Pith. sign in

REVIEW 3 major objections 4 minor 48 references

Gated Spatial Redundancy Projection for Pathology Transformer Attentions

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Gated SRP claims that a post-attention projection along a local neighborhood redundancy axis, with a learned signed gate, improves pathology transformer predictions at negligible parameter cost.

desk verdict A clean, lightweight attention correction with a plausible pathology-specific motivation, but the empirical case is undercut by an undisclosed per-dataset hyperparameter selection and small gains relative to seed noise. read the letter →

arxiv 2608.08374 v2 pith:XRCFGVBP submitted 2026-08-08 cs.CV

Zhiyuan Yang , Jiahao Cheng , Vincent Quoc-Huy Trinh , Mahdi S. Hosseini This is my paper · ORCID
classification cs.CV
keywords whole-slidepathologytransformerattentionlocalspatialredundancyself-attentioncorrectionsurvivalanalysisslide-levelclassificationgatedprojectionC-index
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

Whole-slide pathology images are full of near-duplicate tissue patches, and the paper argues that this local redundancy is a specific failure mode for transformer attention: neighboring patches get mixed into each other and dilute the subtle evidence that matters for diagnosis and prognosis. To counteract this, the paper proposes Gated Spatial Redundancy Projection (Gated SRP), a small post-attention module that, for each patch and attention head, estimates a local redundancy direction from neighboring value vectors, projects the attention output onto that direction, and subtracts a learned signed fraction of the projected component. Because the gate is initialized at zero, standard attention is exactly recoverable during training, and the module adds roughly 0.02% of total parameters. In the paper's experiments, this correction yields the highest mean concordance index among compared attention variants on all five tested survival cohorts and improves the base attention on 12 of 16 slide-level classification metrics. If correct, a negligible-cost geometric correction to attention outputs, rather than a new model or loss, can make pathology transformers more sensitive to diagnostically relevant deviations.

What carries the argument

The load-bearing object is the gated spatial redundancy projection $z_{i,h} = y_{i,h} - \beta_{i,h}\,\mathrm{proj}_{\hat{r}_{i,h}}(y_{i,h})$ applied after each self-attention layer to patch tokens only. The redundancy axis $\hat{r}_{i,h}$ is the L2-normalized mean of the neighboring value vectors in a $3\times3$ spatial window, computed with stop-gradient so it acts as a fixed geometric reference; the signed coefficient $\beta_{i,h} = \delta \tanh(\ell_{i,h})$ comes from a factored gate logit combining a shared token-level MLP (tissue homogeneity, neighbor count, log-confidence) with head-level alignment features (cosine similarity to the axis, absolute alignment, log-magnitude) and per-layer/head biases. The gate is initialized so $\beta=0$, making the original attention update exactly reachable, which means the module can only learn to deviate from base attention when the training signal supports it. This single equation converts a suspected failure mode into a parameterized geometric choice: identity, projection, anti-projection, or reflection of the neighborhood-aligned component.

What would settle it

A decisive check is to hold out the test labels of all ten datasets, select $\delta$ and $h$ using only validation-set search (for example, a small grid over $\delta \in \{0.5,1,1.5,2\}$ and $h \in \{16,32,64,128\}$), and then evaluate once. If Gated SRP no longer improves mean C-index on all five survival cohorts and more than half of classification metrics under this protocol, the central claim of consistent gains is not supported. A complementary mechanistic check: measure the cosine similarity between patch tokens and their neighborhood mean after correction; the method predicts that tokens with large positive $\beta$ should become less similar to the local mean while retaining predictive risk signal.

Watch

Extended reading notes

Core claim

The central claim is that the local spatial redundancy of whole-slide images—neighboring patches that share tissue type, stain, texture, and cellular composition—can repeatedly inject redundant content into patch tokens inside self-attention layers, making subtle diagnostic or prognostic deviations less distinguishable. Gated SRP addresses this by defining, per token and head, a local redundancy axis $\hat{r}$ as the L2-normalized mean of the neighboring value vectors (computed with stop-gradient), projecting the attention output $y$ onto that axis, and writing the corrected output $z = y - \beta_{\mathrm{eff}} \mathrm{proj}_{\hat{r}}(y)$, where $\beta_{\mathrm{eff}} = \delta \tanh(\ell)$ is a bounded signed gate produced by a factored token/head logit. The paper reports that, with this correction, mean case-level C-index improves over the base attention on all five cohorts (for example, from 0.7247 to 0.7648 on one kidney cohort and from 0.5513 to 0.5832 on one lung cohort), and classification performance improves on 12 of 16 reported metrics, while the module contributes an average of +0.02% parameters. The paper also claims that ablations support each design choice: the learned signed gate outperforms fixed projection strengths and a nonnegative sigmoid gate, detached geometric features outperform live gradients, and zero initialization outperforms random initialization, so the correction is learned gradually from an identity start.

Load-bearing premise

The reported gains rest on the assumption that the per-dataset choices of the gate range $\delta$ and gate hidden size $h$ (listed in Tables 10 and 11) were fixed honestly before observing test performance; the paper does not describe the selection procedure, so if those hyperparameters were tuned against the test labels, the mean improvements could be inflated.

Editorial extensions

If this is right

  • Gated SRP improves mean case-level C-index over the base attention on every one of the five survival cohorts, with reported gains such as 0.7247 to 0.7648 on one kidney cohort and 0.5513 to 0.5832 on one lung cohort.
  • Across five classification datasets, it improves 12 of 16 reported metrics, reaches the best mean AUC on three datasets, and takes the best F1 and accuracy on two of them.
  • The identity initialization makes base attention exactly reachable at the first step, so the module can fall back to the original attention whenever correction is not useful, at an average cost of +0.02% of total parameters.
  • Ablations attribute the gain to the learned signed gate: fixed projection strengths of 1, 2, and -1 are less reliable, and a nonnegative sigmoid gate underperforms the signed tanh gate.
  • The correction also improves all reported metrics in the dense-attention vision transformer experiments on two image tasks, indicating the mechanism is not tied to the approximate-attention backbone.

Reading between the lines

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

  • An untested consequence is that the same redundancy-projection mechanism should transfer to other imagery with near-duplicate neighboring patches, such as satellite surveys, wide-field microscopy, or video frames; confidence in the pathology-specific story would grow if Gated SRP improves those tasks too.
  • The signed gate's negative branch lets the model amplify alignment with the neighborhood mean; a concrete test is whether tokens with strongly negative $\beta$ coincide with tumor-stroma boundaries or micro-environmental gradients in survival cohorts, which would connect the geometric correction to histology.
  • The current redundancy axis is unsupervised geometry—the mean of neighboring value vectors; extending it to use tissue-type segmentation labels (for example, marking the $3\times3$ window as homogeneous or mixed in annotated class) could sharpen the axis without changing the architecture.
  • The paper trains on frozen patch features and lists pretraining as open work; if correct, injecting this correction during self-supervised pretraining may improve the base representations themselves, but that is an extrapolation beyond the reported experiments.
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 / 4 minor

Summary. The paper proposes Gated Spatial Redundancy Projection (Gated SRP), a lightweight post-attention module for pathology transformers. For each patch token and attention head, the module computes a local redundancy axis from neighbouring value vectors, projects the attention output onto that axis, and subtracts a signed, token-specific learned coefficient times the projected component. The module is designed to reduce to standard attention at initialization through zero initialization of the gate, and it adds about 0.02% parameters. The authors evaluate Gated SRP against base Nyström attention, XSA, and differential attention on five TCGA survival cohorts and five slide-level classification datasets, reporting the highest mean C-index on all survival cohorts and improvements on 12 of 16 classification metrics. They also provide ablations on the gate range, gate gradients, patch encoder, gate factorization, gate initialization, and architecture choice, plus qualitative pathologist review of attention maps.

Significance. If the empirical claims are supported, Gated SRP is a useful, conceptually clean contribution to computational pathology: it identifies a pathology-specific failure mode (local spatial redundancy), proposes a geometrically interpretable correction, preserves the base attention path exactly at initialization, and adds negligible parameters. The identity initialization and the public code release are particular strengths, and the ablation suite is reasonably comprehensive. However, the headline empirical claim currently rests on mean differences across five seeds that are not accompanied by significance testing, and the per-dataset selection of two gate hyperparameters is not disclosed in a way that rules out test-set influence. These issues make the significance contingent on additional statistical and protocol evidence rather than established by the present manuscript.

major comments (3)
  1. [Section B, Hyperparameter Protocol; Tables 10 and 11] The per-dataset selection of the gate range δ and gate hidden dimension h is not documented beyond the statement that 'the selected values are fixed before final result aggregation.' The paper does not report the search grid, the split used for selection, or the selection criterion. Since the text lists δ ∈ {0.5, 1, 1.5, 2} and h ∈ {16, 32, 64, 128}, this is a 16-configuration search per dataset or cohort; if the test split or test labels influenced this selection, the reported mean gains and the across-cohort ranking could be overstated. This concern is load-bearing because the central claim of 'highest mean C-index on all five cohorts' depends directly on these chosen configurations. The authors should disclose the full selection procedure, use a nested or non-overlapping validation split for configuration selection, and report results for a single fixed configuration as a sensitivity check.
  2. [Appendix D, Tables 13 and 14; Tables 1 and 2] No paired significance tests or confidence intervals are reported, and the magnitudes of the gains are small relative to the five-seed standard deviations. For example, in Table 13, the KIRP mean gain is 0.0401 while Gated SRP's standard deviation is 0.0805, and the LUAD gain is 0.0319 with a standard deviation of 0.0492; all five survival gains are smaller than one standard deviation. Given that the same seeds are used for all methods, paired comparisons are available and should be reported. Without such tests, the claim that Gated SRP 'obtains the highest mean C-index' is fragile, and the '12 of 16 metrics' claim in the abstract is not statistically supported.
  3. [Section 4.1 and Section B, Hyperparameter Protocol] The comparison is not equally matched: standard NA, XSA, and Diff do not receive any per-dataset configuration search, while Gated SRP receives dataset-specific tuning of both the correction range and gate capacity. Even under the benign reading that selection was performed on a validation split, the reported advantage could reflect the benefit of tuning rather than the mechanism. The authors should either run a comparable hyperparameter search for each baseline, or clearly report the selection protocol and demonstrate that the relative ordering is stable across a fixed configuration of δ and h for Gated SRP.
minor comments (4)
  1. [Tables 1 and 2 captions] The captions say green/red indicate 'meaningful' increase/decrease relative to the base NA, but no threshold or statistical definition of 'meaningful' is given; please define this term or remove it.
  2. [Figure 3 and Figure 2] The panel labels in Figure 3 appear garbled ('zre f l zpro j ziden =y zanti'), and Figure 2's panel layout is described in the caption but not clearly labeled in the text; please fix the rendering and refer to panels explicitly where needed.
  3. [Section 3.3, Eq. (6)] The text says n=3 gives an 8-neighbour grid for patch tokens, but the abstract and Figure 1 describe a '3×3-neighbour mean'; for patch-level images the 'analogous fixed grid' is not defined precisely. Please clarify how the neighbourhood is constructed when patch coordinates are unavailable.
  4. [Appendix B, Model architecture] The deviations from the original TransMIL architecture (width, depth, block topology, Nyström landmark count, disabled value-residual depthwise convolution, regularization) are extensive; the main text should point readers to Appendix B at the first mention of 'TransMIL-style' so that the actual baseline is not misread as the official TransMIL.

Circularity Check

0 steps flagged · score 1.0 of 10

No in-construction circularity: Gated SRP's gains are empirical and externally benchmarked; the only adjacent concern is undisclosed per-dataset hyperparameter selection, which is an experimental-validity issue rather than a reduction of the derivation to its inputs.

full rationale

I walked the derivation chain from the proposed correction z_{i,h} = y_{i,h} - beta_{eff,i,h} c_{i,h} (Eq. 5) back to the definition of the local redundancy axis r_{i,h} as a neighbourhood mean of detached value vectors (Eq. 6), the projection c = (y^T r_hat) r_hat (Eq. 4), and the signed gate beta = delta tanh(logit) (Eq. 7). These equations define the module; they do not presuppose the outcome that Gated SRP improves C-index or classification metrics. The identity initialization (all gate output weights and biases zero, so beta=0 at step 1) is an architectural property that recovers base attention; it is not a circular prediction, because the reported gains are measured on held-out test cases over five seeds rather than solved for by the construction. No fitted parameter is renamed as a prediction: the gate parameters are learned, and the reported survival and classification numbers are independent evaluations. The self-citations to AtlasPatch, ADP, ADPv2, and MOOZY are tooling, dataset, and related-work references; none is invoked as an external theorem that forces the Gated SRP design. The closest concern is Section B's Hyperparameter Protocol, which states that 'the selected values are fixed before final result aggregation' without describing the search grid, the split used for selection, or the selection criterion. Per-dataset choices of delta and h (Tables 10-12) could in principle inflate the reported mean gains if test labels influenced selection, and the gains are modest relative to reported standard deviations. However, this is a selection and reporting validity issue, not an in-construction circularity: the paper's equations do not make the reported improvement equal to a fitted quantity or to a self-citation. Under the required quote-and-reduction standard, no circular step is established.

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

No new physical entities are introduced; Gated SRP is a modular addition to a neural network. The local spatial redundancy axis is a computational direction, not a postulated entity with independent evidence. The main free parameters are the per-dataset gate range and hidden size, whose selection procedure is not documented.

free parameters (2)
  • Gate range delta (per dataset or cohort) = Table 11 values, e.g., KIRC 1.5, KIRP 2.0, LUAD 1.0, STAD 0.5, UCEC 1.5, CAM16 1.0, CAM17 2.0, KGH 1.0, PANDA 0.5…
    Chosen per dataset and controls the maximum correction magnitude; selection procedure is not described.
  • Gate hidden dimension h (per dataset or cohort) = Table 11 values, e.g., KIRC 128, KIRP 32, LUAD 128, STAD 128, UCEC 32, CAM16 16, CAM17 64, KGH 64, PANDA 16, BRACS 128
    Controls gate capacity; selected per dataset, selection procedure is not described.
assumptions (4)
  • domain assumption Patch tokens lie on a meaningful 2D grid, so an 8-neighbour 3x3 window defines spatial adjacency.
    Eq. 6 builds N(i) from patch coordinates; if the grid is not spatially meaningful, the redundancy direction is arbitrary.
  • domain assumption The mean of neighbouring value vectors is the correct local redundancy direction to correct.
    Eqs. 4 to 6 assume the neighbourhood mean captures local tissue, stain, and texture redundancy; supported only by the cosine-similarity observation in Figure 1, not by a formal argument.
  • domain assumption A one-dimensional signed correction along r_hat is sufficient to preserve or amplify diagnostic deviations.
    Eq. 5 changes only the component along r_hat; any useful signal orthogonal to the redundancy axis is left untouched, and the paper does not analyse whether diagnostic deviations are typically orthogonal.
  • domain assumption Nyström attention with 64 landmarks is a faithful approximation for evaluating the correction mechanism.
    Main experiments use m=64 landmarks (Appendix B, Model architecture); if the approximation distorts per-token outputs, the correction's effect could differ for dense attention.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Gated Spatial Redundancy Projection for Pathology Transformer Attentions." pith.science (2026). https://pith.science/paper/XRCFGVBP

@misc{pith2026260808374,
  author       = {Pith},
  title        = {Pith review of: Gated Spatial Redundancy Projection for Pathology Transformer Attentions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XRCFGVBP}},
  note         = {Machine review of arXiv:2608.08374}
}
read the original abstract

Transformer models are increasingly used for whole-slide image analysis in computational pathology. Yet, WSIs differ fundamentally from natural images: neighbouring patches often contain highly similar tissue type, stain, texture, and cellular composition. We identify this local spatial redundancy as a pathology-specific failure mode of self-attention, where dominant neighbourhood features can be repeatedly mixed into patch-tokens and weaken subtle diagnostic or prognostic deviations. We propose Gated Spatial Redundancy Projection (Gated SRP), a lightweight drop-in correction module for self-attention layers. For each patch token and attention head, Gated SRP estimates a local redundancy axis from neighbouring value vectors, projects the attention output onto this axis, and applies a learned signed gate to correct the redundancy-aligned component geometrically. Across five TCGA survival cohorts, Gated SRP obtains the highest mean C-index among the compared attention variants in all cohorts, with an average improvement over the base attention, while adding only +0.02% parameters. Across five slide-level classification datasets, it improves the base attention on 12 of 16 reported metrics and achieves the best AUC on three datasets. Code is publicly available at https://github.com/AtlasAnalyticsLab/GatedSRP.

Figures

Figures reproduced from arXiv: 2608.08374 by the authors.

Figure 1
Figure 1. Local spatial redundancy in pathology. (a) Nearby crops in a natural image often contain distinct semantic content. (b) Nearby crops in a whole-slide image (WSI) often show the same tissue type, stain, and texture. (c) In transformer models, after self￾attention layers, each WSI patch token has a high cosine similarity to its 3 × 3-neighbour mean, whereas the corresponding similarities for natural image patches are … view at source ↗
Figure 2
Figure 2. Method overview. (a) Gated SRP is a lightweight drop-in module inserted after a self-attention layer, adding only ~0.02% parameters. (b) For each patch-token attention output y, Gated SRP uses the local 3×3 neighbourhood to estimate a redundancy direction ˆr. It projects y onto this direction to obtain the neighbourhood-aligned component c = projrˆ (y). A learned gate predicts a signed coefficient β, and the correct… view at source ↗
Figure 3
Figure 3. Geometric illustration of the signed gate. (a) For a fixed local redundancy axis ˆr, the redundant component c = (y ·rˆ)rˆ lies along ˆr, and the signed gate writes z = y − βeff c. Varying βeff therefore moves the output only along ˆr. The identity endpoint is labelled y = ziden. The special cases are labelled as z variants: projection, reflection, and anti-projection (amplification). (b) The signed gate places thes… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Effective gate trajectories for representative datasets. We plot βeff = δ tanh(g), where g is the learned gate logit, for the three gate-active Gated SRP layers. Curves show the mean across five global seeds, where each seed-level point averages the logged training￾bat…
Figure 5
Figure 5. Figure 5: Qualitative comparison of attention maps. One representative WSI from some datasets. Each heatmap overlays the final layer CLS-to-patch attention score on the WSI thumbnail and is normalized within each slide and method for visualization. For Diff, the final attention …
Figure 6
Figure 6. Figure 6: Effective gate coefficient trajectories for TCGA survival cohorts. The addi￾tional TCGA survival cohorts. Curves and shaded regions follow the same five-seed mean and ±1 standard-deviation convention. It is important to distinguish the selected range parameter and the …
Figure 7
Figure 7. Figure 7: Effective gate coefficient trajectories for WSI classification tasks. The addi￾tional classification datasets [PITH_FULL_IMAGE:figures/full_fig_p028_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 6 linked inside Pith

  1. [1]

    Hosseini

    Ahmed Alagha, Christopher Leclerc, Yousef Kotp, Omar Metwally, Calvin Moras, Pe- ter Rentopoulos, Ghodsiyeh Rostami, Bich Ngoc Nguyen, Jumanah Baig, Abdelhakim Khellaf, Vincent Quoc-Huy Trinh, Rabeb Mizouni, Hadi Otrok, Jamal Bentahar, and Mahdi S. Hosseini. Atlaspatch: Efficient tissue detection and high-throughput patch ex- traction for computational pa...

    arXiv 2026

  2. [2]

    Tizhoosh

    Areej Alsaafin, Peyman Nejat, Abubakr Shafique, Jibran Khan, Saghir Alfasly, Ghazal Alabtah, and Hamid R. Tizhoosh. Sequential patching lattice for image classifi- cation and enquiry: Streamlining digital pathology image processing.The Ameri- can Journal of Pathology, 194(10):1898–1912, 2024. ISSN 0002-9440. doi: https: //doi.org/10.1016/j.ajpath.2024.06....

    doi:10.1016/j.ajpath.2024.06.007 1912

  3. [3]

    Token merging: Your vit but faster, 2023

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichten- hofer, and Judy Hoffman. Token merging: Your vit but faster, 2023. URLhttps: //arxiv.org/abs/2210.09461

    arXiv 2023

  4. [4]

    Bracs: A dataset for breast carcinoma subtyping in h&e histology images.Database, 2022:baac093, 01 2022

    Nadia Brancati, Anna Maria Anniciello, Pushpak Pati, Daniel Riccio, Giosuè Scog- namiglio, Guillaume Jaume, Giuseppe De Pietro, Maurizio Di Bonito, Antonio Fon- cubierta, Gerardo Botti, Maria Gabrani, Florinda Feroce, and Maria Frucci. Bracs: A dataset for breast carcinoma subtyping in h&e histology images.Database, 2022:baac093, 01 2022. ISSN 1758-04...

    doi:10.1093/database/baac093 2022

  5. [5]

    Steiner, Hester van Boven, Robert Vink, Christina Hulsbergen-van de Kaa, Jeroen van der Laak, Mahul B

    Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinck- aers, Kunal Nagpal, Yuannan Cai, David F. Steiner, Hester van Boven, Robert Vink, Christina Hulsbergen-van de Kaa, Jeroen van der Laak, Mahul B. Amin, Andrew J. Evans, Theodorus van der Kwast, Robert Allan, Peter A. Humphrey, Henrik Grön- berg, Hemamali Samaratunga, Brett Dela...

    2022

  6. [6]

    Péter Bándi, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balken- hol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, Quanzheng Li, Farhad Ghazvinian Zanjani, Svitlana Zinger, Keisuke Fukuta, Daisuke Komura, Vlado Ovtcharov, Shenghua Cheng, Shaoqun Zeng, Jeppe Thagaard, Anders B. Dahl, Huangjing Lin, H...

    arXiv 2019

  7. [7]

    Chen, Ming Y

    Richard J. Chen, Ming Y . Lu, Muhammad Shaban, Chengkuan Chen, Tiffany Y . Chen, Drew F. K. Williamson, and Faisal Mahmood. Whole slide images are 2d point clouds: Context-aware survival prediction using patch-based graph convolutional networks. In Marleen de Bruijne, Philippe C. Cattin, Stéphane Cotin, Nicolas Padoy, Stefanie Spei- del, Yefeng Zheng, and...

    2021

  8. [8]

    Chen, Chengkuan Chen, Yicong Li, Tiffany Y

    Richard J. Chen, Chengkuan Chen, Yicong Li, Tiffany Y . Chen, Andrew D. Trister, Rahul G. Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel im- ages via hierarchical self-supervised learning. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 16144–16155, June 2022

    2022

Show all 48 references
  1. [9]

    Towards a general-purpose foundation model for computational pathology.Na- ture Medicine, 2024

    Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Bowen Chen, Andrew Zhang, Daniel Shao, Andrew H Song, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology.Na- ture Medicine, 2024

    2024

  2. [10]

    Tc-ssa: Token compression via semantic slot aggregation for gigapixel pathology reasoning, 2026

    Zhuo Chen, Shawn Young, and Lijian Xu. Tc-ssa: Token compression via semantic slot aggregation for gigapixel pathology reasoning, 2026. URLhttps://arxiv. org/abs/2603.01143

    2026

  3. [11]

    Convit: Improving vision transformers with soft convolutional in- ductive biases

    Stéphane D’Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional in- ductive biases. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Mach...

    2021

  4. [12]

    A multimodal whole-slide foundation model for pathology.Nature Medicine, pages 1–13, 2025

    Tong Ding, Sophia J Wagner, Andrew H Song, Richard J Chen, Ming Y Lu, Andrew Zhang, Anurag J Vaidya, Guillaume Jaume, Muhammad Shaban, Ahrong Kim, et al. A multimodal whole-slide foundation model for pathology.Nature Medicine, pages 1–13, 2025

    2025

  5. [13]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.ICLR, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Trans- formers for image recognitio...

    2021

  6. [14]

    Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Gin- neken, Nico Karssemeijer, Geert Litjens, Jeroen A. W. M. van der Laak, , and the CAMELYON16 Consortium. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in wome...

    2017

  7. [15]

    Understanding the difficulty of training deep feed- forward neural networks

    Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feed- forward neural networks. In Yee Whye Teh and Mike Titterington, editors,Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, vol- ume 9 ofProceedin...

    2010

  8. [16]

    Focus: Knowledge-enhanced adaptive visual compression for few- shot whole slide image classification

    Zhengrui Guo, Conghao Xiong, Jiabo Ma, Qichen Sun, Lishuang Feng, Jinzhuo Wang, and Hao Chen. Focus: Knowledge-enhanced adaptive visual compression for few- shot whole slide image classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2025

  9. [17]

    Neighborhood attention transformer

    Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6185–6194, June 2023

    2023

  10. [18]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV ’15, page 1026–1034, USA, 2015. IEEE Comp...

    2015 doi

  11. [19]

    Hosseini, Lyndon Chan, Gabriel Tse, Michael Tang, Jun Deng, Sajad Norouzi, Corwyn Rowsell, Konstantinos N

    Mahdi S. Hosseini, Lyndon Chan, Gabriel Tse, Michael Tang, Jun Deng, Sajad Norouzi, Corwyn Rowsell, Konstantinos N. Plataniotis, and Savvas Damaskinos. At- las of digital pathology: A generalized hierarchical histological tissue type-annotated database for deep learning. InPro...

    2019

  12. [20]

    Mahdi S. Hosseini, Babak Ehteshami Bejnordi, Vincent Quoc-Huy Trinh, Lyndon Chan, Danial Hasan, Xingwen Li, Stephen Yang, Taehyo Kim, Haochen Zhang, Theodore Wu, Kajanan Chinniah, Sina Maghsoudlou, Ryan Zhang, Jiadai Zhu, Samir Khaki, Andrei Buin, Fatemeh Chaji, Ala Salehi, Bi...

    2024

  13. [21]

    Integral transformer: Denoising attention, not too much not too little

    Ivan Kobyzev, Abbas Ghaddar, Dingtao Hu, and Boxing Chen. Integral transformer: Denoising attention, not too much not too little. In Christos Christodoulopoulos, Tan- moy Chakraborty, Carolyn Rose, and Violet Peng, editors,Proceedings of the 2025 Conference on Empirical Method...

    2025 doi

  14. [22]

    Hosseini

    Yousef Kotp, Vincent Quoc-Huy Trinh, Christopher Pal, and Mahdi S. Hosseini. Moozy: A patient-first foundation model for computational pathology, 2026. URL https://arxiv.org/abs/2603.27048

    2026 arXiv

  15. [23]

    Eliceiri

    Bin Li, Yin Li, and Kevin W. Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14318–14328...

    2021

  16. [24]

    Pathvq: Reforming computational pathology foundation model for whole slide image analysis via vector quantization

    Honglin Li, Zhongyi Shui, Yunlong Zhang, Chenglu Zhu, and Lin Yang. Pathvq: Reforming computational pathology foundation model for whole slide image analysis via vector quantization. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors,Ad...

    2025

  17. [25]

    Not all patches are what you need: Expediting vision transformers via token reorgani- zations, 2022

    Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorgani- zations, 2022. URLhttps://arxiv.org/abs/2202.07800

    2022 arXiv

  18. [26]

    Swin transformer: Hierarchical vision transformer using shifted win- dows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted win- dows. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012–10022, October 2021

    2021

  19. [27]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URLhttps:// openreview.net/forum?id=Bkg6RiCqY7

    2019

  20. [28]

    Data-efficient and weakly supervised computational pathology on whole-slide images.Nature Biomedical Engineering, 5(6):555–570, 2021

    Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images.Nature Biomedical Engineering, 5(6):555–570, 2021

    2021

  21. [29]

    A visual- language foundation model for computational pathology.Nature Medicine, 30:863—- 874, 2024

    Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual- language foundation model for computational pathology.Nature Medicine, 30:863—- 874, 2024. Y ANG ET AL.: GA TED SPA TIAL...

    2024

  22. [30]

    More expressive attention with negative weights, 2025

    Ang Lv, Ruobing Xie, Shuaipeng Li, Jiayi Liao, Xingwu Sun, Zhanhui Kang, Di Wang, and Rui Yan. More expressive attention with negative weights, 2025. URLhttps: //arxiv.org/abs/2411.07176

    2025 arXiv

  23. [31]

    Differential gated self- attention, 2025

    Elpiniki Maria Lygizou, Mónika Farsang, and Radu Grosu. Differential gated self- attention, 2025. URLhttps://arxiv.org/abs/2505.24054

    2025

  24. [32]

    Alexis Marouani, Oriane Siméoni, Hervé Jégou, Piotr Bojanowski, and Huy V . V o. Revisiting [cls] and patch token interaction in vision transformers, 2026. URLhttps: //arxiv.org/abs/2602.08626

    2026

  25. [33]

    The Cancer Genome Atlas Program (TCGA).https: //www.cancer.gov/ccg/research/genome-sequencing/tcga

    National Cancer Institute. The Cancer Genome Atlas Program (TCGA).https: //www.cancer.gov/ccg/research/genome-sequencing/tcga. Ac- cessed: 2026-04-28

    2026

  26. [34]

    Dynamicvit: Efficient vision transformers with dynamic to- ken sparsification

    Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho- Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic to- ken sparsification. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Informatio...

  27. [35]

    Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia An- gelova

    Michael S. Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia An- gelova. Tokenlearner: What can 8 learned tokens do for images and videos?, 2022. URLhttps://arxiv.org/abs/2106.11297

    2022 arXiv

  28. [36]

    Medgemma technical report.arXiv preprint arXiv:2507.05201, 2025

    Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Tra- verse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, et al. Medgemma technical report.arXiv preprint arXiv:2507.05201, 2025

    2025 arXiv

  29. [37]

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

    Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, and yongbing zhang. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan,...

  30. [38]

    Going deeper with image transformers

    Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV), pages 32–42, October 2021

    2021

  31. [39]

    URLhttps://proceedings.neurips.cc/paper_files/paper/ 2021/file/10c272d06794d3e5785d5e7c5356e9ff-Paper.pdf

    2021

  32. [40]

    Nyströmformer: A nyström-based algorithm for approxi- mating self-attention.Proceedings of the AAAI Conference on Artificial Intelligence, 35(16):14138–14148, May 2021

    Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nyströmformer: A nyström-based algorithm for approxi- mating self-attention.Proceedings of the AAAI Conference on Artificial Intelligence, 35(16):14138–14148, May 2021. doi: ...

    2021 doi

  33. [41]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wal- lach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neura...

  34. [42]

    Hosseini

    Zhiyuan Yang, Kai Li, Sophia Ghamoshi Ramandi, Patricia Brassard, Abdelhakim Khellaf, Vincent Quoc-Huy Trinh, Jennifer Zhang, Lina Chen, Corwyn Rowsell, Sonal Varma, Kostas Plataniotis, and Mahdi S. Hosseini. Adpv2: A hierarchical histological tissue type-annotated dataset for...

    2026

  35. [43]

    Dif- ferential transformer

    Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Dif- ferential transformer. In Y . Yue, A. Garg, N. Peng, F. Sha, and R. Yu, editors,Interna- tional Conference on Learning Representations, volume 2025, pages 144–164, 2025. URLhttps://proceedings.ic...

    2025

  36. [44]

    Wright, Ari Ro- bicsek, Brian Piening, Carlo Bifulco, Sheng Wang, and Hoifung Poon

    Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Nau- mann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, Yanbo Xu, Mu Wei, Wen- hui Wang, Shuming Ma, Furu Wei, Jianwei Yang, Chunyuan Li, Jianfeng Gao, Jaylen Rosemon, Tucker Bower, Soohee Lee, Ros...

    2024

  37. [46]

    bhnt,ht->bhn

    For each seed, we generate an approximate 70%, 10%, and 20% split for training, vali- dation, and testing. The same split is reused by all compared methods under the same dataset and seed, which gives a paired comparison across attention modules. The split unit is chosen to av...

  38. [47]

    Exclusive self attention, 2026

    Shuangfei Zhai. Exclusive self attention, 2026. URLhttps://arxiv.org/abs/ 2603.09078. Y ANG ET AL.: GA TED SPA TIAL REDUNDANCY PROJECTION21 A Dataset Details CAMELYON16.CAMELYON16 [14] is a lymph-node metastasis benchmark for breast cancer. We use it to perform binary WSI-leve...

    2026

  39. [2017]

    20Y ANG ET AL.: GA TED SPA TIAL REDUNDANCY PROJECTION

    URLhttps://proceedings.neurips.cc/paper_files/paper/ 2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf. 20Y ANG ET AL.: GA TED SPA TIAL REDUNDANCY PROJECTION

    2017

  40. [2021]

    URLhttps://proceedings.neurips.cc/paper_files/paper/ 2021/file/747d3443e319a22747fbb873e8b2f9f2-Paper.pdf

    2021

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.