Pith. sign in

REVIEW 5 major objections 6 minor 29 references

GNN-ViTCap: GNN-Enhanced Multiple Instance Learning with Vision Transformers for Whole Slide Image Classification and Captioning

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read GNN-ViTCap claims one architecture can classify microscopic whole slide images and generate pathology captions, reporting F1 0.934, AUC 0.963, and BLEU-4 0.811 on two datasets.

desk verdict The paper combines known pieces into a new pipeline and reports strong numbers, but the graph construction that supposedly provides spatial context is ill-defined and untested. read the letter →

arxiv 2507.07006 v1 pith:WN2CYDZR submitted 2025-07-09 cs.CV cs.LG

classification cs.CVcs.LG
keywords wholeslideimageclassificationhistopathologycaptioningmultipleinstancelearningdeepembeddedclusteringgraphneuralnetworkvisiontransformerlargelanguagemodelmicroscopicWSI
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

This paper sets out to show that microscopic whole slide images, which lack the coordinates and contain redundant patches that scanner-based pipelines assume away, can still support both cancer classification and automatic diagnostic captions. The proposed GNN-ViTCap handles the missing positions by clustering patch embeddings, keeping one attention-selected representative per cluster, and then running a graph neural network over a similarity-built graph before pooling to a slide-level vector. That vector drives a binary benign/malignant classifier and, through a linear projection, conditions large language models that produce captions. On BreakHis the method reports an F1 score of 0.934 and an AUC of 0.963, and on PatchGastric a BLEU-4 of 0.811 with a METEOR of 0.569, numbers the paper presents as outperforming the compared state-of-the-art MIL and captioning methods. If these results hold, the same architecture offers a route to computer-aided pathology on routine microscope images rather than expensive whole-slide scanners.

What carries the argument

The carrying mechanism is the similarity-based graph over attention-selected cluster representatives. Because microscopic images have no absolute coordinates, ordinary spatial neighbor graphs cannot be built; instead, cosine similarity among the representative patch embeddings, binarized into edges with Gumbel-Softmax, defines which patches are treated as neighbors. Graph attention layers then propagate information along these edges, and global mean pooling produces the slide-level vector. This is what supplies the spatial and contextual information the paper credits for its gains, since the graph is the only component that relates patches to one another before classification or captioning.

What would settle it

A controlled ablation that keeps all components identical but replaces the GAT aggregation with simple mean pooling over the same representative embeddings would test whether the graph carries the reported gains; if F1, AUC, and BLEU-4 stay essentially unchanged, the graph is not the active ingredient. A second check is to randomize or shuffle the edge structure while keeping node features fixed, since genuine spatial information should degrade when the edges are destroyed.

Watch

Extended reading notes

Core claim

GNN-ViTCap claims that classification and captioning of microscopic WSIs can be jointly solved by replacing coordinate-based spatial modeling with a clustering-and-graph formulation. A ViT-B/16 encoder produces patch embeddings; deep embedded clustering groups redundant patches and a scaled dot-product attention score selects the most representative embedding per cluster. Those representatives become nodes of a graph whose edges are chosen by cosine similarity through a Gumbel-Softmax operation, after which three graph attention layers update node features and global mean pooling yields the slide representation. The same representation is passed to an MLP for classification and, after a linear projection, is used as a visual prefix to fine-tune four large language models. The paper reports the best variant, ViT with DEC, GNN-MIL, and BiomedGPT, reaches an F1 of 0.934 and AUC of 0.963 on BreakHis and BLEU-4 of 0.811 with METEOR of 0.569 on PatchGastric, and that these figures exceed the compared baselines including ABMIL, DSMIL, TransMIL, DTFD-MIL, PatchCap, PathM3, and SGMT.

Load-bearing premise

The load-bearing premise is that cosine similarity between patch embeddings in feature space stands in for the spatial and contextual tissue relationships a pathologist would actually use; if those similarity edges do not correspond to meaningful tissue topology, the graph reduces to a clustering artifact and the paper's explanation for its reported gains loses its footing.

Editorial extensions

If this is right

  • Microscopy-based diagnosis can exploit inter-patch context even when slide coordinates are unknown, because the graph is constructed from feature similarity rather than positions.
  • A single slide-level representation can serve both a diagnostic classifier and a captioning language model, so classification and report generation no longer need separate pipelines.
  • Removing redundant patches by clustering before graph construction makes graph aggregation and LLM fine-tuning feasible on gigapixel-scale data.
  • Replacing recurrent decoders and traditional transformers with large language models improves captioning on histopathology text, with domain-specific biomedical models giving the largest reported gains.
  • The reported margins over scanner-oriented MIL baselines indicate those baselines are not well matched to subjective microscope captures with unknown patch positions.

Reading between the lines

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

  • Because the graph edges are learned from feature similarity rather than coordinates, the same pipeline could transfer to other settings with missing positional metadata, such as legacy slide archives or intraoperative microscopy; the paper only tests two datasets.
  • A controlled comparison that fixes the language model and toggles the GNN would separate the graph's contribution from the language model's contribution to the captioning scores, an ablation the paper does not report.
  • The paper's stated sensitivity to the cluster count K suggests that adaptive clustering, rather than a fixed K, might preserve rare diagnostic structures that a fixed partition discards.
  • On scanner WSIs where true coordinates exist, comparing a coordinate-based graph with this similarity-based graph would show whether similarity edges recover real tissue topology or only semantic patch similarity.
Share X Bluesky LinkedIn Reddit HN

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 proposes GNN-ViTCap, a pipeline that combines a ViT/ResNet feature extractor, deep embedded clustering with attention-based representative selection, a graph neural network (GNN-MIL) over representative patch embeddings, and fine-tuned large language models, for classification and captioning of microscopic whole-slide images. The authors report F1=0.934 and AUC=0.963 on BreakHis classification and BLEU-4=0.811, METEOR=0.567 on PatchGastric captioning, claiming state-of-the-art performance. The central novelty is the GNN-MIL module, which is intended to recover spatial/contextual structure from microscopy images that lack absolute coordinates.

Significance. If the method and results were reproducible, the paper would be a useful contribution to computational pathology, since it directly addresses redundancy in microscopy patches and combines MIL with LLM-based captioning. The idea of using deep embedded clustering to reduce bag size and a graph over representative embeddings is interesting, and the evaluation is reported on standard metrics. However, the current manuscript does not support the central claim: the graph-construction step is not well defined, the graph is not based on spatial information despite the paper's explanation, and no ablation isolates the contribution of GNN-MIL from the clustering, attention selection, or choice of LLM. The paper also omits the train/test split protocol. These issues need to be resolved before the empirical claims can be accepted.

major comments (5)
  1. [III-E1, Eq. (16)] The edge-selection rule in Eq. (16) is ill-defined and does not implement the k-nearest-neighbor graph promised in the abstract. The condition uses N(i), the neighbor set of node i, which is exactly the set of edges E being constructed, making the definition circular. For a fixed column j, the maximum over k of σ_gsf(S_k,j) does not depend on the row i except through N(i), so the rule selects at most one incoming edge per column rather than connecting each node to its nearest neighbors. This is a load-bearing component of GNN-MIL, and the paper must either give a precise, executable construction (e.g., a row-wise top-k rule) or state that the reported results correspond to some different procedure.
  2. [III-E, V-C] The graph is built solely from cosine similarity between representative feature embeddings (Eq. (15)), not from any spatial or positional information. Yet Q1 and Q2 attribute the performance gains to 'spatial relationships' and 'patch positional encoding.' Since microscopic WSIs lack absolute coordinates, the graph can at best capture feature-space proximity. The paper provides no ablation that separates the GNN's effect from a non-spatial pooling baseline or a random graph, so the claimed spatial-context mechanism is unsupported. A controlled experiment (e.g., same features with global mean pooling over representative embeddings, or a graph with random edges) is necessary to support the explanation.
  3. [III-D2, Eq. (13)] The selection of the representative embedding via arg max in Eq. (13) is non-differentiable. The paper does not explain how gradients flow through this discrete selection to the feature extractor or clustering module, despite the total losses in Eqs. (24) and (26) being written as differentiable objectives. If the selection is detached, the end-to-end training described in Section III-G is inaccurate; if it is not detached, a straight-through or Gumbel-softmax estimator should be specified. This affects the reproducibility of the training procedure.
  4. [IV-A, IV-B] The paper does not specify the train/validation/test split for either dataset. This is critical for BreakHis, where the 7,909 images come from only 82 patients: if images from the same patient are not grouped into the same split, the reported F1 and AUC may be inflated by patient-level leakage. The authors should report the patient-wise split, the number of patients in each fold, and confirm that all baselines use identical splits.
  5. [V-B, Table II] The captioning comparison is confounded by model choice: GNN-ViTCap uses ViT-B/16 plus BiomedGPT/LLaMAV2/ClinicalT5, while PatchCap, PathM3, and SGMT use different visual encoders and decoders. The 26% BLEU-4 improvement is attributed to GNN-MIL, but no ablation replaces GNN-MIL with a simpler aggregation (e.g., mean pooling or ABMIL-style attention) while keeping the same ViT and LLM. Without such a controlled study, the margin cannot be credited to the proposed graph module.
minor comments (6)
  1. [Abstract / V-B] The abstract reports METEOR = 0.569, while Table II reports 0.567 for the same BiomedGPT configuration; please make these consistent.
  2. [III-G, Eq. (24)] In Eq. (24), the classification loss is written as LBCL; it should be LBCE to match Eq. (23).
  3. [III-D, Eq. (14)] Eq. (14) lists r_1 twice; it should be [r_1; r_2; ...; r_K].
  4. [III-D2, Eq. (10)] The symbol P in Eq. (10) is not defined; if it denotes summation over the vector dimensions, please write it explicitly.
  5. [V-B] The text contains a typo, 'BLUE@4' instead of 'BLEU@4'.
  6. [III-A, Eq. (1)] Eq. (1) contains 'iif' and an extra parenthesis; please correct the typo.

Circularity Check

1 steps flagged · score 2.0 of 10

One definitional circularity in Eq. (16) (edges defined via their own neighbor set); headline classification/captioning metrics are otherwise independent held-out results.

  1. self definitional [Section III-E.1, Eq. (16)]
    "E (s) i,j = 1, if S (s) i,j = max_{k∈N (i)} σgsf(S (s) k,j ) 0, otherwise (16) where E (s) i,j ∈ {0, 1} indicates the presence of an edge between node i and node j and N (i) denotes the set of neighbors for node i."

    The edge matrix E is the object being constructed, but the condition uses N(i), which is the set of neighbors of node i — i.e., the set {j : E_i,j = 1} that E itself defines. Thus E appears on both sides of the definition: E_i,j is defined by taking a maximum over k ∈ N(i), while N(i) is defined in terms of E. Without an independent rule fixing N(i) (e.g., a row-wise top-k on S before Eq. 16), the graph construction cannot be executed, and the abstract's 'nearest neighbors in the similarity matrix' operation is not realized. This is a definitional circularity in the graph-construction step, although it does not propagate to the test-set metrics, which are computed on held-out data.

full rationale

The central claim (F1 0.934, AUC 0.963 on BreakHis; BLEU-4 0.811, METEOR 0.569 on PatchGastric) is evaluated on held-out test sets and is not defined in terms of the clustering loss, the GNN adjacency, or any fitted parameter, so the headline results are not circular. The paper cites standard external components (DEC [23], attention [24], GAT [26], pretrained ViT, and published LLMs) rather than a self-citation chain, and no uniqueness theorem is imported from the authors. The only genuine circular feature is Eq. (16), where the adjacency matrix is defined using the neighbor set N(i) that the adjacency matrix itself determines; this is a self-referential construction rather than a fitted-prediction circularity. The additional concern that the graph encodes cosine similarity of embeddings while the text calls it 'spatial relationships' is an interpretability/correctness gap, not a circular derivation: the reported gains may be misattributed to spatial context, but the metrics are not forced by that attribution. The paper also acknowledges its own limitation about K sensitivity, which further supports that the method is not presented as a self-proving derivation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central performance claim rests on standard MIL machinery, a transfer-learning assumption for pretrained vision encoders, and two ad hoc modeling choices: feature-similarity edges treated as spatial context, and a non-differentiable representative-selection step whose training behavior is unspecified. The number of clusters K is a hand-set parameter per dataset and is acknowledged as a weakness.

free parameters (2)
  • K, number of clusters = 8 (BreakHis), 50 (PatchGastric)
    Hand-selected in Section IV-B; no ablation or sensitivity analysis, and Section VI admits the method is sensitive to K and risks information loss.
  • L, number of GAT layers = 3
    Chosen in Section IV-B without ablation; determines how many graph hops are aggregated and is not justified experimentally.
assumptions (4)
  • standard math Standard MIL assumption: a bag is positive if at least one instance is positive (Eq. 1).
    Used to formulate WSI classification as MIL; inherited from prior MIL literature and not re-derived.
  • domain assumption Pretrained ImageNet ViT-B/16 and ResNet-34 features transfer to histopathology patches.
    The vision encoder is not retrained on histopathology; the method assumes its embeddings support clustering and graph construction (Section IV-B).
  • ad hoc to paper Cosine similarity between representative patch embeddings approximates spatial/contextual tissue relations.
    Microscopic WSIs lack coordinates, so graph edges in Eqs. (15)-(16) are feature-similarity edges; the paper labels this spatial information without evidence of correspondence to tissue adjacency.
  • ad hoc to paper The hard argmax selection in Eq. (13) is compatible with end-to-end gradient training.
    Representative patches are picked by a non-differentiable argmax, and the paper does not specify how losses backpropagate through this selection to the clustering or feature extractor.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GNN-ViTCap: GNN-Enhanced Multiple Instance Learning with Vision Transformers for Whole Slide Image Classification and Captioning." pith.science (2026). https://pith.science/paper/WN2CYDZR

@misc{pith2026250707006,
  author       = {Pith},
  title        = {Pith review of: GNN-ViTCap: GNN-Enhanced Multiple Instance Learning with Vision Transformers for Whole Slide Image Classification and Captioning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WN2CYDZR}},
  note         = {Machine review of arXiv:2507.07006}
}
read the original abstract

Microscopic assessment of histopathology images is vital for accurate cancer diagnosis and treatment. Whole Slide Image (WSI) classification and captioning have become crucial tasks in computer-aided pathology. However, microscopic WSI face challenges such as redundant patches and unknown patch positions due to subjective pathologist captures. Moreover, generating automatic pathology captions remains a significant challenge. To address these issues, we introduce a novel GNN-ViTCap framework for classification and caption generation from histopathological microscopic images. First, a visual feature extractor generates patch embeddings. Redundant patches are then removed by dynamically clustering these embeddings using deep embedded clustering and selecting representative patches via a scalar dot attention mechanism. We build a graph by connecting each node to its nearest neighbors in the similarity matrix and apply a graph neural network to capture both local and global context. The aggregated image embeddings are projected into the language model's input space through a linear layer and combined with caption tokens to fine-tune a large language model. We validate our method on the BreakHis and PatchGastric datasets. GNN-ViTCap achieves an F1 score of 0.934 and an AUC of 0.963 for classification, along with a BLEU-4 score of 0.811 and a METEOR score of 0.569 for captioning. Experimental results demonstrate that GNN-ViTCap outperforms state of the art approaches, offering a reliable and efficient solution for microscopy based patient diagnosis.

Figures

Figures reproduced from arXiv: 2507.07006 by the authors.

Figure 1
Figure 1. Overview of the GNN-ViTCap framework for microscopic whole slide images classification and captioning. (a) Scanner WSI where the absolute position of each patch is known. (b) Microscopic WSI lacks patch position and contains redundant patches due to multiple captures from a pathologist’s subjective perspective. (c) GNN-ViTCap architecture: (i) extracting the patches from whole slide image, (ii) extracting image embe… view at source ↗
Figure 2
Figure 2. t-SNE feature visualizations for the GNN-ViTCap for the BreakHis [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 14 canonical work pages

  1. [1]

    Inference of captions from histopatho- logical patches,

    M. Tsuneki and F. Kanavati, “Inference of captions from histopatho- logical patches,” in International Conference on Medical Imaging with Deep Learning. PMLR, 2022, pp. 1235–1250

  2. [2]

    Artificial intelligence for digital and compu- tational pathology,

    A. H. Song, G. Jaume, D. F. Williamson, M. Y . Lu, A. Vaidya, T. R. Miller, and F. Mahmood, “Artificial intelligence for digital and compu- tational pathology,” Nature Reviews Bioengineering , vol. 1, no. 12, pp. 930–949, 2023

  3. [3]

    Pathalign: A vision-language model for whole slide images in histopathology,

    F. Ahmed, A. Sellergren, L. Yang, S. Xu, B. Babenko, A. Ward, N. Olson, A. Mohtashamian, Y . Matias, G. S. Corrado et al., “Pathalign: A vision-language model for whole slide images in histopathology,” arXiv:2406.19578, 2024

  4. [4]

    Attention-based deep multiple instance learning,

    M. Ilse, J. Tomczak, and M. Welling, “Attention-based deep multiple instance learning,” in International conference on machine learning . PMLR, 2018, pp. 2127–2136

  5. [5]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification,

    Z. Shao, H. Bian, Y . Chen, Y . Wang, J. Zhang, X. Ji et al., “Transmil: Transformer based correlated multiple instance learning for whole slide image classification,” Advances in neural information processing systems, vol. 34, pp. 2136–2147, 2021

  6. [6]

    A survey on graph-based deep learning for computational histopathology,

    D. Ahmedt-Aristizabal, M. A. Armin, S. Denman, C. Fookes, and L. Petersson, “A survey on graph-based deep learning for computational histopathology,” Computerized Medical Imaging and Graphics , vol. 95, p. 102027, 2022

  7. [7]

    Enhanced descriptive captioning model for histopathological patches,

    S. Elbedwehy, T. Medhat, T. Hamza, and M. F. Alrahmawy, “Enhanced descriptive captioning model for histopathological patches,” Multimedia Tools and Applications, vol. 83, no. 12, pp. 36 645–36 664, 2024

  8. [8]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar et al. , “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023

Show all 29 references
  1. [9]

    Clinicalt5: A generative language model for clinical text,

    Q. Lu, D. Dou, and T. Nguyen, “Clinicalt5: A generative language model for clinical text,” inFindings of the Association for Computational Linguistics: EMNLP 2022 , 2022, pp. 5436–5443

  2. [10]

    Biogpt: generative pre-trained transformer for biomedical text genera- tion and mining,

    R. Luo, L. Sun, Y . Xia, T. Qin, S. Zhang, H. Poon, and T.-Y . Liu, “Biogpt: generative pre-trained transformer for biomedical text genera- tion and mining,” Briefings in bioinformatics, vol. 23, no. 6, p. bbac409, 2022

  3. [11]

    A generalist vision–language foundation model for diverse biomedical tasks,

    K. Zhang, R. Zhou, E. Adhikarla, Z. Yan, Y . Liu, J. Yu, Z. Liu, X. Chen, B. D. Davison, H. Ren et al., “A generalist vision–language foundation model for diverse biomedical tasks,” Nature Medicine, pp. 1–13, 2024

  4. [12]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large scale visual recognition challenge,” International journal of computer vision, vol. 115, pp. 211–252, 2015

  5. [13]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

  6. [14]

    Scaling up visual and vision-language representation learning with noisy text supervision,

    C. Jia, Y . Yang, Y . Xia, Y .-T. Chen, Z. Parekh, H. Pham, Q. Le, Y .-H. Sung, Z. Li, and T. Duerig, “Scaling up visual and vision-language representation learning with noisy text supervision,” in International conference on machine learning . PMLR, 2021, pp. 4904–4916

  7. [15]

    Pathm3: A multimodal multi-task multiple instance learning framework for whole slide image classification and captioning,

    Q. Zhou, W. Zhong, Y . Guo, M. Xiao, H. Ma, and J. Huang, “Pathm3: A multimodal multi-task multiple instance learning framework for whole slide image classification and captioning,” arXiv preprint arXiv:2403.08967, 2024

  8. [16]

    Dual-stream multiple instance learn- ing network for whole slide image classification with self-supervised contrastive learning,

    B. Li, Y . Li, and K. W. Eliceiri, “Dual-stream multiple instance learn- ing network for whole slide image classification with self-supervised contrastive learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 14 318–14 328

  9. [17]

    Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification,

    H. Zhang, Y . Meng, Y . Zhao, Y . Qiao, X. Yang, S. E. Coupland, and Y . Zheng, “Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification,” in Pro- ceedings of the IEEE/CVF conference on computer vision and patte...

  10. [18]

    Visual language pretrained multiple instance zero-shot transfer for histopathology images,

    M. Y . Lu, B. Chen, A. Zhang, D. F. Williamson, R. J. Chen, T. Ding, L. P. Le, Y .-S. Chuang, and F. Mahmood, “Visual language pretrained multiple instance zero-shot transfer for histopathology images,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern...

  11. [19]

    Cpath-omni: A unified multimodal foundation model for patch and whole slide image analysis in computational pathology,

    Y . Sun, Y . Si, C. Zhu, X. Gong, K. Zhang, P. Chen, Y . Zhang, Z. Shui, T. Lin, and L. Yang, “Cpath-omni: A unified multimodal foundation model for patch and whole slide image analysis in computational pathology,” arXiv:2412.12077, 2024

  12. [20]

    Large language models in healthcare and medical domain: A review,

    Z. A. Nazi and W. Peng, “Large language models in healthcare and medical domain: A review,” in Informatics, vol. 11, no. 3. MDPI, 2024, p. 57

  13. [21]

    What a whole slide image can tell? subtype-guided masked transformer for pathological image captioning,

    W. Qin, R. Xu, P. Huang, X. Wu, H. Zhang, and L. Luo, “What a whole slide image can tell? subtype-guided masked transformer for pathological image captioning,” arXiv preprint arXiv:2310.20607, 2023

  14. [22]

    Patch-based convolutional neural network for whole slide tissue image classification,

    L. Hou, D. Samaras, T. M. Kurc, Y . Gao, J. E. Davis, and J. H. Saltz, “Patch-based convolutional neural network for whole slide tissue image classification,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 2424–2433

  15. [23]

    Unsupervised deep embedding for clustering analysis,

    J. Xie, R. Girshick, and A. Farhadi, “Unsupervised deep embedding for clustering analysis,” in International conference on machine learning . PMLR, 2016, pp. 478–487

  16. [24]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

  17. [25]

    Categorical reparameterization with gumbel-softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,” arXiv preprint arXiv:1611.01144 , 2016

  18. [26]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengio et al., “Graph attention networks,” stat, vol. 1050, no. 20, pp. 10–48 550, 2017

  19. [27]

    Rectifier nonlinearities improve neural network acoustic models,

    A. L. Maas, A. Y . Hannun, A. Y . Ng et al. , “Rectifier nonlinearities improve neural network acoustic models,” in Proc. icml, vol. 30, no. 1. Atlanta, GA, 2013, p. 3

  20. [28]

    A dataset for breast cancer histopathological image classification,

    F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte, “A dataset for breast cancer histopathological image classification,” Ieee transactions on biomedical engineering , vol. 63, no. 7, pp. 1455–1462, 2015

  21. [29]

    A survey of evaluation metrics used for nlg systems,

    A. B. Sai, A. K. Mohankumar, and M. M. Khapra, “A survey of evaluation metrics used for nlg systems,” ACM Computing Surveys (CSUR), vol. 55, no. 2, pp. 1–39, 2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.