Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Explaining Vision GNNs: A Semantic and Visual Analysis of Graph-based Image Classification

T0 review · 5 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A Vision GNN's own graph shows its reasoning diverges from human perception in deep layers.

desk verdict Useful but flawed diagnostic: the local-to-global trend in ViG is real, but the 'semantic coherence' metric is circular and the analysis lacks statistics. read the letter →

arxiv 2504.19682 v1 pith:PMGHM2PH submitted 2025-04-28 cs.CV cs.LG

Nikolaos Chaidos , Angeliki Dimitriou , Nikolaos Spanos , Athanasios Voulodimos , Giorgos Stamou This is my paper · ORCID
classification cs.CVcs.LG
keywords VisionGNNgraphneuralnetworksimageclassificationexplainabilitymodularityembeddingsimilarityadversarialexamplessemanticconsistency
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 tries to establish that a Vision GNN's internal graph can be read as a self-documenting record of how an image classification decision forms, and that reading it reveals a process that diverges from human perception. On standard images, early layers connect nearby patches that look alike, while deeper layers link distant patches that the model considers embedding-similar even when they look nothing alike to a human. A late spike in embedding similarity coincides with a jump in classification confidence, suggesting that the model commits to a class through semantic convergence rather than visual coherence. On natural adversarial images, the object-background separation in the graph is weaker from the start and the final convergence spike is muted, matching the model's low accuracy. A sympathetic reader would take away that graph structure itself is a usable explanation surface for these models, even when it shows the model reasoning unlike a human.

What carries the argument

The central object is the per-layer dynamic graph $G^l = (V, E^l)$ built by connecting each of 196 image patches to its $K$ nearest neighbors in embedding space using cosine similarity. The analysis instruments this graph with five metrics—embedding similarity $S^l_{\text{emb}}$, spatial distance $D^l$, visual similarity $S^l_{\text{vis}}$, layer-wise ground-truth probability $p^l$, and object-based modularity $Q^l$ computed with binary object masks from GroundingDINO and SAM—plus heatmaps showing incoming edges of selected patches. The metrics are what carry the argument: they transform the dynamic graph update into a measurable story about locality, semantic coherence, and object separation across layers.

What would settle it

Recompute $Q^l$ on a few hundred ImageNet images using human-annotated ground-truth object masks instead of the automatic GroundingDINO/SAM masks, and also on a control set of random binary masks. If the modularity gap between ImageNet and ImageNet-a (0.236 versus 0.095) shrinks or disappears, the object-separation claim is an artifact of the mask generation; if random masks produce similar values, the metric is not measuring object structure.

Watch

Extended reading notes

Core claim

On the paper's own terms, the decision-making of Vision GNN image classifiers can be effectively explained by examining the graphs formed at each layer, and those graphs do not behave like human perception. Across layers there is a consistent progression: visual similarity of connected patches falls from 0.700 to 0.306, spatial distance grows from 3.5 to 8.9 grid steps, embedding similarity dips then spikes to 0.900 in the final layers, and this spike aligns with the largest increase in ground-truth class probability. On ImageNet-a adversarial images, graph modularity starts at 0.095 versus 0.236 on ImageNet and stays lower throughout, and the convergence spike is reduced, indicating weaker object/background separation and less coherent class-specific representations. The explanations are therefore effective as diagnostics of the model's own reasoning, not as evidence of human-aligned reasoning.

Load-bearing premise

The binary object masks produced by GroundingDINO and SAM are assumed to segment the ground-truth object correctly for every image, and the object-based modularity scores, including the conclusion that adversarial images have weaker object separation, would shift if those masks are wrong.

Editorial extensions

If this is right

  • Graph edges of a ViG become a faithful explanation artifact: one can trace a patch's receptive field and see the model's shift from local visual features to global semantic ones.
  • The final-layer embedding-similarity spike is a marker of decision consolidation, so monitoring it could serve as a confidence or failure detector without extra supervision.
  • Low graph modularity on adversarial inputs could act as an early signal of likely misclassification, since it correlates with near-zero accuracy on ImageNet-a.
  • Explanations should not be expected to match human intuition in deep layers; methods that align machine explanations to human perception may fight against the model's actual behavior.
  • The dynamic graph construction in ViG is inherently more transparent than fixed receptive fields, because each edge documents a learned relevance relation.

Reading between the lines

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

  • Editorial inference: The same five metrics could serve as a layer-wise diagnostic for other graph-based vision models, but the paper only studies ViG, so transfer is not established.
  • Editorial inference: The final embedding-similarity spike may partly be an artifact of k-NN graph construction in deep embedding space, where mutually nearest neighbors become mechanically similar; the paper does not separate this structural effect from genuine semantic convergence.
  • Editorial inference: If the central claim is right, a regularizer that forces deep-layer edges to be visually similar to human perception should reduce accuracy; that is a direct testable consequence not run in the paper.
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

5 major / 6 minor

Summary. The paper presents a white-box explainability analysis of the Vision GNN (ViG) image classifier. It defines five metrics—embedding similarity, spatial distance, visual similarity, layer-wise class probability, and object-based modularity—and reports their evolution across layer pairs on a 10,000-image ImageNet subset and on 7,500 ImageNet-a adversarial images. The main descriptive finding is that as layers deepen, connected patches become more spatially distant and less visually similar, while embedding similarity rises sharply in the final layers, coinciding with increasing classification confidence. The authors also provide heatmap visualizations and claim that their analysis reveals ViG's decision-making can be explained and that its later-layer reasoning diverges from human perception.

Significance. If validated, this would be a useful contribution to interpretability for graph-based vision models, introducing a suite of quantitative metrics and heatmap visualizations that could be reused by the community. The paper is honest about its white-box assumption and provides code. The non-circular metrics (D and S_vis) give a clean, reproducible description of the model's increasing receptive field, and the comparison between in-domain and adversarial images is a relevant robustness question. However, the paper's central semantic-coherence claim rests on an embedding-similarity metric that is in part a tautology of the KNN graph construction, and the absence of statistical validation and mask-quality checks limits the strength of the additional conclusions.

major comments (5)
  1. [Sec. 3.2, Eq. (2)] The metric S^l_emb is circular with respect to the graph construction. ViG builds E^l by connecting each node to its K nearest neighbors in the same embedding space using cosine similarity (Sec. 3.2), so Eq. (2) simply averages the top-K cosine similarities of each node. The late-layer increase in S_emb (Table 1: 0.842 to 0.900; Fig. 2) is therefore expected as features cluster and does not, by itself, show that the model 'links semantically related regions' (Sec. 4.2) or that confident predictions diverge from human visual intuition (Sec. 5.2). To support the semantic interpretation, the authors should compare S_emb to a baseline (e.g., random edges or non-neighbor pairs) or validate the selected edges with an external semantic signal.
  2. [Table 1, Sec. 5.1] All quantitative conclusions are based on single point estimates. There are no error bars, standard deviations, or significance tests, yet the text makes comparative claims such as a 'sudden increase' in S_emb and 'significantly lower' modularity for ImageNet-a (Q 0.095 vs 0.236). For a dataset of 10,000 and 7,500 images, the authors should report per-image variation and appropriate significance tests (paired or bootstrap) for the layer-to-layer and dataset-to-dataset differences.
  3. [Sec. 4.1] The ImageNet subset is described only as '10,000 validation images' without details of sampling, seed, or class balance, making selection bias and reproducibility impossible to assess. Please specify the exact subset construction or use the full validation set. Also correct the statement in Sec. 5.1 that results are across 'all three datasets', since only ImageNet and ImageNet-a are used.
  4. [Sec. 4.2, Eq. (6)] The modularity formula is not the standard directed-graph modularity and is not derived. For a directed graph, modularity is usually defined with k_in and k_out in the null model (e.g., Leicht and Newman), and the form used here, with (k_in^c k_out^c)/(2|E|)^2, is not justified. In addition, the object/background masks from GroundingDINO and SAM are taken as ground truth with no quality validation; for multi-object images the 'main object' is ambiguous, and the adversarial versus standard difference in Q could be an artifact of segmentation failures. The authors should either adopt the correct modularity definition and validate masks (e.g., IoU on a human-annotated subset) or show that their qualitative conclusions are robust to these choices.
  5. [Sec. 5.1 and abstract] The paper analyzes only ViG-Small, yet the title, abstract, and conclusion refer to 'Vision GNNs' and 'GNN-based vision models' in general. To substantiate the plural claim, the authors should add at least one additional architecture (e.g., MobileViG or GreedyViG) or explicitly restrict the scope and revise the abstract and conclusion accordingly.
minor comments (6)
  1. [Abstract] The phrase 'Explanations from standard and adversarial settings are also compared to assess whether they reflect the classifiers' robustness' is unclear; the comparison is between metric values, not between 'explanations' as objects.
  2. [Table 1] The star and dagger markers for ImageNet S_emb are inconsistent with the stated rule; the third-highest value (0.900, layers 15-16) is not marked, while 0.845 in layers 13-14 is marked as second highest. Please correct the markers.
  3. [Eq. (6)] The symbol L_c is used before being defined; please define it as the number of intra-community edges.
  4. [References] References [6] and [7] are duplicates of the same paper (Hendrycks et al.); remove one.
  5. [Figures] Fig. 2 and Fig. 3 are referenced in the text but appear to be missing from the arXiv preprint; ensure the final submission includes all figures.
  6. [Sec. 4.1] If the ImageNet subset is made available, please provide a download link or the exact selection procedure.

Circularity Check

1 steps flagged · score 5.0 of 10

Embedding-similarity metric is self-definitional: edges are selected by KNN on the same cosine similarity that Eq. (2) then averages.

  1. self definitional [Sections 3.2 and 4.2, Eq. (2)]
    "At each layer l, a graph Gl = (V, El) is constructed by connecting each node to its K nearest neighbors in the embedding space, based on cosine similarity between node features. ... S_l_emb = 1/|E_l| sum_{(i,j) in E_l} (x_l_i · x_l_j)/(||x_l_i|| ||x_l_j||) ... By quantifying the similarity of learned representations between connected patches, we assess if the model links semantically related regions. Higher values indicate stronger semantic coherence in the graph structure."

    The edge set E_l is itself selected as the K nearest neighbors of each node under cosine similarity of the same features x_l that Eq. (2) averages. Thus S_l_emb is simply the mean of the top-K cosine similarities that defined the graph: it is a self-similarity statistic of the KNN construction, not an independent measurement of whether the model 'links semantically related regions.' The late-layer spike in S_emb is therefore partly a mathematical consequence of feature clustering under the same metric, and cannot by itself support the conclusion that the model converges to 'class-specific representations' or that its confident decisions diverge from human visual intuition.

full rationale

The one concrete circular step is the embedding-similarity metric S_emb (Eq. 2). Since ViG's dynamic graph construction (Sec. 3.2) selects edges by K-nearest-neighbor on cosine similarity of the same node features, S_emb is by construction the average of the selection scores, so high values and the final-layer spike are partly tautological rather than evidence of semantic organization. This weakens the semantic side of the central claim, but the paper's other metrics—pixel-space visual similarity S_vis, grid-space Manhattan distance D, and object-based modularity Q—are not defined in terms of the graph-construction criterion and independently support the observed progression from local, visually coherent connections to long-range, visually dissimilar ones, as well as the degraded object separation on ImageNet-a. There is no load-bearing self-citation: the cited ViG work is external and maechine-implemented, and no uniqueness theorem is imported from the authors' own prior work. The unvalidated segmentation-mask assumption behind Q is a correctness risk, not a circularity. Overall, the central claim retains independent content, but one of its key quantitative pillars reduces to the construction itself, so a partial circularity score of 5 is appropriate.

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

The analysis rests on two main domain assumptions: the quality of automatic object masks, and the validity of probing intermediate layers with the final classification head. No free parameters are fitted in this paper; the only design choices are the 10k subset and the choice of external mask generators.

assumptions (3)
  • domain assumption GroundingDINO and SAM produce binary object masks that accurately separate the ground-truth object from background for all images.
    The modularity metric Q (Eq. 6) relies on these masks to define object and background communities; no validation of mask quality is reported. Location: Section 4.2.
  • domain assumption Applying the final classification head to intermediate layer features yields meaningful layer-wise class probabilities p_l.
    The paper uses a frozen linear head on earlier representations (Eq. 5), assuming it tracks the model's decision evolution. Location: Section 4.2.
  • domain assumption ViG-Small's dynamic graph construction uses K-nearest-neighbor with cosine similarity as described in the original ViG paper, and the pretrained weights are used without modification.
    The analysis assumes the model behaves as documented in the cited reference [4]. Location: Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explaining Vision GNNs: A Semantic and Visual Analysis of Graph-based Image Classification." pith.science (2026). https://pith.science/paper/PMGHM2PH

@misc{pith2026250419682,
  author       = {Pith},
  title        = {Pith review of: Explaining Vision GNNs: A Semantic and Visual Analysis of Graph-based Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PMGHM2PH}},
  note         = {Machine review of arXiv:2504.19682}
}
read the original abstract

Graph Neural Networks (GNNs) have emerged as an efficient alternative to convolutional approaches for vision tasks such as image classification, leveraging patch-based representations instead of raw pixels. These methods construct graphs where image patches serve as nodes, and edges are established based on patch similarity or classification relevance. Despite their efficiency, the explainability of GNN-based vision models remains underexplored, even though graphs are naturally interpretable. In this work, we analyze the semantic consistency of the graphs formed at different layers of GNN-based image classifiers, focusing on how well they preserve object structures and meaningful relationships. A comprehensive analysis is presented by quantifying the extent to which inter-layer graph connections reflect semantic similarity and spatial coherence. Explanations from standard and adversarial settings are also compared to assess whether they reflect the classifiers' robustness. Additionally, we visualize the flow of information across layers through heatmap-based visualization techniques, thereby highlighting the models' explainability. Our findings demonstrate that the decision-making processes of these models can be effectively explained, while also revealing that their reasoning does not necessarily align with human perception, especially in deeper layers.

Figures

Figures reproduced from arXiv: 2504.19682 by the authors.

Figure 1
Figure 1. Visual depiction of ViG’s representation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Heatmap Visualization of intermediate graphs (layers 4 and 10), for two [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Heatmap Visualization of intermediate graphs (layers 1, 8 and 15), for [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Prediction via Shapley Value Regression

    cs.LG 2025-05 conditional novelty 6.0 of 10

    ViaSHAP trains one network whose prediction is the sum of its own learned Shapley values, giving tabular accuracy comparable to XGBoost and built-in, fast Shapley explanations.

Reference graph

Works this paper leans on

21 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: 2020 33rd SIBGRAPI Con- ference on Graphics, Patterns and Images (SIBGRAPI)

    Avelar, P.H., Tavares, A.R., da Silveira, T.L., Jung, C.R., Lamb, L.C.: Superpixel image classification with graph attention networks. In: 2020 33rd SIBGRAPI Con- ference on Graphics, Patterns and Images (SIBGRAPI). pp. 203–209. IEEE (2020)

    work page 2020

  2. [2]

    In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV)

    Chattopadhyay, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Grad- cam++: Improved visual explanations for deep convolutional networks. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 839– 847 (2018). https://doi.org/10.1109/WACV.2018.00097 12 N. Chaidos et al

    arXiv 2018

  3. [3]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Fey, M., Lenssen, J.E., Weichert, F., Müller, H.: Splinecnn: Fast geometric deep learning with continuous b-spline kernels. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 869–877 (2018)

    work page 2018

  4. [4]

    In: NeurIPS (2022)

    Han, K., Wang, Y., Guo, J., Tang, Y., Wu, E.: Vision gnn: An image is worth graph of nodes. In: NeurIPS (2022)

    2022

  5. [5]

    2023 IEEE/CVF International Conference on Com- puter Vision (ICCV) pp

    Han, Y., Wang, P., Kundu, S., Ding, Y., Wang, Z.: Vision hgnn: An image is more than a graph of nodes. 2023 IEEE/CVF International Conference on Com- puter Vision (ICCV) pp. 19821–19831 (2023), https://api.semanticscholar. org/CorpusID:267024787

    work page 2023

  6. [7]

    CVPR (2021)

    Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. CVPR (2021)

    work page 2021

  7. [8]

    Huang, Q., Yamada, M., Tian, Y., Singh, D., Yin, D., Chang, Y.: Graphlime: Local interpretable model explanations for graph neural networks (2020),https: //arxiv.org/abs/2001.06216

    work page Pith review arXiv 2020

  8. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023)

    2023

Show all 21 references
  1. [10]

    9266–9275 (2019), https://api.semanticscholar.org/CorpusID:201070021

    Li, G., Müller, M., Thabet, A.K., Ghanem, B.: Deepgcns: Can gcns go as deep as cnns? 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9266–9275 (2019), https://api.semanticscholar.org/CorpusID:201070021

    2019

  2. [11]

    In: European Conference on Computer Vision

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: European Conference on Computer Vision. pp. 38–

  3. [12]

    Luo, D., Cheng, W., Xu, D., Yu, W., Zong, B., Chen, H., Zhang, X.: Parameterized explainer for graph neural network (2020),https://arxiv.org/abs/2011.04573

    2020 arXiv

  4. [13]

    In: 2020 international joint conference on neural networks (IJCNN)

    Muhammad, M.B., Yeasin, M.: Eigen-cam: Class activation map using principal components. In: 2020 international joint conference on neural networks (IJCNN). pp. 1–7. IEEE (2020)

    2020

  5. [14]

    2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp

    Munir, M., Avery, W., Marculescu, R.: Mobilevig: Graph-based sparse attention for mobile vision applications. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp. 2211–2219 (2023),https:// api.semanticscholar.org/CorpusID:259317049

    2023

  6. [15]

    2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Munir, M., Avery, W., Rahman, M.M., Marculescu, R.: Greedyvig: Dynamic axial graph construction for efficient vision gnns. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6118–6127 (2024), https://api.semanticscholar.org/CorpusID:269757862

    2024

  7. [16]

    International Journal of Computer Vision115, 211 – 252 (2014),https://api.semanticscholar.org/CorpusID:2930547

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision115, 211 – 252 (2014),https://api.sema...

    2014

  8. [17]

    In: Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- cam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 618–626 (2017).https://doi.org/10...

    2017 doi

  9. [18]

    arXiv preprint arXiv:1312.6034 (2013) A Semantic and Visual Analysis of Graph-based Image Classification 13

    Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional net- works: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013) A Semantic and Visual Analysis of Graph-based Image Classification 13

    2013 arXiv

  10. [19]

    Vu, M.N., Thai, M.T.: Pgm-explainer: Probabilistic graphical model explanations for graph neural networks (2020),https://arxiv.org/abs/2010.05788

    2020 arXiv

  11. [20]

    Ying, R., Bourgeois, D., You, J., Zitnik, M., Leskovec, J.: Gnnexplainer: Generat- ing explanations for graph neural networks (2019),https://arxiv.org/abs/1903. 03894

    2019

  12. [21]

    In: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13

    Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 818–833. Springer (2014)

    2014

  13. [22]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921–2929 (2016)

    2016

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.