{"id":"f4552fcb-138f-473f-9aa0-bc442859aed0","arxiv_id":"2504.19824","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Gaussian-centered cropping variants are claimed to improve contrastive learning accuracy on small-image benchmarks over random cropping, with gains of 2.7 to 12.4 points on CIFAR-10.","lead":"A self-supervised learning study proposes two new cropping methods that sample image crops from a Gaussian distribution instead of uniformly at random. On three image datasets, the methods report accuracy gains over random cropping, but the evaluation omits the closest prior method and tunes key parameters on the test set.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The label-preserving mechanism is not supported because GCC/MGCC crop outside the image, potentially relying on unspecified padding; the in-bounds corrected variants underperform, so the reported gains may stem from padding or test-set hyperparameter selection rather than label-preserving center…","rationale":"The reader's weakest assumption centers on the lack of object-location statistics and the underperformance of corrected variants. My concern is more specific and more load-bearing: the uncorrected methods are not actually label-preserving because they sample out-of-bounds crops and rely on unspecified padding. This directly undermines the paper's explanatory mechanism, not just its transferability. The empirical accuracy claim could still be true, so I do not recommend moving from CONDITIONAL to REJECT; however, the conditionality is reinforced. The paper does provide a code repository, which makes the proposed ablation feasible. The disagreement with the reader is partial because the reader identified a related weakness but did not isolate the padding confound. The concrete test would settle whether the central claim's mechanism is real or whether the gains vanish once out-of-bounds padding is removed while preserving the same truncated Gaussian center distribution.","tokens_in":9692,"tokens_out":6966,"duration_ms":79533,"concrete_test":"Run the released code on CIFAR-10 under two conditions with fixed alpha=1.5 and crop size 0.4 across five seeds: (1) GCC with unspecified padding as currently implemented; (2) an in-bounds variant that re-samples centers from the same Gaussian truncated to the valid center region, rather than the CGCC center-shift. Compare against RandomCrop and CGCC on a held-out validation split, and record the fraction of crops requiring padding. If truncated-Gaussian in-bounds cropping does not exceed RandomCrop while padded GCC does, the headline gain is attributable to padding or out-of-bounds behavior, not to label-preserving center sampling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithms 1 and 2 draw crop centers from a Gaussian and then directly slice image[t:b, l:r] without any boundary check. Whenever the sampled center is near an edge, the crop lies partly outside the image; the paper acknowledges 'simple padding when necessary' in Sec. 5 but never specifies the padding value or mode. This matters because the paper's stated mechanism is label preservation: centering the crop distribution should reduce false positives by keeping views inside the semantic content. The methods themselves are not label-preserving whenever padding is applied. The two 'corrected' variants (CGCC/MCGCC), which keep crops inside the image, consistently underperform RandomCrop in Table 1 (e.g., CGCC 62.1 vs RandomCrop 61.1 at crop size 0.4, and far below at larger sizes). If the gains were due to label-preserving center bias, these in-bounds variants should at least match RandomCrop; instead, only the out-of-bounds, padding-dependent variants beat it. The paper's own explanation that padding behaves like Multi-Crop is speculative and contradicted by its observation that optimal alpha is small. Consequently, the central claim that GCC/MGCC improve accuracy through label-preserving cropping is not established: the improvement could be a padding artifact or an artifact of selecting the best alpha on the test set without a comparable RandomCrop hyperparameter.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two parameterized random-cropping methods for contrastive learning: Gaussian-Centered Cropping (GCC), which samples crop centers from a multivariate Gaussian centered on the image center, and Multi-Object Gaussian-Centered Cropping (MGCC), which samples the Gaussian mean from a uniform region around the center to accommodate off-center or multiple objects. Two 'corrected' in-bounds variants (CGCC and MCGCC) are also defined. The authors claim that GCC and MGCC improve downstream classification accuracy on CIFAR-10 by between 2.7% and 12.4% relative to the standard RandomCrop, with additional experiments on TinyImageNet and ImageNet64. The mechanism is argued to be label preservation: centering the crop distribution reduces false positives in contrastive learning. The paper includes pseudocode, figures showing accuracy versus the variance scaling parameter α, and a qualitative visual comparison of crops.","tokens_in":9989,"tokens_out":3688,"duration_ms":36756,"significance":"If the results hold, the contribution is practically valuable: GCC and MGCC are simple, computationally free replacements for RandomCrop that can be dropped into existing contrastive learning pipelines, and the authors provide code. However, the significance is materially weakened by evaluation gaps. The headline accuracies are peak values selected post-hoc on the test set, with no validation split or error bars in Table 1; the in-bounds corrected variants underperform RandomCrop, which is inconsistent with the stated label-preserving mechanism unless padding plays a role; and the closest prior parametric method (Contrastive Crop) is not compared experimentally. These issues mean the central claim, while potentially true, is not yet convincingly established.","major_comments":[{"comment":"The crop-center sampling in GCC and MGCC is unconstrained: the algorithms slice image[t:b, l:r] without any boundary check, and the only mention of out-of-bounds handling is the phrase 'simple padding when necessary' in Sec. 5, with no padding mode or value specified. This is load-bearing because the central claim is that the methods are label-preserving; padding destroys that property. Moreover, the corrected in-bounds variants CGCC and MCGCC consistently underperform RandomCrop in Table 1 (e.g., CGCC 62.1 vs RandomCrop 61.1 at crop size 0.4; CGCC 54.8 vs 54.2 at 0.8). If the gains were due to a center bias that avoids label-damaging crops, the in-bounds variants should at least match RandomCrop. I ask the authors to specify the padding, add an ablation comparing in-bounds variants with and without identical padding, and report the fraction of crops that fall out of bounds for each method and α.","section":"3.1 (Algorithms 1 and 2), 5"},{"comment":"The comparison in Table 1 is between the peak accuracy of GCC/MGCC over the α grid and the default RandomCrop without any tuning of RandomCrop's own parameters (e.g., scale/ratio). The text in Sec. 4.1 says experiments were repeated four times, but Table 1 reports only point values with no standard deviations or error bars. Since α is selected post-hoc on the test set, the headline gains are fitted values rather than predictions at a fixed hyperparameter. Please add a validation split or cross-validation for α selection, report mean ± std over the four runs, and either tune RandomCrop over a comparable set of cropping distributions or explicitly state that the baseline is an untuned default.","section":"4.1, Table 1, Fig. 1"},{"comment":"The related work mentions Contrastive Crop (Peng et al. [15]), a parameterized Beta-distribution cropping method, but no experimental comparison with it is provided. Since the paper's contribution is a new parameterized cropping distribution, the empirical advantage over the nearest prior parametric method is essential. Please add a direct comparison with Contrastive Crop under the same pretraining and evaluation protocol on at least CIFAR-10, reporting accuracy as a function of the Beta parameters.","section":"2, 4"},{"comment":"The label-preserving mechanism is asserted but never tested. There are no object-location statistics or foreground-fraction measurements for any dataset, and the paper's own explanation for why the corrected variants underperform is speculative; it is also internally inconsistent with the observation that optimal α is small, which by the authors' own reasoning implies minimal padding. Please provide direct evidence for the mechanism: for example, measure the fraction of crops that intersect the foreground object using bounding-box annotations (available for TinyImageNet) or synthetic center-perturbed data, and compare GCC against an in-bounds uniform crop distribution matched for center bias. Without such evidence, the title's 'Label-Preserving' claim is unsupported.","section":"1, 5"}],"minor_comments":[{"comment":"There are typos and stylistic issues: 'acurated' should be 'curated' in Sec. 4.1, and 'Signification' should be 'Significant' in Sec. 2. The reference list also contains a formatting error on [8] ('Gutmannet al.') that should be fixed.","section":"1, 2"},{"comment":"Several training details necessary for reproducibility are missing: the embedding dimension is called a hyperparameter but never specified, and the optimizer, learning rate schedule, batch size, temperature τ, and number of linear-probe epochs are not reported. Please include a full experimental configuration table.","section":"3.4, 3.5"},{"comment":"The ImageNet64 figure includes a 'Without Pretraining' curve, but its construction and relevance are not explained, and no numerical values or error bars are reported for TinyImageNet or ImageNet64. Please provide a results table for these datasets and clarify the without-pretraining baseline.","section":"4.3, Fig. 4"},{"comment":"The definition of wc,hc = sqrt(crop size) * (w,h) is ambiguous: state explicitly that crop size is a fraction of the image area and that wc,hc are the crop width and height. In Table 1, the α value at which each peak occurs should be reported alongside the peak accuracy, since α is a free parameter.","section":"3.1, Table 1"},{"comment":"Figures 5 and 6 are difficult to interpret: the red boxes are small and the ordering of crops within a row is not self-explanatory. Please provide higher-resolution versions and a clearer caption explaining how positive pairs are formed and how false positives would appear.","section":"5, Figs. 5-6"}],"recommendation":"major_revision","confidential_remarks":"The paper is a workshop-level submission with a simple and potentially useful idea. The evaluation, however, does not yet support the strong claims made in the abstract and conclusion. The combination of post-hoc α selection on the test set, missing error bars, unspecified padding, and absence of a comparison with Contrastive Crop means the reported gains could be artifacts of experimental setup rather than the proposed mechanism. I do not see a fatal flaw; the results may survive a properly controlled evaluation. I would be willing to review a revision that addresses the four major points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes two Gaussian-centered cropping variants for contrastive learning: GCC, which samples crop centers from a Gaussian fixed at the image center, and MGCC, which samples the Gaussian's mean from a uniform box around the center. The methods are simple, drop-in, add no runtime cost, and the code is public. The empirical comparison across CIFAR-10, TinyImageNet, and ImageNet64 is new, and the paper is honest enough to report that the in-bounds corrected variants underperform RandomCrop. That honesty is commendable, but it also exposes the core problem.\n\nThe headline claim—2.7% to 12.4% gains over RandomCrop—rests on numbers in Table 1 that are peak accuracies per method and crop size, with alpha selected on the test set. No validation split, no error bars in the table (the text mentions four repeats, but the figure only shows means), and no comparison with ContrastiveCrop, the closest prior parametric method that the paper itself builds on. So the main effect is a fitted value, not a fixed-hyperparameter result.\n\nThe stress-test note has it right about padding. Algorithms 1 and 2 draw centers near edges and slice out of bounds; Sec. 5 says \"simple padding when necessary\" without specifying the mode. The two corrected variants, which stay in bounds, consistently do worse than RandomCrop. If the gains came from keeping crops label-preserving, the in-bounds variants should at least match RandomCrop. The fact that only the out-of-bounds, padding-dependent variants beat it means the reported improvement could be a padding artifact or a side effect of the test-set selection protocol, not the proposed mechanism. The paper's own Multi-Crop speculation is explicitly contradicted by the low optimal alpha.\n\nThat said, the idea is not absurd. Centering the crop distribution is a reasonable inductive bias for datasets like CIFAR-10 where objects are roughly centered. The paper just does not provide the evidence to separate that bias from the padding effect or from selection artifacts. The missing object-location statistics are exactly what would have supported the mechanism.\n\nBottom line: this deserves a serious referee, but only with major revision. Ask for a fixed alpha protocol on a held-out split, error bars in the main table, a ContrastiveCrop baseline, and a clarification or ablation of the padding behavior. As it stands, I would not cite the reported numbers in my own work.","headline":"Simple, plausibly useful cropping variants, but the evaluation is too loose to credit the label-preserving story.","tokens_in":10498,"tokens_out":1626,"would_cite":false,"duration_ms":17792,"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":"Replacing uniform random cropping with Gaussian-centered crop sampling raises CIFAR-10 accuracy by 2.7%–12.4% in contrastive learning.","keywords":["contrastive learning","self-supervised learning","Gaussian-centered cropping","multi-object cropping","label-preserving augmentation","random cropping","CIFAR-10","linear evaluation"],"falsifier":"Compute the proportion of annotated object area inside crops produced by GCC, MGCC, and RandomCrop on a dataset with bounding boxes (or use a CIFAR-10 subset with object masks). If the Gaussian-centered crops do not enclose a higher fraction of object pixels than RandomCrop, the label-preservation explanation collapses even if accuracy improves.","tokens_in":1588,"feed_emoji":"🎯","tokens_out":2121,"duration_ms":70659,"temperature":0.7,"pith_summary":"The paper sets out to tame the randomness of random cropping in contrastive learning by replacing the uniform crop distribution with a parameterized Gaussian one. Its central claim is that two such samplers—Gaussian-Centered Cropping (GCC), which draws crop centers from a normal distribution fixed at the image center, and Multi-Object Gaussian-Centered Cropping (MGCC), which lets the distribution mean wander uniformly near the center—reduce false positive pairs and thereby improve learned representations. On CIFAR-10 the authors report downstream accuracy gains of 2.7% to 12.4% over RandomCrop depending on crop size, and similar advantages on TinyImageNet and ImageNet64 when hyperparameters are chosen well. The appeal of the claim is that the improvement costs no extra computation or auxiliary networks, only a different random sampler.","feed_headline":"Gaussian-centered cropping beats random crops by up to 12.4%","feed_subtitle":"Two new crop samplers beat RandomCrop on CIFAR-10, TinyImageNet, and ImageNet64.","key_machinery":"The load-bearing object is the two-dimensional Gaussian center sampler. GCC fixes the mean at the image center and sets the covariance to $\\operatorname{diag}(\\alpha w, \\alpha h)$, so the scaling factor $\\alpha$ controls how far apart the two views' centers may lie; MGCC replaces the fixed mean with one drawn uniformly from a specified central band $[a \\cdot w, b \\cdot w] \\times [a \\cdot h, b \\cdot h]$. The variance parameter tunes a trade-off: low $\\alpha$ concentrates crops near the center and yields consistent but low-variance pairs, while high $\\alpha$ drifts toward false positives. The corrected variants add a clamp that shifts out-of-bounds centers toward the image center, and the paper's comparison of corrected versus uncorrected versions isolates the effect of allowing crops to extend beyond the boundary.","core_discovery":"The central discovery is that crop-center geometry, not crop size alone, is a controllable source of false positives in contrastive self-supervision. GCC samples both positive views from a two-dimensional normal distribution whose mean is the image center and whose variance is $\\alpha w$ by $\\alpha h$; MGCC instead samples the distribution mean uniformly in a central region before drawing centers. Compared with uniform RandomCrop, these samplers bias views toward the center of the image, which the paper argues preserves the label more often and keeps the two views at a useful distance. On CIFAR-10, GCC peaks at 67.8% and MGCC at 66.0% linear-evaluation accuracy (200 epochs, 40% crop size) versus 61.1% for RandomCrop. The corrected variants CGCC and MCGCC, which clamp every crop fully inside the image, plateau below RandomCrop; the authors account for this by suggesting the uncorrected methods behave like multi-scale view generation, while cautioning that their own evidence points mainly to well-separated centered pairs rather than padding.","pith_inferences":["If the label-preservation mechanism is what drives the gain, the size of the improvement should track how concentrated the dataset's objects are near the image center; a dataset with consistently off-center objects would be expected to erode GCC's advantage and favour MGCC. This is a testable consequence the paper does not run.","The uncorrected methods' out-of-boundary crops may add a subtle scale augmentation by effectively resampling outside the image; comparing GCC against a version that pads with zeros or reflects pixels rather than shifting centers could separate that effect.","MGCC could be dropped into Multi-Crop style pipelines as the global-view generator, leaving local views random; the paper suggests integrability but does not report such an experiment.","The reported gains are against a RandomCrop baseline that may itself be under-tuned; re-running the comparison with stronger contrastive baselines would reveal how much of the 2.7–12.4% gap is specific to the new samplers."],"forward_implications":["Standard contrastive pipelines can replace RandomCrop with GCC or MGCC by changing only the crop-center distribution, with no architectural or loss-function changes.","The gains widen as crop size grows: at 80% crop size on CIFAR-10, GCC exceeds RandomCrop by 12.4 points (66.6% vs 54.2%), so the method matters most when views are large.","MGCC is the preferable sampler when datasets contain multiple or off-center objects; it outperforms GCC on TinyImageNet at 40% and 60% crop sizes and on ImageNet64 across tested sizes.","Forcing crops to stay fully inside the image is counterproductive in these experiments; the corrected CGCC and MCGCC variants consistently fall below RandomCrop.","The optimal $\\alpha$ and crop size are dataset-dependent, so the method introduces a small hyperparameter search in exchange for compute-free gains."],"supporting_citations":[{"why":"Supplies the contrastive learning setup and NT-Xent loss used for pretraining and evaluation.","marker":"[4]"},{"why":"Provides the parameterized Beta-cropping baseline and general view-crafting setup that this work extends.","marker":"[15]"},{"why":"Establishes that random cropping is a non-label-preserving augmentation, the paper's motivating premise.","marker":"[16]"},{"why":"Underlies the InfoNCE/NT-Xent objective used for contrastive self-labeling.","marker":"[8]"},{"why":"Primary benchmark dataset for the reported 2.7%–12.4% gains.","marker":"[11]"},{"why":"Introduces Multi-Crop view generation that the paper contrasts with and could replace.","marker":"[2]"},{"why":"Prior object-aware cropping baseline compared as a compute-heavy alternative.","marker":"[13]"}],"fun_headline_variants":["Centered crops boost contrastive learning accuracy by 12.4%","Gaussian cropping lifts CIFAR-10 linear eval by 12.4%","New crop samplers add 12.4% to contrastive learning","Crop geometry, not size, drives SSL gains on CIFAR-10","Center-biased samplers reduce false labels in SSL"],"cache_read_input_tokens":12672,"weakest_assumption_plain":"The central assumption is that centering the crop distribution increases the fraction of label-preserving crops; the paper states this mechanism but supplies no object-location statistics showing that GCC and MGCC crops actually contain more of the object than RandomCrop crops do.","fun_headline_variants_meta":{"raw":{"variants":["Centered crops boost contrastive learning accuracy by 12.4%","Gaussian cropping lifts CIFAR-10 linear eval by 12.4%","New crop samplers add 12.4% to contrastive learning","Crop geometry, not size, drives SSL gains on CIFAR-10","Center-biased samplers reduce false labels in SSL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001131,"raw_usage":{"total_tokens":4708,"prompt_tokens":959,"completion_tokens":3749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":3650}},"tokens_in":575,"tokens_out":3749,"duration_ms":26932,"temperature":1.0,"reasoning_tokens":3650,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:42:28.292042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the proportion of annotated object area inside crops produced by GCC, MGCC, and RandomCrop on a dataset with bounding boxes (or use a CIFAR-10 subset with object masks). If the Gaussian-centered crops do not enclose a higher fraction of object pixels than RandomCrop, the label-preservation explanation collapses even if accuracy improves.","supporting_citations":[{"cited_title":"A simple framework for contrastive learn- ing of visual representations","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive learning setup and NT-Xent loss used for pretraining and evaluation."},{"cited_title":"Crafting better contrastive views for siamese rep- resentation learning","cited_arxiv_id":null,"evidence_quote":"Provides the parameterized Beta-cropping baseline and general view-crafting setup that this work extends."},{"cited_title":"A survey on image data augmentation for deep learning","cited_arxiv_id":null,"evidence_quote":"Establishes that random cropping is a non-label-preserving augmentation, the paper's motivating premise."},{"cited_title":"Gutmann and Aapo Hyv ¨arinen","cited_arxiv_id":null,"evidence_quote":"Underlies the InfoNCE/NT-Xent objective used for contrastive self-labeling."},{"cited_title":"Learning multiple layers of features from tiny images","cited_arxiv_id":null,"evidence_quote":"Primary benchmark dataset for the reported 2.7%–12.4% gains."},{"cited_title":"Unsupervised learning of visual features by contrasting cluster assignments","cited_arxiv_id":null,"evidence_quote":"Introduces Multi-Crop view generation that the paper contrasts with and could replace."},{"cited_title":"Object-Aware Cropping for Self-Supervised Learning","cited_arxiv_id":"2112.00319","evidence_quote":"Prior object-aware cropping baseline compared as a compute-heavy alternative."}],"review_version":1}