{"id":"d96da614-3457-4600-bd77-5a1774a373b8","arxiv_id":"2505.24558","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A fixed spatial density mask applied to convolution kernels improves reported CIFAR-100 accuracy and DIV2K denoising PSNR without adding parameters.","lead":"The paper replaces the plain convolution kernel in CNNs with a fixed spatial weighting pattern, keeping the same number of trainable parameters. On CIFAR-100 and DIV2K benchmarks, the authors report higher accuracy and PSNR than standard convolution, with gains up to 10 points.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains are consistent with test-set tuning of alpha: alpha is tuned to improve reported metrics (Sec.","rationale":"The paper's contribution is empirical: a fixed spatial mask is inserted into the convolution, and the evidence for its value is the consistent improvement in Tables 3 and 5. The most load-bearing condition for that claim is that the mask hyperparameters were selected without using the test set. The text repeatedly says the density function was tuned to improve accuracy and the classification/denoising results, but it never documents a held-out validation split. Since setting alpha = 1 reproduces standard convolution exactly, the published result is equivalent to taking one default configuration (alpha = 1) and several tuned configurations (one per method), then reporting that the tuned configuration beats the default. With a small grid (one or two alpha parameters), this would hold often even if the operator had no intrinsic value, especially without error bars. A 10-point accuracy swing on CIFAR-100 from one run is not enough to rule out such an artifact. The DnCNN 5x5 baseline is a separate but reinforcing problem: 12.15 dB PSNR indicates an implementation fault rather than a legitimate standard convolution result. The reader's verdict identifies the same core weakness, so agreement is 'agree.' While the paper does make a concrete, parameter-free proposal that is cheap to test and reproducible code is promised, the current experimental protocol does not support the central claim as written. The right disposition is REJECT: the reported evidence is not valid for the claimed conclusion. A revision with a documented validation-based alpha selection, multiple seeds, and repaired baselines could make the claim testable and potentially acceptable.","tokens_in":11052,"tokens_out":4026,"duration_ms":55212,"concrete_test":"Re-run the comparison with an explicit validation protocol: train on the official training split, choose alpha on a held-out validation set (same grid values reported in Tables 3 and 5), then evaluate once on the untouched test set, with at least three independent seeds per alpha. If the best-validation alpha does not beat alpha = 1 on the untouched test set for VGG and DnCNN, the reported gains are a tuning artifact. Additionally, re-run DnCNN with 5x5 standard convolution using explicit padding='same' and confirm the output size equals the input; a correct baseline should give PSNR near or above the roughly 40 dB noisy-input level, not 12.15 dB.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that replacing standard convolution with the fixed rank-1 mask Phi = alpha*alpha^T (Eq. 4) improves every tested method, without adding trainable parameters. But Phi = 1 reduces exactly to standard convolution, so the comparison in Tables 3 and 5 is only a comparison between the default setting alpha = 1 and one tuned alpha per method. Section 4 explicitly states that the CINECA cluster was used to 'tune the density function to improve the accuracy of the weighted convolution and the results of the classification and denoising.' No validation split, alpha-selection protocol, or repeated runs are documented. If alpha was chosen by looking at the reported test metrics, the improvements are the expected result of selecting the best point of a small hyperparameter grid on the evaluation set, not evidence that the weighted convolution generalizes. The same language appears in Section 1 and in Section 5 ('tune the values of the density function to improve the accuracy of the denoising and classification'), which supports the reading that test-oriented tuning occurred. This makes the headline gains (e.g., VGG 56.89% to 66.94%; DnCNN 3x3 PSNR 20.17 to 22.63) uninterpretable as an evaluation of the operator. A second, independent defect is the DnCNN 5x5 standard baseline in Table 5: PSNR 12.15 dB is far below the noisy input level (with Gaussian noise sigma = 0.01 on [0,1], the noisy input has PSNR around 40 dB), indicating a broken baseline; the corresponding weighted result of 23.35 dB therefore does not establish an improvement over a valid standard 5x5 DnCNN baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weighted convolution operator that multiplies each kernel weight by a fixed spatial density function Phi = alpha*alpha^T (Eq. 4), with no additional trainable parameters, and evaluates it on CIFAR-100 classification and DIV2K denoising. The manuscript reports that the weighted convolution improves accuracy and PSNR/SSIM for every tested architecture compared with standard convolution, with VGG accuracy rising from 56.89% to 66.94% and DnCNN PSNR from 20.17 dB to 22.63 dB. The central claim is empirical: that this parameter-free modification of the convolution kernel consistently improves performance.","tokens_in":11333,"tokens_out":2521,"duration_ms":30891,"significance":"If the reported gains were obtained under sound evaluation protocols, the idea would be of practical interest because it is architecture-agnostic, adds no trainable parameters, and has a low-overhead implementation. The authors provide a public PyTorch implementation and report training times showing only modest overhead. However, the evaluation as presented does not support the central claim: the density parameter appears to have been selected on the test set, at least one baseline is implausibly broken, and no repeated-run statistics are reported. These issues make the headline improvements uninterpretable as evidence about the operator itself.","major_comments":[{"comment":"The paper states in the introduction, in Section 4, and in Section 5 that the values of the density function were 'tuned to improve the accuracy' of the weighted convolution and the results of classification and denoising. No validation split, selection grid, or selection protocol is documented. Since Phi = 1 (alpha = 1) reduces the weighted convolution exactly to standard convolution, every comparison in Tables 3 and 5 is a comparison between the default alpha = 1 and a per-method, per-dataset alpha chosen after seeing performance. The reported gains (e.g., VGG 56.89% to 66.94%; DnCNN 3x3 PSNR 20.17 to 22.63 dB) are therefore consistent with test-set hyperparameter fitting, not with a general property of the operator. A proper held-out validation procedure for alpha is required before these numbers can support the paper's claim.","section":"Sec. 4.1, 4.2; Tables 3 and 5"},{"comment":"The standard-convolution DnCNN baseline with a 5x5 kernel reports PSNR 12.15 dB, NRMSE 0.266, and SSIM 0.402. With Gaussian noise of sigma = 0.01 added to images in the [0,1] range, the noisy input itself has PSNR around 40 dB, so this baseline is far below the input quality and indicates a broken training or evaluation pipeline for that configuration. Because the paper's conclusion that weighted convolution 'improves the standard convolution under every metric' relies on this row, the corresponding improvement (12.15 to 23.35 dB) is not evidence for the method; it likely reflects a failed baseline rather than an improved operator.","section":"Table 5, DnCNN 5x5 row"},{"comment":"All results are reported as single values with no repeated runs, seeds, or variance estimates. The claimed improvements are often a few percentage points (e.g., NiN 51.96% to 52.35%, gMLP 32.21% to 32.66%), and with the stochasticity of deep learning training these differences may not be significant. Without repeated runs or at least seed information, the reader cannot distinguish a genuine improvement from random variation, especially given the small alpha grid used.","section":"Sec. 4.1, 4.2"}],"minor_comments":[{"comment":"The title and some running text contain the typo 'Denosing'; it should be 'Denoising'.","section":"Title and running text"},{"comment":"The confusion-matrix figure uses 'weighed' instead of 'weighted', and Tables 4 and 6 contain the spacing artifact 'W eighted convolution'.","section":"Fig. 1 and Tables 4, 6"},{"comment":"The dataset name is written inconsistently as 'CIF AR-100' in the text and 'CIFAR-100' in the abstract and related work; please unify the spelling.","section":"Sec. 4.1"},{"comment":"The sentence 'The density function is shared across both the image and the kernels' is unclear; the density multiplies kernel weights, so a more precise statement would be that it is shared across all kernels and all spatial positions of the input.","section":"Sec. 3.2"}],"recommendation":"reject","confidential_remarks":"The manuscript's own text admits that alpha was tuned to improve the reported metrics, and one baseline is physically implausible. Even a sympathetic reading cannot separate the operator's effect from test-set selection. I recommend rejection; if the authors redo the evaluation with a proper validation protocol, repeated runs, and corrected baselines, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick take on arXiv:2505.24558. The paper proposes a fixed rank-1 symmetric mask that rescales convolution kernel elements, with no added trainable parameters, and reports that it improves classification and denoising across several architectures. The idea is simple and the implementation is clean, with code released. But the empirical evidence, which is the entire basis of the contribution, is not sound.\n\nThe central problem is the tuning of the density parameter α. The paper explicitly states that the CINECA cluster was used to 'tune the density function to improve the accuracy of the weighted convolution and the results of the classification and denoising.' There is no validation protocol, no description of how α was selected, and no repeated runs. Since Φ=αα^T with α=1 reduces exactly to standard convolution, the tables are essentially comparing one default setting (α=1) against a per-method, per-dataset value of α chosen to maximize the reported metrics. That makes the reported gains look like fitted values, not evidence of generalization.\n\nSecond, the denoising baselines appear broken. With Gaussian noise σ=0.01 on [0,1], the noisy input has a PSNR around 40 dB. Any reasonable denoiser, including one that simply copies the input, should be at or above that level. Yet all standard-convolution PSNRs in Table 5 are well below 40 dB—for DnCNN with 5×5 kernels, it's 12.15 dB. That is orders of magnitude worse than the input, which indicates a bug in the data pipeline, the loss, or the evaluation. The corresponding weighted results are meaningless if the baseline is broken.\n\nTo the paper's credit, the method is clearly defined, the pre-computation of the mask is sensible, and the authors test across a broad set of architectures with shared hyperparameters. There is also prior art (variable-weight convolution, dynamic convolution) that makes the novelty incremental rather than fundamental.\n\nWho is this for? A reader who wants a simple convolution variant to try might find the code useful. But as a paper, it doesn't establish its central claim. I would not send this to peer review in its current form; the experimental methodology needs a major overhaul before the results can be trusted. If the authors redo the experiments with a proper validation split for α, fix the baselines, and report error bars, there might be a modest conference-level contribution. As is, the numbers are not credible.","headline":"The paper's weighted convolution is a simple rank-1 mask, but the empirical claims are undermined by acknowledged test-set tuning of the mask parameter and by denoising baselines that are clearly broken.","tokens_in":11914,"tokens_out":4389,"would_cite":false,"duration_ms":49750,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A fixed position-dependent density mask multiplied into convolution kernels improves classification and denoising across every tested network, with no added trainable parameters.","keywords":["Convolution","Density function","Denoising","Classification","Deep learning","weighted convolution","rank-one mask","convolutional neural networks"],"falsifier":"Re-run the same eight models with $\\alpha$ selected on a held-out validation split and with the standard baselines re-implemented from their original public code, then report test metrics only for the chosen $\\alpha$. If VGG's 10-point gain shrinks or disappears and DnCNN's $5\\times5$ standard PSNR moves from 12.15 dB upward toward the noisy-input level, the claimed superiority would be explained by test-set peeking and a broken baseline rather than by the operator.","tokens_in":10788,"feed_emoji":"🎯","tokens_out":11161,"duration_ms":120756,"temperature":0.7,"pith_summary":"The paper asks whether a convolution can be improved simply by remembering where each pixel sits relative to the pixel being updated. Its answer is yes: multiply each kernel element by a fixed spatial density value before applying the kernel, and every tested network improves on its target metric. The density mask is chosen as a symmetric rank-one outer product, so a $3\\times3$ kernel needs one tunable scalar and a $5\\times5$ kernel needs two, and it does not add any trainable parameters. On CIFAR-100 classification, VGG accuracy rises from $56.89\\%$ to $66.94\\%$; on DIV2K denoising, DnCNN PSNR rises from $20.17$ to $22.63$ dB. If the claim holds, the operator is a cheap, architecture-agnostic upgrade that any CNN can adopt with only a small runtime overhead.","feed_headline":"Spatial mask lifts VGG accuracy 10 points, zero new parameters","feed_subtitle":"Classification and denoising improve in all eight tested networks, with zero extra trainable parameters.","key_machinery":"The central object is the weighted convolution $(I*W^{\\Phi})^f_{ij}=\\sum_{a,b}\\Phi_{ab}w^f_{ab}I_{i+a-\\cdots,j+b-\\cdots}$ of Eq. (4): a standard discrete convolution whose kernel $w^f$ is elementwise multiplied by a density mask $\\Phi$ before the inner product with the image neighbourhood. The mask is built as a symmetric rank-one outer product $\\Phi=\\alpha\\alpha^\\top$, with $\\alpha$ symmetric about the center, so a $K\\times K$ kernel is controlled by $(K-1)/2$ scalar hyperparameters, and setting $\\alpha=(1,\\dots,1)$ recovers standard convolution. The implementation precomputes $\\Phi$ and forms $W^{\\Phi}=\\Phi\\circ W$ once per iteration, adding $O(K^2F)$ operations per layer, so the operator carries its gains through a fixed spatial prior rather than through new trainable weights.","core_discovery":"On the paper's own terms, the discovery is that the position of a pixel inside the receptive field is usable signal that standard convolution throws away, and a fixed nonuniform density can harvest it. The paper claims that for every architecture and dataset tested there exists a density matrix $\\Phi=\\alpha\\alpha^\\top$ that improves the target metric: all five CIFAR-100 classifiers gain in accuracy and F1 score, and all three DIV2K denoisers gain in PSNR and most secondary quality metrics. Because $\\Phi$ multiplies existing kernel weights and is not learned, the improvement comes without increasing the number of trainable parameters; the cost is only a small runtime overhead. The method is presented as a drop-in generalization of standard convolution, which is exactly the special case $\\Phi=\\mathbf{1}$.","pith_inferences":["The paper tunes $\\alpha$ as a hyperparameter and never learns it; a natural next step is to backpropagate through $\\alpha$ so each layer learns its own position mask, which would remove the manual search entirely.","Because $\\Phi=\\alpha\\alpha^\\top$ is a separable rank-one mask, the weighted convolution is equivalent to standard convolution with a kernel premultiplied by a separable profile, which suggests the same construction transfers to 3D or 1D grids by taking products of one-dimensional profiles.","The much larger gain for VGG than for the other classifiers suggests the practical benefit may depend on how close the standard baseline is to its full potential; an ablation with equally well-tuned baselines would clarify where the gain really comes from.","If the gains replicate, the density mask can be viewed as a cheap fixed spatial prior encoding that center pixels matter more, and it could be compared with learned attention or positional encodings to see what the prior captures."],"forward_implications":["Any existing CNN can adopt the weighted convolution as a drop-in replacement for standard convolution layers, since the number of trainable parameters is unchanged and the extra per-layer cost is only $O(K^2F)$ operations.","On CIFAR-100, all five classifiers improve in accuracy and F1 score, with VGG's accuracy rising from $56.89\\%$ to $66.94\\%$.","On DIV2K denoising, all three networks improve in PSNR and most quality metrics, with DnCNN's PSNR rising from $20.17$ to $22.63$ dB under a $3\\times3$ kernel.","A weighted $5\\times5$ kernel outperforms both standard and weighted $3\\times3$ kernels, so the density mask can make larger receptive fields more useful without adding parameters.","Because uniform density recovers standard convolution, the weighted operator is a strict generalization of the standard one and can serve as the default convolution in future experiments."],"supporting_citations":[{"why":"Supplies the CIFAR-100 dataset that all classification experiments measure against.","marker":"[KH+09]"},{"why":"Supplies the DIV2K dataset that all denoising experiments measure against.","marker":"[AT17]"},{"why":"Provides the VGG architecture whose accuracy rises from 56.89% to 66.94%.","marker":"[SZ15]"},{"why":"Provides the ResNet baseline, the deepest classifier in the comparison.","marker":"[HZRS16]"},{"why":"Provides the Network-in-Network baseline used in the classification tests.","marker":"[LCY13]"},{"why":"Provides the gMLP baseline used in the classification tests.","marker":"[LDSL21]"},{"why":"Provides the GAC-SNN baseline used in the classification tests.","marker":"[QZC+24]"},{"why":"Provides the DnCNN baseline whose PSNR rises from 20.17 to 22.63 dB.","marker":"[ZZC+17]"},{"why":"Provides the NAFNet baseline used in the denoising tests.","marker":"[CCZS22]"},{"why":"Provides the CGNet baseline used in the denoising tests.","marker":"[GJS+24]"}],"fun_headline_variants":["Pixel position boosts CNN accuracy 10 points, no new weights","Weighted convolution lifts VGG by 10 accuracy points, zero added parameters","Where a pixel sits in a kernel is free signal for CNNs","Fixed spatial weights improve every tested CNN, zero extra training","Density-weighted kernels: +10% accuracy, +2.5 dB PSNR, same params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that the density values $\\alpha$ reported in Tables 3 and 5 were chosen through a validation-based search without looking at the test metrics, and that the standard-convolution baselines were implemented correctly.","fun_headline_variants_meta":{"raw":{"variants":["Pixel position boosts CNN accuracy 10 points, no new weights","Weighted convolution lifts VGG by 10 accuracy points, zero added parameters","Where a pixel sits in a kernel is free signal for CNNs","Fixed spatial weights improve every tested CNN, zero extra training","Density-weighted kernels: +10% accuracy, +2.5 dB PSNR, same params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1349,"prompt_tokens":1040,"completion_tokens":309,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":211}},"tokens_in":656,"tokens_out":309,"duration_ms":4374,"temperature":1.0,"reasoning_tokens":211,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:18:08.241017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same eight models with $\\alpha$ selected on a held-out validation split and with the standard baselines re-implemented from their original public code, then report test metrics only for the chosen $\\alpha$. If VGG's 10-point gain shrinks or disappears and DnCNN's $5\\times5$ standard PSNR moves from 12.15 dB upward toward the noisy-input level, the claimed superiority would be explained by test-set peeking and a broken baseline rather than by the operator.","supporting_citations":[],"review_version":1}