{"id":"885aef22-3349-45f7-9ae6-465e36ce1bd9","arxiv_id":"2505.05943","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new attention module combining Triplet Attention and Squeeze-and-Excitation reports 78.27% on FER2013, a small gain over its own ConvNeXt baseline.","lead":"This paper proposes TripSE, an attention block that combines Triplet Attention with Squeeze-and-Excitation, and plugs it into ResNet, DenseNet, and ConvNeXt, reporting a new state-of-the-art 78.27% on the FER2013 facial expression benchmark. The gains are modest and inconsistent across datasets, and the headline number is a single run without code, seeds, or error bars.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 78.27% FER2013 SOTA rests on an unspecified test/validation split and a single run selected from eight configurations; the claim needs seeded reproduction on the official split.","rationale":"I agree with the reader's weakest assumption. The paper's own data show non-monotonic sensitivity to the SE reduction ratio and variant choice, with the best configuration surrounded by configurations below baseline. Since the authors do not specify the evaluation split and do not report seeded runs, the 78.27% number is not established as a genuine SOTA. The proposed test would settle whether the number is reproducible. The abstract's blanket claim that TripSE boosts performance is also contradicted by Table II entries (ResNet18+CIFAR100 67.08 vs 67.56; DenseNet+AffectNet 61.10 vs 62.16) and by multiple Table III entries below baseline, but that issue is secondary to the central SOTA concern. I therefore leave the reader's conditional verdict unchanged.","tokens_in":12022,"tokens_out":7846,"duration_ms":72014,"concrete_test":"Obtain the standard FER2013 split (training/validation/test) and finetune ConvNeXt-S + TripSE4 (r=1) using the paper's protocol (Table I) with at least five different random seeds, measuring accuracy on the official test split (and separately on validation). If the mean accuracy minus one standard error does not exceed the ConvNeXt-S baseline (77.19) or the 76.82 reference RMN result, or if the test/validation accuracy differs by more than about 1 point, the SOTA claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim (Section V.C, Table III) is that ConvNeXt-S + TripSE4 (r=1) reaches 78.27% on FER2013, described as an unprecedented feat. For this claim to land, 78.27 must be a valid estimate of generalization under the standard FER2013 protocol. The paper does not state whether the reported accuracy is on the official test split or the validation split, which matters because FER2013 papers often use the validation split as a proxy. The result is also the maximum over four TripSE variants and two reduction ratios (eight configurations), with no multiple-seed evidence and no code. The surrounding entries in Table III undermine the robustness interpretation: the same architecture with r=16 gives 76.71, below the ConvNeXt-S baseline of 77.19, and TripSE1 r=1 gives 76.61, also below baseline. Thus the 1.08-point win over the best single baseline is selected, not stable, and may reflect test-set overfitting or noise. The paper also does not report the standard deviation or any repetition, so the reader cannot assess whether 78.27 is distinguishable from the 77.19 baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a family of attention blocks, TripSE1-4, that combine Triplet Attention (TA) and Squeeze-and-Excitation (SE) to produce what the authors call 3D attention maps. The blocks are inserted after each group of blocks in ResNet18, DenseNet, and ConvNeXt, and the resulting models are evaluated on CIFAR100, ImageNet, FER2013, and AffectNet. The central claim is that ConvNeXt-S with TripSE4 (r=1) achieves 78.27% accuracy on FER2013, described as a new state-of-the-art and 'an unprecedented feat' (Section V.C, Table III). The paper also reports a controlled comparison on FER2013 in which TripSE1 outperforms both TA-only and SE-only variants under the same finetuning protocol.","tokens_in":12270,"tokens_out":3424,"duration_ms":35577,"significance":"If the central claim holds, the work offers a simple, parameter-efficient way to improve CNN-based facial expression recognition: the TripSE block adds only about 0.1M parameters to ConvNeXt-T (28.6M vs 28.7M) while reportedly gaining over one point on FER2013 relative to the ConvNeXt-S baseline and 1.45 points over the stated reproducible baseline of 76.82% (Section V.C, Table III). The controlled comparison on FER2013 is a genuine strength: under the same finetuning protocol, TripSE1 reaches 78.08% against 76.51% for TA and 76.93% for SE, which supports the claim that the proposed combination, rather than a generic attention module, is responsible for the gain. The significance is conditional, however, on the 78.27% headline number being a reproducible estimate of generalization; the paper provides no seeds, no error bars, and no explicit statement of whether the result is measured on the official test split. The paper does not include code or any machine-checked derivations, so the contribution rests entirely on the empirical results.","major_comments":[{"comment":"The state-of-the-art claim of 78.27% rests on a single run with no seeds, no standard deviations, and no explicit statement of whether FER2013 accuracy is measured on the official test split (3,589 images) or the validation split. Because the headline number is the maximum over four variants and two reduction ratios (eight configurations reported in Table III), the result may reflect selection on the test set rather than a genuine improvement. Please report mean and standard deviation over at least three seeds, state clearly which split is used, and either preselect the variant and r before evaluation or apply an appropriate multiple-comparison correction.","section":"Section V.C, Table III"},{"comment":"The internal evidence for the robustness of the proposal is mixed. ConvNeXt-S + TripSE1 (r=1) scores 76.61%, below the 77.19% baseline, and ConvNeXt-S + TripSE4 (r=16) scores 76.71%, also below baseline, while the r=16/r=1 counterparts achieve 78.08% and 78.27%. The text states that TripSE1 'consistently demonstrated the effectiveness' of the block and that the original ConvNeXt-S failed to gain when scaled up, but the table shows the improvement disappears or reverses for some configurations. This instability requires explanation; at minimum, all runs, the selection procedure, and the sensitivity to r should be reported and discussed honestly.","section":"Table III and Section V.C"},{"comment":"The method is specified only verbally and through figures. Terms such as 'rotational channels', 'rotational maps', 'multiplicative or translational broadcasting', and the exact position of the SE block in each variant are not defined with equations or tensor-shape diagrams. This makes the four variants impossible to reproduce or formally check. Please provide precise mathematical definitions of TripSE1-4, including where the SE block is applied, how the 1D SE vector is expanded and combined with the 2D TA attention map, and what operation occurs at branch unification.","section":"Section III.C and III.D"},{"comment":"The implementation details omit key protocol elements needed to interpret the FER2013 numbers: the evaluation split (validation vs test), any normalization used for the resized 224x224 images, and whether the learning-rate scheduler is driven by validation accuracy and then followed by evaluation on the test set. Given that the FER2013 test labels are public and that validation-set evaluation is common in this literature, the paper must state explicitly which split produced the reported numbers and how early stopping was applied.","section":"Section IV.B"}],"minor_comments":[{"comment":"The abstract says 'a new feet for this dataset' and should read 'a new feat'.","section":"Abstract"},{"comment":"The footnote contains a typo: 'Trained from scrtach' should be 'Trained from scratch'.","section":"Table II footnote"},{"comment":"The text reports the ConvNeXt-T baseline as 77.20%, while Table III reports 77.19% for the same model; the inconsistency should be corrected.","section":"Section V.C"},{"comment":"The rows for ConvNeXt-T + SE and ConvNeXt-T + TA are labeled inconsistently; the SE row has no reduction ratio, and the TA row is listed as ConvNeXt-T + TA1 but discussed as the TA block. Please add the exact hyperparameters used for these baselines.","section":"Table III"},{"comment":"Figure 2's caption contains 'visa-versa', which should be 'vice versa'.","section":"III.D"},{"comment":"The sentence 'we consolidate the resultant 3 3D rotational attention maps' appears to contain a duplicated '3'; it should read 'three 3D'.","section":"III.C"},{"comment":"Reference [33] is malformed: the author list contains a garbled fragment ('and D. H., \"Lee,\"Y. Zhou, ...') that should be cleaned up.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's contribution is incremental but potentially useful if the empirical results survive scrutiny. The main risk is that the headline SOTA number is a selected single-run result with an unspecified evaluation split; this is fixable with additional experiments and clearer reporting, so I do not recommend rejection. In revision, the authors should provide code or detailed pseudocode, multiple seeds, and explicit split disclosure. I would also encourage the editor to check whether the claimed SOTA comparison is current, since the paper appears to omit several recent FER2013 results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an honest engineering paper with a useful controlled comparison, but the headline SOTA number is not robust as reported. The new block (TripSE) is a straightforward combination of Triplet Attention and SE in four placement variants. That is combinatorial novelty, but it is real; I haven't seen these exact configurations in the literature. The parameter overhead is tiny (~0.1M), and the controlled finetuning comparison on FER2013 under a fixed protocol shows ConvNeXt-S+TripSE1 beating both TA-only and SE-only versions of the same model. That controlled evidence is the best part of the paper and supports the claim that the combination can help.\n\nWhere it gets weak is the headline 78.27% claim. That number is the maximum over four variants and two reduction ratios (eight configurations), reported from a single run, with no seeds, no standard deviation, and no explicit statement of whether the accuracy is on the FER2013 test or validation split. Under the standard protocol this matters. Several neighboring configurations score below the plain ConvNeXt baseline (76.61, 76.71, 76.54, 76.66), so the selected 78.27 could easily be noise or test-set overfitting. The abstract's blanket claim that TripSE 'boosts' performance is also contradicted by the paper's own Table II: ResNet18 on CIFAR100 drops, DenseNet on AffectNet drops. So the SOTA number should be labeled provisional, and the general claim should be softened.\n\nThere are minor inconsistencies: Section V.C says ConvNeXt-T baseline is 77.20 but Table III lists 77.19; the implementation details are thin for AffectNet (10 epochs). These are fixable.\n\nBottom line: the controlled comparison is worth a serious referee, but the paper needs seeds, a clear test/validation statement, and code before the 78.27 claim can be taken at face value. It deserves peer review with major revision, not desk rejection.","headline":"An honest engineering paper whose controlled comparison supports the TripSE block, but the headline 78.27% FER2013 SOTA is a single selected run and needs seeds and a clear test/validation statement before it can be taken at face value.","tokens_in":12769,"tokens_out":1897,"would_cite":false,"duration_ms":18037,"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":"A new attention block combining Triplet Attention and Squeeze-and-Excitation reports 78.27% on FER2013.","keywords":["attention mechanism","Triplet Attention","Squeeze-and-Excitation","3D attention","facial expression recognition","FER2013","ConvNeXt","transfer learning"],"falsifier":"Re-run ConvNeXt-S plus TripSE4 ($r=1$) on FER2013 under the standard protocol: fix the configuration before evaluation, report the split, run multiple seeds, and compare the mean and variance against the 76.82% baseline. If the mean is below 78.27% or within noise of the plain ConvNeXt-S score of 77.19%, the state-of-the-art claim is not sustained.","tokens_in":11819,"feed_emoji":"😊","tokens_out":5327,"duration_ms":48209,"temperature":0.7,"pith_summary":"The paper claims that fusing two existing attention ideas, Triplet Attention (TA) and Squeeze-and-Excitation (SE), into a single block produces a form of 3D attention that improves CNN image classifiers, especially for facial expression recognition. The proposed TripSE block is inserted after each group of blocks in ResNet18, DenseNet121, and ConvNeXt, adding only about 0.1 million parameters. On the FER2013 benchmark the authors report that ConvNeXt-S with the TripSE4 variant and reduction ratio $r=1$ reaches 78.27% accuracy, which they call a new state of the art for this dataset. The same block also yields gains on AffectNet, CIFAR100, and ImageNet, and the paper argues this shows CNNs with cheap attention remain competitive in vision tasks.","feed_headline":"Attention block lifts FER2013 accuracy to 78.27%","feed_subtitle":"Combining triplet attention with squeeze-and-excitation sets a new state of the art on facial expression recognition.","key_machinery":"The central object is the TripSE attention block. It takes TA's three rotational branches, operating over the $C \\times W$, $C \\times H$, and $W \\times H$ faces of the tensor, each producing a 2D attention map through z-pooling, convolution, batch normalisation, and sigmoid, and fuses each map with the output of an SE module that computes global channel weights from the permuted tensor. Depending on the variant, the SE sits before, after, or parallel to the TA branch; TripSE3 and TripSE4 run SE in parallel with each branch, and TripSE4 adds a shift before the sigmoid and a final SE at the end, yielding a full 3D weighting tensor used to scale the branch input. The block's work is to let inter-dimensional attention and global channel importance modulate each other before the three branches are combined.","core_discovery":"On its own terms, the paper discovers that the gap between TA's cross-dimension attention and SE's global channel weighting can be closed by a multiplicative, and in TripSE4 affine, fusion of the two, producing a 3D attention tensor. In a controlled comparison on FER2013, ConvNeXt-T with TripSE1 reaches 77.86%, above the reported reproducible baseline of 76.82% from a residual masking network ensemble, and ConvNeXt-S with TripSE4 ($r=1$) reaches 78.27%. The authors attribute the success of TripSE4 to its parallel SE branches with a shift-and-scale transformation per rotational channel and a final SE at branch unification, which they say preserves information better than simple averaging. They also report that adding SE alone or TA alone to ConvNeXt lowers FER2013 accuracy relative to the plain backbone, whereas TripSE raises it.","pith_inferences":["The reported 78.27% comes from a configuration selected after seeing multiple variant and ratio results, so independent multi-seed reproduction with a pre-registered configuration is needed to know how much of the gap over baselines is real.","The same fusion idea may transfer to other small-sample vision tasks, such as medical image classification, where global channel statistics are informative and overfitting risk is high.","A natural testable extension is to vary the reduction ratio and branch placement systematically across more backbones, and to measure whether TripSE's benefit grows with the number of inserted blocks.","Because the paper does not state whether the FER2013 number is measured on the test or validation split, any comparison against published leaderboard results is provisional until the protocol is clarified."],"forward_implications":["If the FER2013 result holds, a simple plug-in attention block can lift existing CNN backbones above published facial-expression state of the art without an architectural redesign.","The controlled comparisons imply that TA alone and SE alone are not enough on this task; their fusion is what produces the reported gain.","The near-zero parameter overhead (28.6 to 28.7 million on ConvNeXt-T) means the block is cheap enough to add to deployed models.","The AffectNet improvement, from 62.87% to 64.16%, suggests the benefit is not specific to a single dataset.","The underperformance of TripSE2 and TripSE3 relative to TripSE1 and TripSE4 is presented as evidence that 3D scaling and a final SE at unification are load-bearing design choices."],"supporting_citations":[{"why":"Supplies the Triplet Attention module whose three rotational branches TripSE extends.","marker":"[12]"},{"why":"Supplies the Squeeze-and-Excitation channel attention block that TripSE fuses with TA.","marker":"[9]"},{"why":"Provides the ConvNeXt architecture and pretrained weights used for the headline FER2013 results.","marker":"[16]"},{"why":"Provides the 76.82% reproducible state-of-the-art baseline that the paper's FER2013 comparisons must beat.","marker":"[31]"},{"why":"Shows a prior ConvNeXt-plus-SE design for facial expression recognition that motivates combining SE with attention.","marker":"[17]"},{"why":"Prior study of attention mechanisms on CNNs for facial expression recognition that motivates using TA as a strong baseline.","marker":"[30]"}],"fun_headline_variants":["TripSE attention hits 78.27% on FER2013","New attention block sets FER2013 record: 78.27%","Triplet + squeeze-excitation achieves 78.27% FER","3D attention block outperforms on FER2013","Combined attention mechanism reaches 78.27% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline 78.27% result rests on the assumption that a single run, whose split is not stated and whose configuration was chosen after seeing all variant results, is a trustworthy estimate of generalization on FER2013.","fun_headline_variants_meta":{"raw":{"variants":["TripSE attention hits 78.27% on FER2013","New attention block sets FER2013 record: 78.27%","Triplet + squeeze-excitation achieves 78.27% FER","3D attention block outperforms on FER2013","Combined attention mechanism reaches 78.27% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1409,"prompt_tokens":924,"completion_tokens":485,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":394}},"tokens_in":540,"tokens_out":485,"duration_ms":3978,"temperature":1.0,"reasoning_tokens":394,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:52:07.216081+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run ConvNeXt-S plus TripSE4 ($r=1$) on FER2013 under the standard protocol: fix the configuration before evaluation, report the split, run multiple seeds, and compare the mean and variance against the 76.82% baseline. If the mean is below 78.27% or within noise of the plain ConvNeXt-S score of 77.19%, the state-of-the-art claim is not sustained.","supporting_citations":[{"cited_title":"Rotate to Attend: Convolutional Triplet Attention Module","cited_arxiv_id":"2010.03045","evidence_quote":"Supplies the Triplet Attention module whose three rotational branches TripSE extends."},{"cited_title":"A convnet for the 2020s,","cited_arxiv_id":null,"evidence_quote":"Provides the ConvNeXt architecture and pretrained weights used for the headline FER2013 results."},{"cited_title":"Facial expression recognition using residual masking network,","cited_arxiv_id":null,"evidence_quote":"Provides the 76.82% reproducible state-of-the-art baseline that the paper's FER2013 comparisons must beat."},{"cited_title":"Emonext: an adapted convnext for facial emotion recognition,","cited_arxiv_id":null,"evidence_quote":"Shows a prior ConvNeXt-plus-SE design for facial expression recognition that motivates combining SE with attention."},{"cited_title":"Best fit activation functions for attention mechanism: Comparison and enhancement,","cited_arxiv_id":null,"evidence_quote":"Prior study of attention mechanisms on CNNs for facial expression recognition that motivates using TA as a strong baseline."}],"review_version":1}