applied beyond images — to text or audio — where augmentation families differ, and the method's success would depend on whether label distributions there faithfully reflect semantic content.","Editorial inference: A direct test of the paper's motivation would be to replace the memory bank with class prototypes; if performance holds, the relative contrastive formulation, not the bank, is the essential ingredient.","Editorial inference: Because the paper reports that aggressive augmentations such as Sobel edge detection are included, an interesting stress test is whether TCL suppresses texture-based features that are useful for some novel classes but absent after edge filtering."],"forward_implications":["A single fully supervised pre-training phase on base classes, without episode sampling or meta-tuning, is enough to reach state-of-the-art 5-way 1-shot accuracy on all four tested benchmarks.","Existing two-stage methods such as FEAT, Meta-Baseline, and STV AE inherit substantial gains simply by starting from the MGRCL backbone rather than a plain supervised one.","On fine-grained datasets like CUB, where class differences are small, the class-pushing CCL component contributes more than the transformation-alignment TCL component.","Unsupervised contrastive losses that treat every other image as negative can hurt few-shot pre-training because they push same-class samples apart; MGRCL avoids that by contrasting against a same-class positive from a memory bank.","Accuracy improvements come without extra network layers beyond a small MLP for the self-supervised transformation-prediction branch, keeping inference identical to standard feature-extraction-based few-shot classifiers."],"supporting_citations":[{"why":"Supplies the baseline paradigm of fully supervised pre-training on base classes followed by a frozen feature extractor, and the RFS method that MGRCL compares against.","marker":"[17]"},{"why":"Provides the supervised contrastive formulation and temperature convention that CCL adapts.","marker":"[25]"},{"why":"Supplies the unsupervised contrastive baseline (SimCLR) that the paper shows can hurt few-shot pre-training, and the temperature setting for CCL.","marker":"[26]"},{"why":"Contributes the memory-bank update rule ($r = 0.99$) and the invariant/equivariant contrastive FSL method that MGRCL builds on and compares against.","marker":"[18]"},{"why":"Motivates the momentum-encoder alternative discussed as a way to avoid storing a full memory bank.","marker":"[27]"},{"why":"Provides the dataset splitting protocol used for all four benchmarks.","marker":"[39]"},{"why":"FEAT is a two-stage meta-learning method used as a combination target to show MGRCL's pre-trained backbone transfers.","marker":"[53]"},{"why":"Meta-Baseline is a two-stage meta-learning method used as a combination target to show MGRCL's pre-trained backbone transfers.","marker":"[54]"},{"why":"STV AE is a generative augmentation method used as a combination target to show MGRCL's pre-trained backbone transfers.","marker":"[14]"},{"why":"Supplies the Jensen-Shannon divergence used as the TCL loss.","marker":"[36]"}],"fun_headline_variants":["Multi-grained contrastive learning improves few-shot classification","A three-way relation split lifts few-shot performance","Pre-training with TCL and CCL beats meta-learning on few-shot","New contrastive pre-training boosts few-shot methods as backbone","Rethinking sample relations: one pass beats two-stage few-shot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the classifier's predicted label distribution faithfully represents a sample's semantic content, so forcing identical label outputs across transformations preserves semantics rather than discarding useful information.","fun_headline_variants_meta":{"raw":{"variants":["Multi-grained contrastive learning improves few-shot classification","A three-way relation split lifts few-shot performance","Pre-training with TCL and CCL beats meta-learning on few-shot","New contrastive pre-training boosts few-shot methods as backbone","Rethinking sample relations: one pass beats two-stage few-shot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2614,"prompt_tokens":1037,"completion_tokens":1577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":1496}},"tokens_in":653,"tokens_out":1577,"duration_ms":13913,"temperature":1.0,"reasoning_tokens":1496,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:58:03.661303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same base learner with and without TCL on a dataset where one of the used augmentations (for instance, 90-degree rotation on digit images) flips the semantic label; if the TCL-trained model does not beat the no-TCL baseline on held-out 5-way 1-shot tasks, the premise that label outputs capture semantic content and that alignment preserves it is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the baseline paradigm of fully supervised pre-training on base classes followed by a frozen feature extractor, and the RFS method that MGRCL compares against."},{"cited_title":"Khosla, P","cited_arxiv_id":null,"evidence_quote":"Provides the supervised contrastive formulation and temperature convention that CCL adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the memory-bank update rule ($r = 0.99$) and the invariant/equivariant contrastive FSL method that MGRCL builds on and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the momentum-encoder alternative discussed as a way to avoid storing a full memory bank."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the dataset splitting protocol used for all four benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FEAT is a two-stage meta-learning method used as a combination target to show MGRCL's pre-trained backbone transfers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Meta-Baseline is a two-stage meta-learning method used as a combination target to show MGRCL's pre-trained backbone transfers."},{"cited_title":"Zhang, S","cited_arxiv_id":null,"evidence_quote":"STV AE is a generative augmentation method used as a combination target to show MGRCL's pre-trained backbone transfers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Jensen-Shannon divergence used as the TCL loss."}],"review_version":1}