{"id":"3f8979fc-8bbf-493e-8e98-90a76526090f","arxiv_id":"2411.13582","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ResCNet adds Gaussian CDF-based confidence values to residual features, giving small and sometimes inconsistent accuracy improvements over ResNet and SENet.","lead":"This paper introduces ResCNet, a plug-in module that recalibrates ResNet feature responses using a Gaussian cumulative distribution, and reports modest accuracy gains on several image benchmarks. A generalist might read it because it is a simple architectural variant of widely used ideas, GELU and SENet, with results that are small and not consistent across network depths.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that ResCNet improves accuracy over SENet is contradicted by Table 1: on CIFAR-100, ResCNet-56 (73.81%) is 0.87 points below SENet-56 (74.68%), and most reported gains are small, unrepeated, and not epoch-matched on ImageNet.","rationale":"The reader's weakest-assumption pick (the Gaussian distribution in Section 3) is a genuine weakness in the theoretical framing, but it is not the most load-bearing condition for the central claim. The RC layer's empirical behavior does not collapse if the true distribution is non-Gaussian: the learned mean and standard deviation still define a smooth, learnable nonlinear calibration, so the accuracy comparison can be assessed independently of the Gaussianity assumption. The decisive issue is experimental evidence. Table 1 internally contradicts the claim of superiority over SENet, and the ImageNet comparison is not epoch-matched, so the claimed advantage could be an artifact of training protocol or noise. The reader's overall CONDITIONAL verdict is appropriate: the method is plausible and reproducible, but the comparisons must be controlled and the claims qualified. My concern reinforces the reader's verdict rather than changing it, so the verdict remains CONDITIONAL (UNCHANGED).","tokens_in":13060,"tokens_out":4504,"duration_ms":47428,"concrete_test":"Rerun the exact Table 1 protocol for ResNet-56, SENet-56, and ResCNet-56 on CIFAR-100 with at least 5 random seeds, reporting mean ± standard deviation and paired differences. If SENet-56 still exceeds ResCNet-56 by more than the seed noise, the claim that ResCNet outperforms SENet is false for at least one depth, and the paper must be revised to state a qualified or restricted version of its central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim—that ResCNet improves classification accuracy over both ResNet and SENet—requires controlled, consistent comparisons. The paper's own Table 1 contradicts this claim: on CIFAR-100, ResCNet-56 reaches 73.81% while SENet-56 reaches 74.68%, and even ResNet-56 reaches 73.79%. So for at least one architecture depth, the proposed method is worse than the SENet baseline it claims to beat. Elsewhere the reported gains are tiny (0.02–0.19 points) and no standard deviations, repeated seeds, or significance tests are reported, making it impossible to distinguish a real effect from run-to-run noise. Table 2 is not an epoch-matched comparison: ResCNet-50 is trained for 90 epochs while the SENet-50 row is the published 100-epoch result, and the ResNet rows mix TorchVision and re-implementation numbers. The Gaussian assumption in Section 3 is indeed weakly justified, but it is not the most load-bearing issue: even if the responses are not Gaussian, the calibration layer is still a smooth learned nonlinearity whose empirical effect can be evaluated independently. The load-bearing issue is that the comparative evidence does not currently support the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Response Calibration Networks (ResCNet), a plug-in module that estimates a per-channel Gaussian model for convolutional feature responses, converts each response into a confidence weight via the standard normal CDF (with symmetric complementary weighting for above-mean responses), multiplies the original response by this weight, and adds the calibrated value to the original feature. The module is inserted after the second BatchNorm in a ResNet block, in a configuration analogous to but distinct from the SE layer. Experiments on CIFAR-10, CIFAR-100, SVHN, and ImageNet compare ResCNet with ResNet and SENet; a related activation-function variant, GCLU, is also evaluated across several architectures. The paper claims improved classification accuracy and improved feature discriminability.","tokens_in":13365,"tokens_out":9820,"duration_ms":97347,"significance":"If the central empirical comparison were clean, this would be a simple, plug-compatible feature-calibration module and a modestly interesting alternative to channel attention. The paper has concrete strengths: the code is publicly available, the experimental coverage spans four datasets, and the ablation study is fairly broad, including integration schemes, CDF approximations, and an activation-function variant. However, the current evidence does not establish the central claim. One of the key comparisons in the paper's own Table 1 contradicts the claim that ResCNet outperforms SENet; the ImageNet comparison is not epoch-matched; and no repeated-seed statistics or significance tests are reported. The conceptual novelty is also limited by the fact that the mean and standard deviation are learned parameters, so the calibrated transformation is a learned nonlinearity whose Gaussian interpretation is motivational unless quantitative distributional evidence is supplied.","major_comments":[{"comment":"The central claim that 'as the depth of the model increases, the proposed ResCNet outperforms both ResNet and SENet' is contradicted by Table 1: on CIFAR-100, ResCNet-56 reaches 73.81%, which is 0.87 points below SENet-56 (74.68%) and only 0.02 points above ResNet-56 (73.79%). On CIFAR-10, ResNet-56 and SENet-56 both reach 94.51% while ResCNet-56 reaches 94.61%, a difference too small to count as an improvement in the absence of repeated runs. Please reconcile the stated claim with these numbers or explicitly narrow the scope of the claimed advantage.","section":"Section 5.2.1, Table 1"},{"comment":"The ImageNet comparison is not controlled. SENet-50 is a published 100-epoch result, whereas ResCNet-50 is trained for 90 epochs; the ResNet-18 and ResNet-34 rows mix TorchVision pretrained checkpoints with in-house re-implementations; and parameter counts differ across methods. Without all baselines retrained under identical epochs, schedulers, augmentation, and hyperparameters, the statement that ResCNet-50 'outperformed SENet-50' is not justified. Please provide an epoch-matched comparison or both 90- and 100-epoch SENet baselines.","section":"Table 2 and Section 5.2.1"},{"comment":"No standard deviations, multiple seeds, or significance tests are reported anywhere in the paper. Most reported gains are fractions of a percentage point, for example Table 1 shows CIFAR-10 ResCNet-32 at 93.79% versus SENet-32 at 93.73%, and Table 3 shows a baseline difference of 80.79% versus 80.87%. These differences are within typical run-to-run variation for CIFAR training. Please report at least three independent seeds with means and standard deviations, or provide a significance test, before claiming superiority over the baselines.","section":"Section 5.2.1 and Tables 1-3"},{"comment":"The Gaussian assumption is supported only by visual inspection of a single violin plot and a CLT argument that does not account for dependencies, ReLU truncation, or the fact that the mean and standard deviation are learned parameters. Because Eq. (4) defines the calibration weights as Gaussian CDF integrals, this assumption is load-bearing for the stated rationale of the mechanism. Please provide a quantitative distributional test (for example, a goodness-of-fit check across multiple layers and datasets), or explicitly reframe the RC Layer as a learned nonlinearity whose Gaussian interpretation is only heuristic.","section":"Section 3 and Eqs. (1)-(4)"},{"comment":"The claim that the calibrated features are 'more discriminative' is supported only by red dashed bounding boxes in a single visualization. Since discriminability is central to the paper's title and contribution, please report a quantitative metric, such as between-class versus within-class feature variance, k-nearest-neighbor classification accuracy on frozen features, or a margin-based measure, to substantiate the claim.","section":"Section 5.2.2, Fig. 5"}],"minor_comments":[{"comment":"The parameter-count argument for the three-FC-layer RC Layer states that 'the total parameter count becomes 3 * C * r'. For the described C -> C/r -> C architecture, the dominant term is 2*C^2/r, so the formula and the accompanying reasoning should be corrected and recomputed.","section":"Section 5.3.1"},{"comment":"The table header 'Params (K) FLOPs (M)' is shared between the 'CIFAR-10 & SVHN' and 'CIFAR-100' dataset columns in a confusing way. Please separate architecture statistics from dataset-specific classifier statistics and clarify how the FLOPs are computed.","section":"Table 1"},{"comment":"AlexNet with GELU on CIFAR-100 is reported at 1.23% accuracy, which strongly suggests a training failure. This row should either be discussed in the text or removed, because it distorts the activation-function comparison.","section":"Table 5"},{"comment":"The claim of being 'the first to achieve over 86% accuracy on the CIFAR-100 dataset with a 50-layer ResNet model variant' needs a verified, dated comparison with other published results rather than a leaderboard footnote, especially because the hyperparameters were selected on the same test set.","section":"Section 5.2.3"},{"comment":"Momentum, weight decay, and exact data-augmentation protocols for CIFAR and SVHN are not reported, although these choices materially affect the reported accuracy values. Please specify them or refer explicitly to the released code.","section":"Section 5.1.2"},{"comment":"There are numerous typographical and formatting issues, including 'Comparision' in Section 5.2.1, 'CIF AR' in place of 'CIFAR', 'T op-1' in Figure 4, and inconsistent capitalization of 'CaIT' versus 'Cait'. These should be corrected in a final revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of an applied deep learning journal and is accompanied by public code, but the central comparative claim needs substantial rework. I would ask the editor to verify the 'first to achieve over 86%' statement against current CIFAR-100 leaderboards before publication. In addition, the related-work section contains several citations (for example, references to fractal-wavelet modeling and metaheuristic optimizers) that appear only loosely connected to the technical content; please ask the authors to justify their relevance or remove them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea is a small, clearly described variant of GELU applied per-channel, and the authors are honest that it is GELU-like. The RC Layer adds a learned mean/variance Gaussian CDF weight to feature responses, which is close to SENet but with additive calibration rather than multiplicative scaling. The GCLU activation is a natural extension, and the experiments across AlexNet, VGG, ResNet, ViT, Swin, and CaiT are a useful empirical sweep. Code is linked.\n\nThe soft spot is the headline empirical claim. Table 1 shows ResCNet-56 at 73.81% on CIFAR-100, below SENet-56's 74.68% and basically tied with ResNet-56. So \"outperforms both ResNet and SENet\" is false for that configuration. Elsewhere gains are 0.02–0.19 points, with no error bars or repeated seeds, so they could easily be noise. ImageNet comparisons are not epoch-matched: the SENet-50 row is the published 100-epoch result while ResCNet-50 runs 90 epochs, and the ResNet-50 rows mix TorchVision and re-implementation numbers. That is not a fair comparison. The \"first to exceed 86% on CIFAR-100 with a 50-layer ResNet variant\" claim is unsupported and likely wrong; the footnote to PapersWithCode does not establish priority.\n\nThe Gaussian distributional assumption in Section 3 is weakly supported (one violin plot plus CLT handwaving), but I agree with your stress-test that it is not the load-bearing issue: even if the responses are not Gaussian, the layer is a smooth learned nonlinearity that could help or hurt, and the empirical results are what matter. So the real flaw is the comparative evidence, not the assumption.\n\nWho is this for? Someone working on activation functions or channel calibration might find the GCLU ablation interesting as a data point. But the central contribution—the ResCNet module—is not shown to be reliably better than a well-tuned SENet baseline. I would send it to peer review, but with a clear request to fix the comparisons, add multiple seeds with error bars, and remove the overclaims. A reviewer should not reject on the basis of novelty alone; the method is a legitimate variant and the code is a plus.\n\nRecommendation: conditional acceptance, but the authors need to do a controlled, epoch-matched comparison with proper uncertainty quantification.","headline":"A clearly described GELU/SENet variant whose headline accuracy claim is contradicted by its own Table 1; the GCLU ablation is a useful empirical sweep, but the comparisons need fixing before the central claim can be taken at face value.","tokens_in":13907,"tokens_out":2424,"would_cite":false,"duration_ms":22480,"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":"Calibrating feature responses with Gaussian confidence values improves classification accuracy and feature distinctiveness over ResNet and SENet.","keywords":["Deep Neural Network","Response Value Calibration","Gaussian Distribution","Feature Discriminability","ResCNet","Activation Function","Image Classification"],"falsifier":"Retrain ResCNet-32 on CIFAR-100 with the learnable Gaussian CDF weight replaced by a fixed, symmetric, non-Gaussian weight of the same shape (for instance, a triangular function of the standardized value); if accuracy is unchanged, the specific Gaussian assumption is not the source of the improvement.","tokens_in":12857,"feed_emoji":"🎯","tokens_out":10039,"duration_ms":92401,"temperature":0.7,"pith_summary":"The paper proposes that a neural network's classification accuracy can be improved by calibrating each convolutional kernel's response values against a per-kernel Gaussian model. It claims that response values for a kernel are approximately Gaussian-distributed, and therefore a confidence weight can be computed for each response as a symmetric integral of the standard normal CDF around the learned mean. Multiplying the original response by this weight and adding the product back into the feature map yields a plugin 'Response Calibration Layer' that, inserted into ResNet blocks, produces ResCNet. The paper reports that ResCNet improves accuracy and feature discriminability over ResNet and SENet on CIFAR-10, CIFAR-100, SVHN, and ImageNet, and that with standard training tricks it exceeds 86% on CIFAR-100 with a ResNet-50 variant. A sympathetic reader would care because the method offers a distribution-based, parameter-lean alternative to scaling-based attention modules like SENet.","feed_headline":"Calibrating responses with Gaussian confidence improves CNN accuracy","feed_subtitle":"A plug-in layer adds per-neuron Gaussian confidence to residual features, improving CIFAR and ImageNet accuracy.","key_machinery":"The central object is the response calibration weight $w_k$ defined by the folded Gaussian CDF: for a feature response $a_k$ with per-kernel mean $\\mu_k$ and standard deviation $\\sigma_k$, let $x=(a_k-\\mu_k)/\\sigma_k$; then $w_k = \\Phi(x)$ for $a_k\\le\\mu_k$ and $w_k = 1-\\Phi(x)$ for $a_k>\\mu_k$, where $\\Phi$ is the standard normal CDF. This weight is close to 1 near the mean and falls toward 0 in the tails, so responses that are typical of the kernel are kept at nearly full strength while atypical responses are suppressed. The calibration value $c_k = a_k w_k$ is added to the original response, and a module containing a global-pooling step plus two or three fully-connected layers learns the per-kernel $\\mu_k$ and $\\sigma_k$. This module is inserted after the final BatchNorm in each residual block, mirroring the SE layer's position, but with addition rather than multiplicative scaling, giving the RC layer and the ResCNet architecture.","core_discovery":"The central claim is that each convolutional kernel's response values follow a Gaussian distribution $\\mathcal{N}(\\mu_k, \\sigma_k^2)$, and that this law can be exploited by defining a confidence weight for each response $a_k$ as the Gaussian CDF evaluated at the standardized value $x=(a_k-\\mu_k)/\\sigma_k$, with the weight mirrored for values above the mean: $w_k = \\Phi(x)$ if $a_k \\le \\mu_k$ and $w_k = 1-\\Phi(x)$ otherwise. The calibration value is $c_k = a_k w_k$, and the RC layer adds this calibration value to the original feature (not just scales it, as SENet does). The paper develops this into a plugin module placed after the last BatchNorm in a residual block, with two or three fully-connected layers fitting $\\mu_k$ and $\\sigma_k$ per kernel, and calls the resulting architecture ResCNet. Across CIFAR-10/100, SVHN, and ImageNet, ResCNet is reported to beat both ResNet and SENet baselines in accuracy, and the paper claims stronger feature distinctiveness as shown by response-value distributions. On CIFAR-100 with ResNet-50 and a series of standard tricks, ResCNet reaches 86.31%, which the authors state is the first time a 50-layer ResNet variant has exceeded 86% without extra data or pre-training.","pith_inferences":["The Gaussian assumption is asserted from a single violin plot and an appeal to the Central Limit Theorem; a natural extension would be to test other distributional families (e.g., logistic or heavy-tailed) for the same folded-weight construction, since the method's rationale would change if the true distributions deviate markedly from Gaussian.","Because the calibration values are added to the original features rather than multiplied, the RC layer can be viewed as injecting a learned 'typicality' signal; this is conceptually similar to a residual attention mechanism, and one could test whether the improvement comes from the distributional modeling or simply from the extra non-linearity and parameters.","The method could be probed on tasks beyond classification, such as object detection or segmentation, where feature distinctiveness matters, to see if the calibration transfers."],"forward_implications":["Because the RC layer is a plugin, the calibration mechanism can in principle be inserted into any convolutional backbone, not only ResNet.","The paper's activation-function variant, GCLU, applies the same symmetric CDF weighting as a standalone activation and is reported to improve accuracy on shallow CNNs like AlexNet that lack BatchNorm.","The reported 86.31% on CIFAR-100 with ResNet-50, if reproducible, marks a new performance point for a 50-layer ResNet variant without extra training data.","The method suggests that feature discriminability can be improved by distribution-based response correction rather than by the scaling-only approach of SENet."],"supporting_citations":[{"why":"ResNet serves as the base architecture that ResCNet modifies; the RC layer is inserted into residual blocks, so the comparison and the architecture itself depend on it.","marker":"[4]"},{"why":"SENet is the primary baseline and the SE layer is the direct contrast: the paper argues the RC layer improves on SENet's scaling-only calibration by adding calibration values.","marker":"[26]"},{"why":"GELU provides the Gaussian CDF weighting that the paper's calibration formula extends (for below-mean responses the weight is exactly GELU's); it also supplies the approximation forms used in the GCLU experiments.","marker":"[30]"},{"why":"CIFAR-10 and CIFAR-100 are the main datasets; the distributional pre-analysis is performed with ResNet-32 on CIFAR-100, establishing the Gaussian-response premise.","marker":"[39]"},{"why":"SVHN is one of the three small-scale benchmarks used to compare ResNet, SENet, and ResCNet.","marker":"[41]"},{"why":"ImageNet is the large-scale benchmark validating that the method transfers beyond small datasets.","marker":"[42]"},{"why":"ResNet-D provides the downsampling block used in the performance-limits exploration that yields the reported 86.31% CIFAR-100 result.","marker":"[44]"}],"fun_headline_variants":["Gaussian confidence weights improve CNN feature calibration","ResCNet plugin adds Gaussian CDF to residual features","Per-kernel Gaussian calibration beats ResNet and SENet","Calibrating features with Gaussian CDF lifts accuracy","Gaussian-based plugin module for discriminative feature calibration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each convolutional kernel's response values follow a Gaussian distribution; if that distributional assumption fails, the confidence values that drive the calibration are no longer meaningful and the module becomes an arbitrary nonlinearity.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian confidence weights improve CNN feature calibration","ResCNet plugin adds Gaussian CDF to residual features","Per-kernel Gaussian calibration beats ResNet and SENet","Calibrating features with Gaussian CDF lifts accuracy","Gaussian-based plugin module for discriminative feature calibration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000515,"raw_usage":{"total_tokens":2543,"prompt_tokens":1031,"completion_tokens":1512,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":1436}},"tokens_in":647,"tokens_out":1512,"duration_ms":11861,"temperature":1.0,"reasoning_tokens":1436,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:19:08.027937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain ResCNet-32 on CIFAR-100 with the learnable Gaussian CDF weight replaced by a fixed, symmetric, non-Gaussian weight of the same shape (for instance, a triangular function of the standardized value); if accuracy is unchanged, the specific Gaussian assumption is not the source of the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ResNet serves as the base architecture that ResCNet modifies; the RC layer is inserted into residual blocks, so the comparison and the architecture itself depend on it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SENet is the primary baseline and the SE layer is the direct contrast: the paper argues the RC layer improves on SENet's scaling-only calibration by adding calibration values."},{"cited_title":"Krizhevsky, G","cited_arxiv_id":null,"evidence_quote":"CIFAR-10 and CIFAR-100 are the main datasets; the distributional pre-analysis is performed with ResNet-32 on CIFAR-100, establishing the Gaussian-response premise."},{"cited_title":"Netzer, T","cited_arxiv_id":null,"evidence_quote":"SVHN is one of the three small-scale benchmarks used to compare ResNet, SENet, and ResCNet."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ImageNet is the large-scale benchmark validating that the method transfers beyond small datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ResNet-D provides the downsampling block used in the performance-limits exploration that yields the reported 86.31% CIFAR-100 result."}],"review_version":1}