{"id":"7bdc1e71-5677-42d1-bbdd-72ec1f42f0ce","arxiv_id":"2505.11538","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-branch MLP that reads lower-triangular functional connectivity and low-pass filtered BOLD spectra matches or beats transformer baselines on ABIDE and HCP with far fewer parameters and FLOPs.","lead":"A pure MLP architecture called BrainNetMLP classifies functional brain networks from resting-state fMRI using one branch for connectivity correlations and one branch for frequency spectra. It reports accuracy competitive with or above larger transformer and graph models on ABIDE and HCP while using about 10x fewer FLOPs and 2x fewer parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on cross-paper baseline numbers whose protocol match is unverified; on HCP the quoted 2.1% gap is over a baseline the paper itself flags as requiring unavailable inputs.","rationale":"The reader's weakest assumption correctly identifies the comparability of Table 1 baselines as the load-bearing point. My stress-test confirms and sharpens this concern: the HCP headline margin is computed against STGCN, which the manuscript itself marks as not applicable to the datasets, and the ABIDE margin is within one standard deviation with no paired significance test. The architecture itself is internally coherent, the ablations support the dual-branch design, and the efficiency numbers are plausible for an MLP, so the concern is evidentiary rather than methodological. CONDITIONAL remains the right verdict because the claims are plausible but not yet verifiable from the manuscript; no change to the reader's verdict is needed.","tokens_in":8338,"tokens_out":6861,"duration_ms":73925,"concrete_test":"Download the released code and any supplementary file, then reproduce Table 1 by using the authors' exact atlas, ROI count, low-pass filter, and data splits for ABIDE and HCP, including any unrelated-subject or site stratification. Retrain GBT, BrainNetTF, and, if feasible, STGCN on those same splits instead of importing numbers from the original papers, and run a paired McNemar test between BrainNetMLP and each baseline on identical folds. If the ABIDE accuracy gap becomes statistically insignificant, or if the HCP margin relative to the valid baselines changes materially when STGCN is excluded, then the state-of-the-art claim is not supported by the current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, that BrainNetMLP achieves state-of-the-art accuracy, depends entirely on Table 1 comparisons with baseline results that appear to be imported from their original papers. Section 3.1 only says the experimental settings are in the supplementary material, and no supplementary file or code is included in the arXiv submission. This is not a cosmetic omission because both headline margins are fragile. On ABIDE, BrainNetMLP reaches 72.6±1.7% accuracy versus GBT's 71.5±2.1%, a 1.1-point gap within one standard deviation, while its AUCROC (78.4) is lower than both GBT (79.2) and BrainNetTF (79.1); no paired significance test is reported. On HCP, the quoted 2.1% improvement is measured against STGCN (77.7%), but the Table 1 footnote states that STGCN requires fixed adjacency matrices and community priors that are not provided for ABIDE and HCP. If STGCN's number came from a different experimental setup, it is not a valid comparator, and the valid comparison set is unverified. HCP also raises a known family-structure leakage risk: if siblings or twins are split across training and test folds, accuracy can be inflated, and the paper does not state how subjects were partitioned. Until the exact atlas, ROI count, preprocessing steps, fold assignments, and subject exclusion rules are disclosed and baselines are rerun under identical conditions, the 1-2% accuracy advantages are unverified cross-paper comparisons rather than established facts.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BrainNetMLP, a pure MLP-based architecture for functional brain network classification. It uses a dual-branch design: a Spatial Connectivity Mixer (SCMixer) operating on the lower-triangular Pearson correlation matrix and a Spectrum ROI Mixer (SRMixer) operating on the low-pass-filtered FFT amplitude spectrum of the BOLD time series, followed by feature concatenation and a linear predictor. An unstructured pruning scheme, Edge Degree Guided Pruning (EDGP), is introduced to remove parameters based on weight magnitude and node degree. The model is evaluated on ABIDE and HCP against CNN, GNN, and Transformer baselines, reporting state-of-the-art accuracy (72.6% on ABIDE, 79.8% on HCP) with about 10x fewer FLOPs and 2x fewer parameters than transformer competitors. Ablation studies indicate that both branches contribute to the final performance.","tokens_in":8630,"tokens_out":4682,"duration_ms":43811,"significance":"The proposed architecture is simple, computationally efficient, and the ablation study (Table 2) supports the utility of both the spatial and spectral branches. The efficiency measurements (Fig. 1 and Fig. 3) suggest a practically useful baseline for functional brain network classification, and the paper usefully questions whether increasing model complexity is necessary. However, the central empirical claim of state-of-the-art accuracy is not yet demonstrated: the comparison relies on baseline numbers that appear to be imported from their original papers without a verified matched protocol, one of the headline HCP improvements is against a baseline the paper itself flags as non-comparable, and no significance tests are reported. If the comparison issues are fixed, the paper would be a solid contribution; in its current form, the SOTA claim is not supported by the presented evidence.","major_comments":[{"comment":"The central SOTA claim depends on Table 1 baseline values that appear to be taken from the original papers, with the only protocol description being 'The experimental settings can be found in our supplementary material' (Section 3.1); however, no supplementary file or code is included in the arXiv submission. On ABIDE, the 1.1-point accuracy gain over GBT is within one standard deviation of both methods (72.6±1.7 vs 71.5±2.1), and BrainNetMLP's AUCROC (78.4) is below both GBT (79.2) and BrainNetTF (79.1). Unless the authors demonstrate identical data splits, atlas, preprocessing, and ROI definitions, and report paired significance tests, the accuracy advantage over GBT is not established.","section":"Section 3.1, Table 1"},{"comment":"The claimed 2.1% HCP accuracy improvement is calculated over STGCN (77.7), but the table footnote states that STGCN requires fixed adjacency matrices and community priors that are not provided for ABIDE and HCP. If the STGCN result was obtained under a different experimental setting, it cannot be considered the second-best model. Additionally, the paper does not state how HCP subjects were partitioned with respect to family structure; because HCP includes siblings and twins, family-related leakage across training and test folds can inflate accuracy. Please report the partition policy and, if relevant, evaluate with a family-aware split.","section":"Table 1 footnote, Section 3.1"},{"comment":"The low-pass cutoff k in Eq. (4) and the pruning balance coefficient λ in Eq. (8) are free hyperparameters with no reported values or sensitivity analysis in the main text. These choices affect the spectral features and the pruning criterion, respectively, and without their values the results are not reproducible. Please report the chosen values and a brief sensitivity analysis (or state that they are fixed by prior knowledge) so that the reported performance does not depend on undisclosed tuning.","section":"Section 2.2, Eq. (4); Section 2.3, Eq. (8)"}],"minor_comments":[{"comment":"There is a formatting error in the HCP row for BrainNetCNN: '72.5±10.477.0±11.6' should read '72.5±10.4  77.0±11.6'.","section":"Table 1"},{"comment":"The phrase 'state-of-the-art accuracy and efficiency' is too strong given that Table 1 shows BrainNetMLP's ABIDE AUCROC (78.4) is lower than GBT (79.2) and BrainNetTF (79.1); consider saying 'competitive or state-of-the-art accuracy'.","section":"Abstract"},{"comment":"The sentence 'only one third/a half time consumed for classifying an functional brain network' is ungrammatical; suggest 'one-third to one-half the time'.","section":"Section 3.1"},{"comment":"The statement 'The size of the circles corresponds to the size of the models' is ambiguous; please specify whether the area or the radius represents the parameter count, and label the axes clearly.","section":"Fig. 1 caption"},{"comment":"The notation 'u∈ R N(N+1)/2' is missing the superscript in the rendered text; it should be 'u ∈ R^{N(N+1)/2}'.","section":"Section 2.1, Eq. (1)"},{"comment":"The paper repeatedly defers details to 'supplementary material' (Sections 3.1 and 3.2), but no supplementary file is included in the submission; these references should be completed or the essential settings should be moved to the main text.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is not supported by the evidence as presented. The main concerns are the absence of the supplementary file, the unverified cross-paper protocol matching, and the use of a flagged non-comparable baseline on HCP. The architecture and ablations are otherwise sound, but the SOTA claim needs to be reworked with a matched-protocol comparison and significance testing. I would also suggest the authors verify the novelty claim of being the 'first pure MLP-based method' given the popularity of MLP-mixer style models in adjacent domains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clean pure-MLP baseline for functional brain network classification: it flattens the lower triangle of the Pearson correlation matrix and feeds it through an MLP-Mixer-style block, and in parallel runs FFT-based low-pass spectral features through another mixer. That combination is new for this task, and the architecture is simple enough that the efficiency numbers (10x fewer FLOPs than transformer baselines) are credible.\n\nWhat the paper does well: the ablations actually support the design. Removing SCMixer collapses ABIDE accuracy to 52%, and removing the spectral branch hurts both datasets, so the dual-branch idea earns its keep. The triangular flattening exploits symmetry without losing information, which is the right kind of simple trick. The edge-degree-guided pruning is a minor but reasonable addition; it shows a small gain over L1-only pruning on ABIDE.\n\nThe soft spot is the headline SOTA claim. Table 1 compares against baseline numbers that appear to be imported from their original papers, and the only stated protocol details are deferred to a supplementary file that is not in the arXiv submission. That matters because the margins are fragile. On ABIDE, the accuracy gain over GBT is 1.1 points with overlapping standard deviations, and the AUC is actually lower (78.4 vs 79.2). On HCP, the largest accuracy gain (2.1% over STGCN) is against a baseline the footnote itself says requires fixed adjacency matrices and community priors that are not available for these datasets, so it is not a valid comparator. No significance tests are reported, and the HCP subject split does not address family structure, which is a known leakage risk in that dataset. These are not cosmetic omissions; they affect whether the central claim is true.\n\nThe paper is honest enough to flag the STGCN caveat, but then still uses that number in the comparison, which is a bit inconsistent. The architecture and ablations are solid; the empirical support for the SOTA claim is not yet.\n\nThis is a paper for people who want a cheap baseline and for reviewers who want a protocol-matched comparison. I would send it to peer review, but with a request for code, hyperparameters, exact atlas and ROI count, fold assignment, and subject exclusion details, and a rerun of the key baselines under identical conditions. If the 1-2% margins survive that, it is a genuinely useful reference point. If not, it is still a fine baseline paper, just with more modest claims.","headline":"A clean, efficient pure-MLP baseline whose architecture and ablations hold up, but whose SOTA accuracy claim rests on protocol-unmatched baselines and should be treated as unverified until the supplementary details and code appear.","tokens_in":9162,"tokens_out":2146,"would_cite":true,"duration_ms":20893,"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":"BrainNetMLP, a pure MLP with parallel spatial-connectivity and spectral branches, claims state-of-the-art accuracy on ABIDE and HCP brain-network classification while using roughly ten times fewer FLOPs and about half the parameters of…","keywords":["functional brain network classification","resting-state fMRI","functional connectivity","multi-layer perceptron","spectral feature learning","low-pass filtering","network pruning","ABIDE and HCP datasets"],"falsifier":"Run BrainNetMLP and every baseline under one shared protocol with identical folds, atlas, preprocessing, and hyperparameter budgets, then check whether the reported 1-2% accuracy gaps and the roughly tenfold FLOPs reduction persist; a simpler check is to ablate the spectral branch while matching parameter counts to see whether the accuracy gap closes.","tokens_in":1855,"feed_emoji":"🧠","tokens_out":2677,"duration_ms":56577,"temperature":0.7,"pith_summary":"The paper asks whether increasingly complex deep learning models are actually necessary for functional brain network classification. It answers no, proposing BrainNetMLP, a pure multi-layer perceptron architecture with a dual-branch design: one branch mixes the lower-triangular functional connectivity matrix, exploiting its symmetry, and the other mixes low-pass filtered spectral features of the BOLD time series. On the ABIDE autism and HCP datasets, the model reports accuracy gains of about 1.1% and 2.1% over the second-best baselines, alongside a roughly tenfold reduction in FLOPs and a twofold reduction in parameters compared with transformer models. The paper also introduces Edge Degree Guided Pruning, a topology-aware pruning method that further trims parameters without sacrificing accuracy. If these results hold under a fully matched experimental protocol, they suggest that model simplicity, rather than architectural complexity, may be the more reliable path for this application.","feed_headline":"Plain MLP beats transformer brain-network models by 1-2%","feed_subtitle":"BrainNetMLP needs a tenth of the FLOPs and still sets the top accuracy on ABIDE and HCP.","key_machinery":"The two load-bearing components are the Spatial Connectivity Mixer (SCMixer) and the Spectrum ROI Mixer (SRMixer). SCMixer extracts the strictly lower-triangular part of the symmetric functional connectivity matrix, flattens it into a vector of length N(N+1)/2, adds a positional embedding, and passes it through an MLP mixer that globally reweights all ROI-pair correlations in one shot. SRMixer performs a fast Fourier transform on the ROI time series, applies a low-pass filter that keeps only the first k frequency components, takes the amplitude spectrum, and mixes across ROIs with another MLP, followed by mean pooling over frequencies. The two branches are fused by concatenation and a small nonlinear predictor. The Edge Degree Guided Pruning (EDGP) technique ranks each SCMixer weight by the L1 norm of the weight plus λ times ED(e_i), where ED(e_i) is the sum of node degrees of the two ROIs connected by that edge, thereby biasing pruning toward connections deemed less important by brain-network topology.","core_discovery":"The central claim is that a pure MLP can outperform CNN, GNN, and Transformer models on functional brain network classification, provided it is designed around the structure of the data. BrainNetMLP treats the symmetric Pearson correlation matrix as a flattened lower-triangular vector, halving the input dimension without losing information, and processes it with a global MLP mixer. In parallel, it applies a fast Fourier transform to the ROI time series, keeps only the low-frequency components, and mixes those spectral ROI features with another MLP. The two feature sets are concatenated and passed to a simple predictor. This design jointly captures spatial connectivity and temporal dynamics while remaining computationally light, and the reported experiments show that it reaches the best accuracy and AUC on both ABIDE and HCP. The paper further claims that pruning parameters according to an edge-degree-weighted importance score can even improve accuracy at moderate pruning ratios, indicating that some learned connections in the ABIDE data are spurious.","pith_inferences":["A likely editorial inference: the magnitude of the claimed advantage (1-2% accuracy, 10x FLOPs) could shrink if every baseline were retrained under the exact same folds, atlas, and preprocessing; the paper defers those settings to supplementary material, so a fully controlled re-benchmark is the natural next check.","The spectral branch is a modular idea: it could be grafted onto existing GNN or transformer models, and the paper's ablation suggests that such a hybrid would likely outperform any single-branch design.","The observation that edge-degree pruning helps most on ABIDE, where spurious correlations are expected, suggests that topology-aware pruning might serve as a debugging tool for noisy connectomes beyond pure efficiency.","One could test whether the MLP's advantage is driven by the lower-triangular symmetry reduction or by the spectral branch alone; the paper's ablation indicates both contribute but does not isolate which matters more at matched parameter counts."],"forward_implications":["If BrainNetMLP's reported accuracy holds, then MLP-only models should be treated as strong baselines in future functional brain network classification studies, rather than being overshadowed by transformer or graph models.","The symmetry of the Pearson correlation matrix is a cheap and generally applicable structural prior: any model that consumes a full N by N connectivity matrix could halve its input dimension by using only the triangle.","Low-pass filtered spectral features appear to capture discriminative temporal dynamics that pure spatial connectivity models miss, and do so more robustly than raw time series.","The edge-degree-guided pruning results suggest that brain-network topology can guide parameter sparsity, meaning efficiency gains need not come at the cost of accuracy.","The reported scalability from 22 ROIs (HCP) to 200 ROIs (ABIDE) indicates that MLP-based architectures may remain practical for higher-resolution atlases, where transformer complexity grows rapidly."],"supporting_citations":[{"why":"BrainNetCNN is the principal CNN baseline whose edge-convolution design the MLP is compared against on both datasets.","marker":"[13]"},{"why":"BrainNetTF is the transformer baseline whose FLOPs, parameter count, and accuracy serve as the main efficiency and performance reference.","marker":"[12]"},{"why":"GBT is a recent geometric transformer baseline compared on ABIDE and HCP, and its complexity scaling is contrasted with the proposed model.","marker":"[16]"},{"why":"STGCN supplies the spatiotemporal graph convolution baseline and motivates the use of temporal dynamics, though its HCP results are reported despite the missing fixed-adjacency prerequisites.","marker":"[4]"},{"why":"ComBrainTF is the community-aware transformer baseline on ABIDE, included in the comparison table with a footnote about missing community priors.","marker":"[1]"},{"why":"FBNETGEN is the GNN baseline that generates functional brain networks and is compared on both datasets.","marker":"[11]"},{"why":"MLP-Mixer supplies the all-MLP architecture idea that motivates replacing separate token and channel mixing with global MLP mixing.","marker":"[19]"},{"why":"This pruning work supplies the prune-after-training scheme that the proposed EDGP method follows.","marker":"[15]"}],"fun_headline_variants":["Pure MLP outperforms GNN and Transformer on brain networks","BrainNetMLP: simple beats complex for brain network classification","Efficient MLP baseline sets new benchmark for brain nets","MLP with dual-branch design wins brain network challenge","Complex models lose to plain MLP on ABIDE and HCP"],"cache_read_input_tokens":11264,"weakest_assumption_plain":"The central claim stands on the assumption that all baselines were evaluated under exactly the same data splits, atlas, preprocessing, and configuration as the proposed model; the paper defers those details to supplementary material and even lists STGCN and ComBrainTF results although those methods require fixed adjacency matrices and community priors that the datasets do not provide.","fun_headline_variants_meta":{"raw":{"variants":["Pure MLP outperforms GNN and Transformer on brain networks","BrainNetMLP: simple beats complex for brain network classification","Efficient MLP baseline sets new benchmark for brain nets","MLP with dual-branch design wins brain network challenge","Complex models lose to plain MLP on ABIDE and HCP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1535,"prompt_tokens":994,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":610,"tokens_out":541,"duration_ms":5205,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:36:02.369836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BrainNetMLP and every baseline under one shared protocol with identical folds, atlas, preprocessing, and hyperparameter budgets, then check whether the reported 1-2% accuracy gaps and the roughly tenfold FLOPs reduction persist; a simpler check is to ablate the spectral branch while matching parameter counts to see whether the accuracy gap closes.","supporting_citations":[{"cited_title":"NeuroImage146, 1038–1049 (2017) 10 Hou et al","cited_arxiv_id":null,"evidence_quote":"BrainNetCNN is the principal CNN baseline whose edge-convolution design the MLP is compared against on both datasets."},{"cited_title":"NIPS35, 25586–25599 (2022)","cited_arxiv_id":null,"evidence_quote":"BrainNetTF is the transformer baseline whose FLOPs, parameter count, and accuracy serve as the main efficiency and performance reference."},{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"GBT is a recent geometric transformer baseline compared on ABIDE and HCP, and its complexity scaling is contrasted with the proposed model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"STGCN supplies the spatiotemporal graph convolution baseline and motivates the use of temporal dynamics, though its HCP results are reported despite the missing fixed-adjacency prerequisites."},{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"ComBrainTF is the community-aware transformer baseline on ABIDE, included in the comparison table with a footnote about missing community priors."},{"cited_title":"In: MIDL","cited_arxiv_id":null,"evidence_quote":"FBNETGEN is the GNN baseline that generates functional brain networks and is compared on both datasets."},{"cited_title":"NIPS34, 24261–24272 (2021)","cited_arxiv_id":null,"evidence_quote":"MLP-Mixer supplies the all-MLP architecture idea that motivates replacing separate token and channel mixing with global MLP mixing."}],"review_version":1}