{"id":"f2d103cb-0c6b-4c49-ad06-c95d7173d0cf","arxiv_id":"2501.08199","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EmoNeXt, a ConvNeXt variant using spatial transformers, squeeze-and-excitation blocks, and a self-attention variance regularizer, reports 76.12% accuracy on FER2013.","lead":"The paper describes a neural network for recognizing facial emotions that combines an existing image model with attention modules and a new regularization term. On the FER2013 benchmark it reports 76.12% accuracy, slightly above the previous best published result.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is a 0.15-point edge in a mixed-protocol table; even if protocols were matched, the gap is below one standard error and cannot establish superiority.","rationale":"The paper is a clean architecture write-up with public code, but its main empirical assertion — a new SOTA on FER2013 — is not secured by Table II. The reader's conditional verdict already captures the protocol-comparability issue; my stress test sharpens it: even granting comparability, 76.12% versus 75.97% on N=3589 is within binomial noise. The absence of lambda and the lack of ablations further prevent evaluating the novel SA term. Because the code is public and the proposed rerun is straightforward, conditional accept remains the right verdict; my read does not change the reader's assessment.","tokens_in":8597,"tokens_out":4239,"duration_ms":43416,"concrete_test":"Fetch the released repository and rerun the closest competitor, Segmentation VGG-19, under the authors' exact protocol (ImageNet-22k pretraining, resize to 224, RandomCrop/RandomRotation, AdamW 1e-4, cosine schedule, label smoothing, EMA, mixed precision) on FER2013, with at least three seeds. Then compare EmoNeXt-XLarge and this rerun using a two-proportion z-test or bootstrap on test predictions, and report Wilson 95% confidence intervals. If the intervals overlap, or if the difference does not survive seed variation, the sentence claiming EmoNeXt 'surpasses the current best state-of-the-art' should be replaced with a claim about matched-protocol improvement and statistical equivalence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — EmoNeXt-XLarge at 76.12% beats Segmentation VGG-19 at 75.97% (Section IV-C, Table II) — is the entire basis for 'superiority over state of the art.' Two independent weaknesses combine. First, the introduction states comparisons are made 'under the same experimental setup,' but Table II mixes published numbers from at least six independent papers ([14], [15], [16], [19], [36]–[40]) with different augmentations, pretraining, and evaluation details; only the ConvNeXt and EmoNeXt rows are generated by this work. Second, even if the protocols were identical, the reported margin is 0.15 percentage points on a 3,589-image test set, which is about five images. With accuracy near 76%, the standard error of the difference of two proportions is roughly sqrt(2*0.76*0.24/3589) ≈ 1.0 percentage point, so 0.15 points is around 0.15 standard errors. The claimed SOTA advantage is therefore not statistically resolvable from the numbers reported, independent of protocol comparability. The paper also never states the value of lambda in Eq. (3), so the SA regularization term cannot be reproduced or evaluated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EmoNeXt, an adaptation of the ConvNeXt architecture for facial expression recognition on FER2013. The model adds a Spatial Transformer Network (STN), Squeeze-and-Excitation (SE) blocks after each ConvNeXt stage, and a self-attention regularization term (SA) added to the cross-entropy loss. The authors report results for five model sizes (Tiny through XLarge) and claim that EmoNeXt-XLarge reaches 76.12% accuracy, surpassing the previous best of 75.97% from Segmentation VGG-19. The manuscript includes architecture descriptions, training details, and a comparison table, plus a link to code.","tokens_in":8857,"tokens_out":2611,"duration_ms":26903,"significance":"If the reported accuracy is reproducible under matched experimental conditions, EmoNeXt would represent a modest but legitimate state-of-the-art result on FER2013, a widely used benchmark. The paper's strengths include a clear architectural description, the release of code, and evaluation across multiple model scales. However, the central claim rests on a single accuracy number per model with no variance estimates, and the comparison mixes results from independently published baselines while claiming a common experimental setup. The significance of the claimed improvement is therefore not established by the present evidence; the contribution is better viewed as an architecture proposal whose empirical validation needs substantial strengthening.","major_comments":[{"comment":"The claim of state-of-the-art performance relies on comparing EmoNeXt rows against published baseline numbers from at least six different papers ([14], [15], [16], [19], [36]-[40]), yet the Introduction states that comparisons are made \"under the same experimental setup\" and Section IV-C repeats this wording. Only the ConvNeXt and EmoNeXt rows appear to be generated by this work. Different pretraining, augmentation, and evaluation protocols across these baselines make the 0.15 percentage point margin over Segmentation VGG-19 non-comparable as reported. The authors should either reproduce the baselines in their own pipeline or temper the superiority claim accordingly.","section":"Section IV-C and Table II"},{"comment":"No error bars, multiple seeds, or statistical tests are provided for any accuracy number. With a test set of 3,589 images and accuracy near 76%, the standard error of the difference between two proportions is roughly sqrt(2 * 0.76 * 0.24 / 3589) ≈ 1.0 percentage point, so the reported margin of 0.15 percentage points is about 0.15 standard errors. The claimed advantage over Segmentation VGG-19 is thus not statistically resolvable from the reported numbers, independent of protocol comparability. The authors should report the mean and standard deviation over several runs and perform a significance test for the key comparison.","section":"Section IV-C and Table II"},{"comment":"The value of the regularization weight lambda in Eq. (3) is never reported, nor is the selection procedure. Without this value, the self-attention regularization term cannot be reproduced or evaluated. Additionally, no ablation study isolates the contributions of the STN, the SE blocks, and the SA term; since the paper's contribution is precisely these additions, an ablation is necessary to support the claim that each component helps.","section":"Section III-E, Eq. (3), and Section IV-B"},{"comment":"Training details are incomplete: the batch size, number of epochs, exact learning rate schedule, data augmentation hyperparameters, and STN/SE configuration are not specified, making exact reproduction difficult. Furthermore, the Conclusion states that \"a detailed study is underway and will be published in a forthcoming journal paper,\" which indicates that the present results are preliminary. The authors should either present the full experimental protocol or explicitly frame the current manuscript as a preliminary report.","section":"Section IV-B and Conclusion"}],"minor_comments":[{"comment":"The dataset image size is given as \"482 pixels\"; the FER2013 images are 48×48 pixels. This appears to be a typographical error.","section":"Section IV-A"},{"comment":"The text mentions resizing images to \"2242\"; this is likely intended as 224×224 pixels.","section":"Section IV-B"},{"comment":"There is a typo: \"acheive\" should be \"achieve\".","section":"Section III-B"},{"comment":"The caption reads \"Architecture designes for ConvNeXt and EmoNeXt\"; \"designes\" should be \"designs\".","section":"Figure 4 caption"},{"comment":"The ordering of rows in Table II is partly non-monotonic. While this may reflect grouping by accuracy, the mixing makes visual scanning harder; a strictly sorted table with confidence intervals would be clearer.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like a short conference paper with a strong architecture-design idea but insufficient empirical validation. The main claim of state-of-the-art accuracy is not statistically supported as presented, and the comparison protocol is questionable. I would encourage the editors to require a revision that adds matched baseline reproduction, variance reporting, ablations, and the missing hyperparameters. The authors' own statement that a detailed study is forthcoming supports the view that this version is preliminary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean architecture study with a new SOTA claim that probably doesn't survive contact with statistics. The paper's real contribution is the specific combination of STN, SE blocks, and a self-attention variance penalty on ConvNeXt, plus a public code release. But the claimed edge over the previous best is 0.15 percentage points, which is about five images on the FER2013 test set, and the standard error of that difference is roughly 1 point. The margin is well within noise.\n\nWhat it does well: the write-up is clear, the components are standard and well explained, and the authors report consistent wins of EmoNeXt over their own ConvNeXt baselines across all five sizes (Tiny through XLarge). That internal consistency is a genuine, if modest, signal. The code is public, which is real evidence and should be credited.\n\nThe soft spots are proportionate to the claim. First, the headline superiority over Segmentation VGG-19 (76.12 vs 75.97) is not statistically resolvable from the numbers given, independent of protocol comparability. Second, the paper claims comparisons 'under the same experimental setup,' but Table II mixes published numbers from at least six different papers with different augmentations, pretraining, and evaluation details. Only the ConvNeXt and EmoNeXt rows are generated by this work. Third, the value of lambda in Eq. (3) is never stated, so the SA regularization term can't be reproduced or evaluated. Fourth, there's no ablation isolating STN, SE, or SA, so we don't know which component earns the gain. These are fixable, but they're real gaps.\n\nThe paper is clearly written and the citations are on point; I don't see any missing major references. It's an honest empirical study, just with a comparison table that overreaches.\n\nWho it's for: anyone tracking FER2013 leaderboards or looking for a ConvNeXt variant to try. It shouldn't be cited as established SOTA until the protocol-matching and statistics are addressed, but the code makes verification possible. A serious referee could sort this out. I'd send it to review with a request for matched-protocol comparisons, error bars or multiple seeds, the lambda value, and an ablation.","headline":"Clean architecture study with a SOTA claim that's within noise; worth a referee but not a citation as established.","tokens_in":9369,"tokens_out":2001,"would_cite":false,"duration_ms":18616,"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":"EmoNeXt claims 76.12% on FER2013, a new state of the art","keywords":["Facial expression recognition","FER2013","ConvNeXt","Spatial transformer network","Squeeze-and-excitation","Self-attention regularization","Emotion classification","Deep learning"],"falsifier":"Re-run EmoNeXt-XLarge and the strongest Table II baselines, especially Segmentation VGG-19, on the same FER2013 test images using identical preprocessing, augmentation, pretrained weights, and evaluation code. If any baseline then matches or exceeds 76.12%, or if EmoNeXt-XLarge drops below 75.97% under the shared protocol, the paper's central superiority claim is refuted.","tokens_in":1806,"feed_emoji":"😊","tokens_out":2369,"duration_ms":80422,"temperature":0.7,"pith_summary":"This paper claims that a convolutional architecture built from ConvNeXt, with a Spatial Transformer Network in front, Squeeze-and-Excitation blocks between stages, and a self-attention regularization term in the loss, sets a new state of the art on the FER2013 facial emotion recognition benchmark. The largest variant, EmoNeXt-XLarge, reports 76.12% test accuracy, edging past the previous best reported by Segmentation VGG-19 at 75.97%. The same recipe improves all five model sizes over their plain ConvNeXt counterparts, so the authors argue that the gains come from the added components rather than from raw scale. A sympathetic reader would care because FER2013 is a widely used, unbalanced, in-the-wild benchmark, and any reproducible improvement on it is useful for applications such as human-machine interaction, healthcare, and driver monitoring.","feed_headline":"EmoNeXt claims 76.12% on FER2013, a new state of the art","feed_subtitle":"Spatial transformers and channel recalibration push a ConvNeXt variant 0.15 points past the best FER2013 score.","key_machinery":"The central object is the EmoNeXt architecture: a ConvNeXt backbone preceded by a Spatial Transformer Network and followed at each stage by a Squeeze-and-Excitation block, trained with cross-entropy plus a self-attention regularization term. The STN is a differentiable module that predicts a spatial transformation and resamples the input, letting the network align faces despite scale, rotation, and translation variation. The SE blocks squeeze each channel to a scalar with global average pooling and excite channel-wise weights through fully connected layers, adaptively emphasizing informative channels. The self-attention regularization term is the variance of the dot-product attention weights, added to the loss with a trade-off hyperparameter $\\lambda$; it drives the attention distribution toward uniformity so features are used more evenly. Together these components are what the paper credits for the accuracy gains over plain ConvNeXt and earlier FER models.","core_discovery":"The central claim is that adding three mechanisms to ConvNeXt yields a better facial emotion classifier: an STN that learns affine-style spatial transformations to align faces, SE blocks that recalibrate channel-wise feature responses after each ConvNeXt stage, and a self-attention regularization term. The regularization term is $L_{\\mathrm{SA}} = \\frac{1}{N}\\sum_{i=1}^{N}(W_i - \\bar{W})^2$, the variance of dot-product self-attention weights, added to cross-entropy as $L_{\\mathrm{final}} = L_{\\mathrm{CE}} + \\lambda L_{\\mathrm{SA}}$. Minimizing this variance pushes attention weights toward their mean, encouraging balanced, compact feature vectors. On the FER2013 test set, EmoNeXt-XLarge reaches 76.12% accuracy, surpassing the 75.97% reported for Segmentation VGG-19, and every EmoNeXt size outperforms the corresponding ConvNeXt size.","pith_inferences":["The comparison in Table II mixes accuracy numbers reported by different groups with different training protocols; a direct head-to-head re-run on the same FER2013 test split is the cleanest way to see whether the 0.15-point edge survives.","Because the self-attention regularization term is defined on feature-vector attention weights and added to any cross-entropy loss, it is not tied to ConvNeXt; it could be tested as a plug-in auxiliary loss for ResNet, VGG, or transformer-based FER models.","The reported improvements are all on FER2013. The paper's stated next step, emotion recognition for Alzheimer's patients, would require cross-dataset and population-specific validation before the architecture's practical value in clinical settings is established.","If the SA regularization indeed works by flattening attention distributions, one testable extension is to ablate $\\lambda$ and visualize the attention maps: the claim predicts less concentrated attention and more evenly distributed feature use as $\\lambda$ grows."],"forward_implications":["EmoNeXt-XLarge claims 76.12% on FER2013, a 0.15 percentage-point gain over Segmentation VGG-19's 75.97%, making it the paper's new single-model state of the art.","The same three additions lift every size of ConvNeXt: EmoNeXt-Tiny reaches 73.34%, Small 74.33%, Base 74.91%, Large 75.57%, and XLarge 76.12%.","Smaller EmoNeXt variants outperform larger plain ConvNeXt models, e.g., EmoNeXt-Small (74.33%) beats ConvNeXt-XLarge (74.15%), suggesting the architectural additions matter more than parameter count.","The full training recipe, including ImageNet-22k pretrained weights, AdamW, cosine decay, stochastic depth, label smoothing, EMA, and mixed precision, is what produced these numbers, so the result is tied to modern training practices, not only the architecture."],"supporting_citations":[{"why":"Supplies the ConvNeXt backbone that EmoNeXt adapts; all accuracy claims are measured against ConvNeXt variants.","marker":"[9]"},{"why":"Defines the Spatial Transformer Network placed at the input of EmoNeXt to align faces.","marker":"[10]"},{"why":"Defines Squeeze-and-Excitation blocks inserted after each ConvNeXt stage for channel recalibration.","marker":"[11]"},{"why":"Provides the FER2013 dataset and its train/validation/test split used in all experiments.","marker":"[13]"},{"why":"Gives the CNNs + BOVW baseline (75.42%) that EmoNeXt-Large surpasses.","marker":"[14]"},{"why":"Provides the LHC-Net baselines (74.28% and 74.42%) that EmoNeXt-Small and Base surpass.","marker":"[15]"},{"why":"Provides the Ad-Corre loss baseline (72.03%) that EmoNeXt-Tiny beats.","marker":"[16]"},{"why":"Reports the previous best single-model result, Segmentation VGG-19 at 75.97%, which EmoNeXt-XLarge claims to surpass.","marker":"[19]"},{"why":"Defines the AdamW optimizer used in training, part of the training recipe behind the reported scores.","marker":"[30]"},{"why":"Supplies the ImageNet-22k pretrained weights that initialize the ConvNeXt backbones.","marker":"[35]"}],"fun_headline_variants":["EmoNeXt sets new FER2013 benchmark at 76.12%","Spatial transformer and SE blocks push EmoNeXt to 76.12%","Self-attention regularization propels ConvNeXt to top FER2013","EmoNeXt outperforms prior SOTA on FER2013 by 0.15 points","Adapted ConvNeXt with STN and SE achieves 76.12% on FER2013"],"cache_read_input_tokens":11520,"weakest_assumption_plain":"The load-bearing premise is that the published accuracy numbers for the earlier models in Table II are accurate and directly comparable to EmoNeXt's, even though they come from different groups with different training protocols; the claimed superiority rests on a mere 0.15 percentage-point margin over the previous best.","fun_headline_variants_meta":{"raw":{"variants":["EmoNeXt sets new FER2013 benchmark at 76.12%","Spatial transformer and SE blocks push EmoNeXt to 76.12%","Self-attention regularization propels ConvNeXt to top FER2013","EmoNeXt outperforms prior SOTA on FER2013 by 0.15 points","Adapted ConvNeXt with STN and SE achieves 76.12% on FER2013"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000653,"raw_usage":{"total_tokens":2967,"prompt_tokens":893,"completion_tokens":2074,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":1957}},"tokens_in":509,"tokens_out":2074,"duration_ms":13321,"temperature":1.0,"reasoning_tokens":1957,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:28:40.054758+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run EmoNeXt-XLarge and the strongest Table II baselines, especially Segmentation VGG-19, on the same FER2013 test images using identical preprocessing, augmentation, pretrained weights, and evaluation code. If any baseline then matches or exceeds 76.12%, or if EmoNeXt-XLarge drops below 75.97% under the shared protocol, the paper's central superiority claim is refuted.","supporting_citations":[{"cited_title":"Chal- lenges in representation learning: A report on three machine learning contests,","cited_arxiv_id":null,"evidence_quote":"Provides the FER2013 dataset and its train/validation/test split used in all experiments."},{"cited_title":"Local learning with deep and handcrafted features for facial expression recognition,","cited_arxiv_id":null,"evidence_quote":"Gives the CNNs + BOVW baseline (75.42%) that EmoNeXt-Large surpasses."},{"cited_title":"Local multi-head channel self- attention for facial expression recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the LHC-Net baselines (74.28% and 74.42%) that EmoNeXt-Small and Base surpass."},{"cited_title":"Ad-corre: Adaptive correlation-based loss for facial expression recognition in the wild,","cited_arxiv_id":null,"evidence_quote":"Provides the Ad-Corre loss baseline (72.03%) that EmoNeXt-Tiny beats."},{"cited_title":"A novel facial emotion recognition model using segmentation vgg-19 architecture,","cited_arxiv_id":null,"evidence_quote":"Reports the previous best single-model result, Segmentation VGG-19 at 75.97%, which EmoNeXt-XLarge claims to surpass."}],"review_version":1}