Pith. sign in

REVIEW 3 major objections 6 minor 29 references

BrainNetMLP: An Efficient and Effective Baseline for Functional Brain Network Classification

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 2505.11538 v2 pith:BFEG57MR submitted 2025-05-14 q-bio.NC cs.CV

Jiacheng Hou , Zhenjie Song , Ercan Engin Kuruoglu This is my paper · ORCID
classification q-bio.NCcs.CV
keywords functionalbrainnetworkclassificationresting-statefMRIconnectivitymulti-layerperceptronspectralfeaturelearninglow-passfilteringpruningABIDEandHCPdatasets
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

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.

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 (3)
  1. [Section 3.1, Table 1] 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.
  2. [Table 1 footnote, Section 3.1] 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.
  3. [Section 2.2, Eq. (4); Section 2.3, Eq. (8)] 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.
minor comments (6)
  1. [Table 1] 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'.
  2. [Abstract] 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'.
  3. [Section 3.1] 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'.
  4. [Fig. 1 caption] 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.
  5. [Section 2.1, Eq. (1)] 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}'.
  6. [Throughout] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the evaluation is empirical against external benchmarks, self-citations are background only, and the MLP pipeline is an architecture proposal rather than a fitted-input prediction.

full rationale

The paper contains no derivation chain whose output is equivalent to its own inputs. BrainNetMLP is defined by explicit equations (SCMixer in Eq. 1-2, SRMixer in Eq. 3-6, fusion in Eq. 7, pruning in Eq. 8), and its reported accuracy, AUCROC, specificity, and sensitivity are measured on held-out test portions of the external ABIDE and HCP datasets. The baseline numbers in Table 1 are imported from the cited original papers; this raises protocol-match and reproducibility concerns, but it is not circular because those numbers are independent experiments, not renamings, fits, or algebraic rearrangements of the present model's outputs. The footnote that STGCN and ComBrainTF require fixed adjacency matrices and community priors not provided in ABIDE/HCP is a comparability caveat, not a circularity step. The authors' self-citations (refs. 6, 7, 9, 25) appear in unrelated video-inpainting and graph-signal-processing contexts; no load-bearing argument in this paper is justified only by those self-citations, and no uniqueness theorem or fitted parameter is invoked to make the model's predictions match its inputs. The omitted supplementary details are a reproducibility weakness, not evidence of circularity. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical or biological entities are posited. The ledger consists of unreported hyperparameters and two domain assumptions about symmetry and spectral noise, plus an assumption about fair baseline comparison.

free parameters (3)
  • Low-pass cutoff k in SRMixer = not reported
    Eq. (4) keeps only the first k FFT components; k controls how much spectral content enters classification and is not specified.
  • Pruning balance coefficient lambda in EDGP = not reported
    Eq. (8) uses lambda to scale edge degree against the L1 weight norm; the value is not given.
  • Architecture and training hyperparameters = not reported
    Hidden dimensions, mixer depth, learning rate, epochs, batch size, and split strategy are deferred to a supplementary file; the SOTA table and efficiency numbers depend on them.
assumptions (3)
  • standard math Lower-triangular extraction of a Pearson correlation matrix discards no information because the matrix is symmetric.
    Section 2.1 relies on this to halve the SCMixer input dimension.
  • domain assumption The low-frequency part of the BOLD spectrum is more noise-robust than the raw time series, so a fixed cutoff can be applied.
    Section 2.2 introduces low-pass filtering without dataset-specific evidence; the cutoff k is not reported.
  • domain assumption The comparison protocol, including splits, atlas, preprocessing, and baseline configurations, matches the original reports for all methods in Table 1.
    Section 3.1 defers settings to a supplementary file, and the table lists methods the paper itself flags as requiring priors not provided in these datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BrainNetMLP: An Efficient and Effective Baseline for Functional Brain Network Classification." pith.science (2026). https://pith.science/paper/BFEG57MR

@misc{pith2026250511538,
  author       = {Pith},
  title        = {Pith review of: BrainNetMLP: An Efficient and Effective Baseline for Functional Brain Network Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BFEG57MR}},
  note         = {Machine review of arXiv:2505.11538}
}
read the original abstract

Recent studies have made great progress in functional brain network classification by modeling the brain as a network of Regions of Interest (ROIs) and leveraging their connections to understand brain functionality and diagnose mental disorders. Various deep learning architectures, including Convolutional Neural Networks, Graph Neural Networks, and the recent Transformer, have been developed. However, despite the increasing complexity of these models, the performance gain has not been as salient. This raises a question: Does increasing model complexity necessarily lead to higher classification accuracy? In this paper, we revisit the simplest deep learning architecture, the Multi-Layer Perceptron (MLP), and propose a pure MLP-based method, named BrainNetMLP, for functional brain network classification, which capitalizes on the advantages of MLP, including efficient computation and fewer parameters. Moreover, BrainNetMLP incorporates a dual-branch structure to jointly capture both spatial connectivity and spectral information, enabling precise spatiotemporal feature fusion. We evaluate our proposed BrainNetMLP on two public and popular brain network classification datasets, the Human Connectome Project (HCP) and the Autism Brain Imaging Data Exchange (ABIDE). Experimental results demonstrate pure MLP-based methods can achieve state-of-the-art performance, revealing the potential of MLP-based models as more efficient yet effective alternatives in functional brain network classification. The code will be available at https://github.com/JayceonHo/BrainNetMLP.

Figures

Figures reproduced from arXiv: 2505.11538 by the authors.

Figure 2
Figure 2. (a) the whole pipeline (b) the data preprocessing scheme (c) the structure [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Runtime comparison between our method and the other compared methods, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Pruning results in different ratios with and without ED. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

  1. [1]

    In: MICCAI

    Bannadabhavi, A., Lee, S., Deng, W., Ying, R., Li, X.: Community-aware transformer for autism prediction in fmri connectome. In: MICCAI. pp. 287–297. Springer (2023)

    work page 2023

  2. [2]

    Nature Reviews Neuro- science 12(1), 43–56 (2011)

    Deco, G., Jirsa, V.K., McIntosh, A.R.: Emerging concepts for the dynamical organization of resting-state activity in the brain. Nature Reviews Neuro- science 12(1), 43–56 (2011)

    work page 2011

  3. [3]

    Academic Press (2016)

    Fornito, A., Zalesky, A., Bullmore, E.: Fundamentals of Brain Network Analysis. Academic Press (2016)

    work page 2016

  4. [4]

    Gadgil, S., Zhao, Q., Pfefferbaum, A., Sullivan, E.V., Adeli, E., Pohl, K.M.: Spatio-temporalgraphconvolutionforresting-statefmrianalysis.In:MICCAI. pp. 528–538. Springer (2020)

    work page 2020

  5. [5]

    arXiv preprint arXiv:1606.08415 (2016)

    Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)

    arXiv 2016

  6. [6]

    IEEE Transactions on Image Processing33, 1095–1108 (2024)

    Hou, J., Ji, Z., Yang, J., Wang, C., Zheng, F.: Mcd-net: toward rgb-d video inpainting in real-world scenes. IEEE Transactions on Image Processing33, 1095–1108 (2024)

    work page 2024

  7. [7]

    IEEE Transactions on Circuits and Systems for Video Technology (2024)

    Hou, J., Ji, Z., Yang, J., Zheng, F.: Bidirectional error-aware fusion network for video inpainting. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    work page 2024

  8. [8]

    arXiv preprint arXiv:2106.04051 (2021)

    Hu,Y.,You,H.,Wang,Z.,Wang,Z.,Zhou,E.,Gao,Y.:Graph-mlp:Nodeclas- sification without message passing in graph. arXiv preprint arXiv:2106.04051 (2021)

    arXiv 2021

Show all 29 references
  1. [9]

    IEEE Transactions on Image Processing (2025)

    Ji, Z., Su, Y., Zhang, Y., Hou, J., Pang, Y., Han, J.: Raformer: Redundancy- aware transformer for video wire inpainting. IEEE Transactions on Image Processing (2025)

    2025

  2. [10]

    PloS one5(8), e12200 (2010)

    Joyce, K.E., Laurienti, P.J., Burdette, J.H., Hayasaka, S.: A new measure of centrality for brain networks. PloS one5(8), e12200 (2010)

    2010

  3. [11]

    In: MIDL

    Kan, X., Cui, H., Lukemire, J., Guo, Y., Yang, C.: Fbnetgen: Task-aware gnn-based fmri analysis via functional brain network generation. In: MIDL. pp. 618–637. PMLR (2022)

    2022

  4. [12]

    NIPS35, 25586–25599 (2022)

    Kan, X., Dai, W., Cui, H., Zhang, Z., Guo, Y., Yang, C.: Brain network transformer. NIPS35, 25586–25599 (2022)

    2022

  5. [13]

    NeuroImage146, 1038–1049 (2017) 10 Hou et al

    Kawahara, J., Brown, C.J., Miller, S.P., Booth, B.G., Chau, V., Grunau, R.E., Zwicker, J.G., Hamarneh, G.: Brainnetcnn: Convolutional neural networks for brain networks; towards predicting neurodevelopment. NeuroImage146, 1038–1049 (2017) 10 Hou et al

    2017

  6. [14]

    bioRxiv (2020)

    Li, X., Zhou, Y., Dvornek, N., Zhang, M., Gao, S., Zhuang, J., Scheinost, D., Staib, L., Ventola, P., Duncan, J.: Braingnn: Interpretable brain graph neural network for fmri analysis. bioRxiv (2020)

    2020

  7. [15]

    arXiv preprint arXiv:1810.05270 (2018)

    Liu, Z., Sun, M., Zhou, T., Huang, G., Darrell, T.: Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270 (2018)

    2018 arXiv

  8. [16]

    In: MICCAI

    Peng, Z., He, Z., Jiang, Y., Wang, P., Yuan, Y.: Gbt: Geometric-oriented brain transformer for autism diagnosis. In: MICCAI. pp. 142–152 (2024)

    2024

  9. [17]

    Statistics Surveys7, 1 (2013)

    Simpson, S.L., Bowman, F.D., Laurienti, P.J.: Analyzing complex functional brain networks: Fusing statistics and network science to understand the brain. Statistics Surveys7, 1 (2013)

    2013

  10. [18]

    NeuroImage54(2), 875–891 (2011)

    Smith, S.M., Miller, K.L., Salimi-Khorshidi, G., Webster, M., Beckmann, C.F., Nichols, T.E., Ramsey, J.D., Woolrich, M.W.: Network modelling methods for fmri. NeuroImage54(2), 875–891 (2011)

    2011

  11. [19]

    NIPS34, 24261–24272 (2021)

    Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An all-mlp architecture for vision. NIPS34, 24261–24272 (2021)

    2021

  12. [20]

    Computer Methods and Programs in Biomedicine 247, 108065 (2024)

    Wang, Y., Long, H., Bo, T., Zheng, J.: Residual graph transformer for autism spectrum disorder prediction. Computer Methods and Programs in Biomedicine 247, 108065 (2024)

    2024

  13. [21]

    Computers in Biology and Medicine163, 107184 (2023)

    Wang, Y., Long, H., Zhou, Q., Bo, T., Zheng, J.: Plsnet: Position-aware gcn-based autism spectrum disorder diagnosis via fc learning and rois sifting. Computers in Biology and Medicine163, 107184 (2023)

    2023

  14. [22]

    In: ICLR

    Wang, Z., Jiang, W., Zhu, Y.M., Yuan, L., Song, Y., Liu, W.: Dynamixer: A vision mlp architecture with dynamic mixing. In: ICLR. pp. 22691–22701. PMLR (2022)

    2022

  15. [23]

    Annals of the New York Academy of Sciences 1224(1), 126–146 (2011)

    Wig, G.S., Schlaggar, B.L., Petersen, S.E.: Concepts and principles in the analysis of brain networks. Annals of the New York Academy of Sciences 1224(1), 126–146 (2011)

    2011

  16. [24]

    In: MICCAI

    Yan, J., Chen, Y., Yang, S., Zhang, S., Jiang, M., Zhao, Z., Zhang, T., Zhao, Y., Becker, B., Liu, T., et al.: Multi-head gagnn: A multi-head guided attention graph neural network for modeling spatio-temporal patterns of holistic brain functional networks. In: MICCAI. pp. 564–...

    2021

  17. [25]

    arXiv preprint arXiv:2412.00462 (2024)

    Yan, Y., Hou, J., Song, Z., Kuruoglu, E.E.: Signal processing over time- varying graphs: A systematic review. arXiv preprint arXiv:2412.00462 (2024)

    2024

  18. [26]

    arXiv preprint arXiv:2406.17086 (2024)

    Yang, Y., Mao, Y., Liu, X., Liu, X.: Brainmae: A region-aware self-supervised learning framework for brain signals. arXiv preprint arXiv:2406.17086 (2024)

    2024 arXiv

  19. [27]

    IEEE transactions on circuits and systems for video technology30(12), 4467–4480 (2019)

    Yu, J., Li, J., Yu, Z., Huang, Q.: Multimodal transformer with multi-view visual representation for image captioning. IEEE transactions on circuits and systems for video technology30(12), 4467–4480 (2019)

    2019

  20. [28]

    NIPS34, 7267– 7280 (2021)

    Zhang, J., Chang, W.C., Yu, H.F., Dhillon, I.: Fast multi-resolution trans- former fine-tuning for extreme multi-label text classification. NIPS34, 7267– 7280 (2021)

    2021

  21. [29]

    In: International conference on modelling, simulation and applied mathematics

    Zhang, J., Luo, Y.: Degree centrality, betweenness centrality, and close- ness centrality in social network. In: International conference on modelling, simulation and applied mathematics. pp. 300–303. Atlantis press (2017)

    2017

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.