acceptodds
Under review as a conference paper at ICLR 2027

Every Scale Has Its Geometry: Rethinking Positional Encoding for Multi-Resolution Multiple Instance Learning

Abstract

Multiple Instance Learning (MIL) enables weakly supervised analysis of Whole Slide Images (WSIs) without costly patch-level annotations, but its patch-based formulation inherently disrupts the spatial organization of tissue structures. Positional encodings (PEs) have therefore been introduced to restore spatial information lost during patchification. More recently, Multi-Resolution MIL (MRMIL) has leveraged complementary information across magnifications by incorporating patches at multiple resolutions, better reflecting the multi-scale visual assessment performed in clinical diagnosis. However, existing PEs are not well suited to the geometry of multi-resolution WSIs (MR-WSIs): they either model spatial relationships independently within each resolution or inadequately account for the heterogeneous spatial scales induced by different magnifications. In particular, topology-aware PEs typically treat cross-resolution spatial relationships uniformly, even though the spatial domain represented by patches expands quadratically with increasing magnification. To address this limitation, we propose Hyperbolic Hierarchical Positional Encoding (H2PE), a positional representation that jointly captures spatial topology and resolution hierarchy within the hyperbolic upper half-space. By leveraging the exponential capacity of hyperbolic geometry, H2PE naturally represents the hierarchical, scale-dependent structure of multi-resolution patches while preserving their spatial relationships across magnifications. H2PE is model-agnostic and can be incorporated into arbitrary MRMIL aggregation frameworks without modifying their underlying aggregation architectures. Extensive experiments on three public benchmarks demonstrate consistent improvements over existing positional encodings across multiple MRMIL frameworks. Further geometric analyses and visualizations validate that H2PE effectively captures the hierarchical organization and cross-resolution spatial structure of MR-WSIs, while providing interpretable positional representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multiple Instance Learning, Whole Slide Images, Positional Encoding

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cbXUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes H2PE, a positional encoding for multi-resolution multiple instance learning on whole slide images. H2PE uses MPP to map patches at multiple magnifications onto a shared hyperbolic upper half-space to capture spatial and cross-scale relationships. Experiments with three MRMIL frameworks on CAMELYON16, BRACS, and BCNB show consistent improvements over existing positional encodings.

Critical strengths
  • The paper points out a gap that has not been fully explored. Existing location encoding either models each resolution independently or ignores the scale-dependent geometry of multi-resolution WSI.
  • H2PE is model-independent (just for the multi-resolution MIL methods) and provides injection schemes for different types of aggregators. Also, its runtime overhead is small, making it practical to adopt.
  • H2PE uses MPP to convert patches from multiple resolution levels into real physical coordinates, so that every level shares a unified space. This lets the resulting hyperbolic distance capture both spatial neighborhood and scale hierarchy in a simple closed form.
Critical weaknesses
  1. All experiments use only three levels (~5×/10×/20×). Under this setting, the distortion produced by Euclidean embedding was small, so the theoretical advantage of hyperbolic space growing exponentially was hardly verified.

  2. H2PE is evaluated only on ZoomMIL, H2MIL, and DSMIL. No hierarchical sequence-modeling encoders like [1,2,3] have been validated. Since recent MIL methods and slide encoders are increasingly built on such architectures, validating the effectiveness of H2PE on these architectures will be of great value and interesting in demonstrating its capability.

[1] "Scaling vision transformers to gigapixel images via hierarchical self-supervised learning." In CVPR.

[2] "Retmil: Retentive multiple instance learning for histopathological whole slide image classification." In MICCAI.

[3] "Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology."In MICCAI.

Questions for authors
  1. Could the authors report results with an additional 0.25 mpp level, and show whether the gap between H2PE and its Euclidean variant grows as the hierarchy gets deeper? This would directly test whether the exponential-growth advantage of hyperbolic space holds in practice.
  2. Does H2PE benefit fine-grained tasks that depend on nuclear-level morphology at high magnification, such as biomarker prediction (e.g., ER/PR/HER2 prediction in BCNB)?
  3. How would H2PE be integrated into hierarchical sequence-modeling encoders, which already have built-in positional mechanisms? Should H2PE replace or complement them?
4 comments
AuthorsOpenReview ↗

cbXU#W1Q1: Evaluation with an Additional Resolution Level

cbXU#W1Q1: Evaluation with an Additional Resolution Level

Following your valuable feedback, we conducted additional experiments incorporating 40x magnification (0.25 MPP).

We would like to clarify the following regarding these results. At 40x magnification, the number of patches is 64 times greater than at 5x, resulting in approximately 64-fold increases in preprocessing time and storage requirements. We initially attempted to train and evaluate MR-MIL models using all 40x patches but encountered GPU out-of-memory (OOM) errors due to the excessive memory demands. We therefore imposed an upper limit on the number of instances in the image pyramid for both training and inference. Consequently, the results may differ slightly from those in the main experiments. However all methods were evaluated under identical conditions to ensure fair comparisons.

  1. The experiments were conducted with progressively deeper MPP hierarchies: (2, 1), (2, 1, 0.5), and (2, 1, 0.5, 0.25). We used the HIPT backbone and CAMELYON16 dataset, reporting ACC/AUC.
Method(2,1)MPPS(2,1,0.5)MPPs(2,1,0.5,0.25)MPPS
H2PE86.8/87.788.8/89.491.1/92.7
Normalized Euclidean86.4/86.488.7/89.391.0/91.9
Euclidean86.7/87.088.1/89.190.6/91.8
None86.7/86.888.5/88.989.8/92.1

At shallow hierarchies (i.e., (2, 1) MPP), performance differences among PE strategies are relatively small. As the hierarchy deepens, all distance exhibit performance improvements. Notably, normalized Euclidean distance, as suggested in Reviewer PaCM's Weakness 2, provides a slight performance advantage over naive Euclidean distance. While deeper hierarchies generally benefit all PE strategies, H2PE achieves the highest performance, demonstrating its effectiveness in deeper multi-resolution hierarchies.

  1. We further evaluated geometric fidelity across hierarchy depths using a synthetic image pyramid with a 4 × 4 top-level grid and seven levels. We define GG as the shortest-path hop distance between two instances at different levels and DD as their distance in the PE space. We randomly sampled 150 source and 400 target instances and computed the distortion ratio D/GD/G for each cross-level pair. To quantify distortion variability, we measured the ratio of the 99th to the 1st percentile of D/GD/G. This metric equals 1 when DD is perfectly proportional to GG.
Method234567
H2PE1.91.81.61.41.31.3
Normalized Euclidean2.33.04.36.39.011.3
Euclidean2.83.65.46.28.09.3

At seven levels, H2PE achieves a distortion ratio of 1.3, indicating that its distance-to-hop ratios vary by only a factor of 1.3 between the 1st and 99th percentiles, closely preserving proportionality between distance and hop count. In contrast, normalized Euclidean distance yields 11.3, indicating substantially greater distortion. As your feedback, these results highlight H2PE's increasing advantage as hierarchy depth grows. We present the corresponding visualization in Fig. 6(b) of the revised manuscript.

  1. Geometric distortion increases with depth. We provide a detailed proof in PaCM#W2,Q2 and kindly ask the reviewer to refer to it.
AuthorsOpenReview ↗

cbXU#W2: H2PE on Hierarchical Sequence Models

cbXU#W2: H2PE on Hierarchical Sequence Models

We appreciate your valuable suggestion, which helped us further improve our work. Following your suggestion, we conducted additional experiments using HIPT and have included the results in Table 1 of the revised manuscript. Since this question is also related to your Question #3, we kindly invite you to refer to our response to that question for further details.

We sincerely appreciate your valuable comments, which provided useful insights for improving our manuscript and designing additional experiments.

AuthorsOpenReview ↗

cbXU#Q2: Nuclear-level Morphology

cbXU#Q2: Nuclear-level Morphology

We sincerely appreciate your suggestion and are pleased to share an interesting finding from our additional experiments.

  • We conducted HER2 prediction on the BCNB dataset using the (1, 0.5) MPP setting and HIPT MIL. As shown in the table below, all PE methods perform within the variation range of the performance without PE. This suggests that positional information may offer limited benefits for this particular task. One possible interpretation is that, for labels such as HER2 status, which are determined by the overall morphology of tumor cells, the spatial arrangement of patches may provide little additional information.
PEACCAUC
none72.6$\pm$2.072.4$\pm$3.1
PPEG72.2$\pm$1.872.2$\pm$2.6
RoPE73.6$\pm$2.671.2$\pm$3.2
TNAM72.3$\pm$2.071.1$\pm$2.2
Sin.PE72.6$\pm$3.171.4$\pm$4.0
HyPE72.5$\pm$3.670.5$\pm$4.3
TreePE72.4$\pm$1.970.8$\pm$3.4
GPSE72.7$\pm$2.371.3$\pm$3.6
H2PE73.1$\pm$1.572.1$\pm$3.1

We believe this finding may provide a useful starting point for future MIL research on more advanced biomarker prediction tasks beyond conventional diagnosis. We also discuss this direction in the Future Work section. We sincerely thank you for this insightful suggestion.

AuthorsOpenReview ↗

cbXU#Q3: Intergrating into Hierarchical Sequence-modeling Encoders

cbXU#Q3: Intergrating into Hierarchical Sequence-modeling Encoders

We appreciate your insightful question, which has helped us further develop our work. Our primary motivation is to provide positional information for MRMIL. As described in Section 3.2.4, H2PE is designed to be flexibly integrated into the general MRMIL architecture (encoder).

  • Accordingly, when positional information is provided by a replaceable module within an architecture, H2PE can replace that module. HIPT, which we evaluated following your suggestion, is an example of this case. HIPT incorporates absolute positional encoding (APE), which we replaced with H2PE and other PE methods for comparison. In contrast, RetMIL and MambaMIL rely on sequence order itself for positional information rather than a separate PE module. In these architectures, H2PE can serve as a complementary component rather than a direct replacement. Specifically, the node term eie_i can be added to the input features, while the pairwise term bijb_{ij} can be incorporated through Equation 6.

  • We sincerely thank you for bringing up this insightful point, which we had not fully considered during our initial investigation. We believe this question provides an opportunity to enrich the manuscript with a broader discussion of H2PE's applicability. Accordingly, we have added a relevant discussion to the Appendix.

Reviewer_PaCMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes H2PE for multi-resolution MIL in WSI analysis. H2PE leverages hyperbolic geometry to jointly model spatial relationships and resolution hierarchies, incorporating asymmetric cross-resolution interactions into different MIL architectures. Experiments on three public datasets across three MIL frameworks demonstrate performance improvements over existing positional encodings, supported by ablation studies and visualization analyses.

Critical strengths
  1. The paper addresses an important and relatively underexplored problem of positional encoding in multi-resolution MIL, with a clear motivation for jointly modeling spatial relationships and resolution hierarchies.
  2. The proposed H2PE is evaluated across three public datasets and three representative MIL architectures, demonstrating its applicability to different aggregation frameworks and generally competitive performance.
Critical weaknesses
  1. As described in Appendix B.1, the authors we use h = 2 for H2PE on ZoomMIL and H2MIL and h = 1 otherwise. This inconsistency may provide H2PE with access to a larger spatial context, making it difficult to attribute the performance gains solely to the proposed positional encoding.
  2. The justification for the necessity of hyperbolic geometry is insufficient. While hyperbolic geometry is a reasonable choice for modeling hierarchical structures, normalizing Euclidean distances by the physical patch sizes corresponding to different magnification levels seems to offer a similar scale-invariant property.
Questions for authors
  1. Why does H2PE use h=2 for ZoomMIL and H2MIL, while other methods use h=1?
  2. Could normalizing Euclidean distances by the physical patch sizes corresponding to different magnification levels serve as an alternative to hyperbolic distance? How would such a baseline perform if all other components were kept unchanged?
  3. For architectures such as ZoomMIL, H2PE introduces an additional relational residual module (Eq. 6), which may itself contribute to the performance improvements. In such case, how much of the improvement comes from H2PE itself?
  4. Some of the improvements reported in Table 1 are relatively marginal. Are these improvements statistically significant?
4 comments
AuthorsOpenReview ↗

PaCM#W1,Q1: Inconsistent Across Models

PaCM#W1,Q1: Inconsistent hh Across Models

We sincerely apologize for the confusion caused by this inconsistency. We fully acknowledge the discrepancy between the value of hh used in H2PE and those used in the other baselines. Following your observation, we carefully inspected our implementation and identified the following issue.

  • Although H2PE was configured with h=2h=2, a coding error caused it to effectively operate with h=1h=1. Consequently, all PEs were compared using the same effective value of hh, despite the different configurations.

  • Specifically, lines 66–89 of interaction.py implement an algorithm that identifies node pairs separated by hh hops in the BHS graph. Our intention was to include pairs such as (A,C) in the expanded set when the graph contains a path A→B→CA \rightarrow B \rightarrow C and h=2h=2. However, the current implementation uses expanded.add(_undirected(u, v)), which adds the pair consisting of the currently explored node and its immediate neighbor. Thus, it adds (B,C) instead of the intended pair (A,C). Since (B,C) is already a one-hop pair, it is already captured when h=1h=1. To implement the intended behavior, the starting node AA (represented by start in the code) should have been used, as in expanded.add(_undirected(start, v)). As a result, the implementation failed to capture relationships with h>1h>1 and only incurred additional BFS computation without expanding the intended set of pairs.

  • Consequently, the previously reported results for H2PE are equivalent to those obtained with h=1h=1. Your insightful observation prompted us to take a closer look at the fairness of the comparison and, importantly, helped us identify this implementation error. We are grateful for this opportunity to correct our implementation and sincerely thank you again for bringing this issue to our attention.

AuthorsOpenReview ↗

PaCM#W2,Q2: The Necessity of Hyperbolic Geometry

PaCM#W2,Q2: The Necessity of Hyperbolic Geometry

Thank you for this insightful observation. We clarify the relationship between the normalized Euclidean distance you suggested and the hyperbolic geometry adopted in our method as follows.

  1. Relationship Between the Two Distances

The normalized Euclidean distance is defined as rij=∣∣ci−cj∣∣/(Pmi⋅mj)r_{ij}=||c_i-c_j||/(P\sqrt{{m_i}\cdot{m_j}}). Given the squared Euclidean distance ∣∣ci−cj∣∣2||c_i-c_j||^2 in the numerator and 2mimj2m_im_j in the denominator of Equation 3, the hyperbolic distance can be expressed as:

dij=arcosh(1+rij2/2)=2×asinh(rij/2)d_{ij}=arcosh(1+r_{ij}^2/2)=2\times{asinh}(r_{ij}/2)

In other words, the hyperbolic distance is a monotonic transformation of the normalized Euclidean distance. Specifically, dij≈rijd_{ij} \approx r_{ij} for small distances, whereas dij≈2ln⁡rijd_ {ij}\approx 2\ln{r_{ij}} for large distances.

  1. Differences from Hyperbolic Geometry

The normalized Euclidean distance rijr_{ij} does not define a metric. For distant pairs of patches, it does not correspond to the length of any path and therefore fails to satisfy the following properties.

  • (a) Triangle inequality: Since dijd_{ij} is a Riemannian distance, no path from ii to jj through an intermediate point kk can be shorter than the direct distance from ii to jj. In contrast, the detour distance rik+rkjr_{ik}+r_{kj} can be smaller than rijr_{ij}. Consider the following three patches:
PatchNotec(μm)c({\mu}m)mm (MPP)z=P×mz = P{\times}m (μm)(\mu{m})
ifinest patch(0, 0)0.25256 x 0.25 = 64
jfinest patch, Eight grid units to the right of ii(8 x 64, 0) = (512, 0)0.2564
kcoarse patch, Directly above the midpoint between ii and jj(256, 0)1.0256 x 1.0 = 256

The distances are given as follows:

Pairrr}dd
(i,j)8$\approx$4.189
(i,k)2.5$\approx$2.095
(k,j)2.5$\approx$2.095

For rr, the detour path i→k→ji \rightarrow k \rightarrow j has a total length of 2.5+2.5=52.5+2.5=5, which is shorter than the direct distance rij=8r_{ij}=8.

  • (c) Depth-dependent Distortion: Let GijG_{ij} denote the minimum number of pyramid steps (grid or parent–child edges) between two finest patches ii and jj with mi=mj=mm_i = m_j = m. Since the shortest route climbs and descends the pyramid, Gij≈2log⁡2rijG_{ij} \approx 2\log_2 r_{ij} where rij=∥ci−cj∥/(Pm)r_{ij} = \|c_i - c_j\|/(Pm). From Part 1, dij≈2ln⁡rij=ln⁡2⋅Gijd_{ij} \approx 2\ln r_{ij} = \ln 2 \cdot G_{ij}, so dijd_{ij} is proportional to the step count, whereas rij=2Gij/2r_{ij} = 2^{G_{ij}/2} grows exponentially in it. Deeper pyramids admit larger GijG_{ij}, so the distortion of rijr_{ij} grows with depth while that of dijd_{ij} does not.

Therefore, we adopt the hyperbolic UHS to turn the magnification-normalized distance into a consistent metric across resolution levels. We also invite you to refer to our responses to wxFH#W4 and cbXU#W1 for related discussions and impact of further magnitude level on performance.

AuthorsOpenReview ↗

PaCM#Q3: ZoomMIL Performance Improvements and Equation 6

PaCM#Q3: ZoomMIL Performance Improvements and Equation 6

Thank you for your valuable feedback. We understand your concern to be that ZoomMIL incorporates additional parameters through Equation 6, and that their effects should be distinguished from those of the positional information introduced by H2PE. To investigate this, we conducted additional experiments using BRACS-F.

ConditionRelational Residual ModuleLocation Info.ACCAUC
ZoomMIL as is--43.780.8
+ Equ. 6, bij=0b_{ij}=0 (i.e., uniform softmax weights)v-45.580.6
+ H2PEvv47.581.7
  • When bijb_{ij}, which incorporates positional information, is fixed to zero, the softmax assigns a uniform weight of 1/N1/N to each of the NN components. This allows us to isolate the effect of WW.

  • The results show that applying H2PE in its entirety leads to improved performance. The other two settings, viewed as ablations, indicate that removing either component negatively affects ZoomMIL's WSI-level analysis.

We sincerely thank the reviewer for this insightful suggestion, which motivated these additional experiments.

AuthorsOpenReview ↗

PaCM#Q4: Statistical Significance

PaCM#Q4: Statistical Significance

  • We sincerely appreciate your valuable comment. We have addressed the same concern in our response to Reviewer wxFH's Weakness #1. We kindly ask you to refer to that response for further details.
Reviewer_wxFHOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

H2PE proposes a novel positional embedding based on hyperbolic geometry for MRMIL architectures, which uniformly models spatial location, physical scale, and directional relationships between multi-resolution patches. In comparison with 7 positional encoding approaches across 3 MRMIL architectures and 3 WSI datasets, H2PE provides accurate, efficient, and stable performance across different resolutions.

Critical strengths
  • The idea is a well-motivated geometric formulation.
  • Different adaptations for different backbone architectures are clearly explained, and H2PE consistently provides the best performance across the examined architectures (except in 3 out of 12 cases).
  • Well-organized evaluations: AUC/ACC table, quantitative and qualitative reports on ROI explanations, data/inference efficiency validation, and ablation study.
Critical weaknesses
  1. The significance of the classification improvement is not statistically investigated.
  2. The evaluated MIL backbones date back to 2021–2022. More recent architectures, such as DAS-MIL and CS-MIL (2024), could be considered. Furthermore, the categorization overlooks Cross-Scale Attention-based MRMIL architectures, which are particularly relevant for evaluating H2PE.
  3. Figure 4, while informative, needs further clarification. It has no legend to explain the colors in the visualization. Why do the red points disappear in (2, 0.5)? What does the last column above the label (f) H2PE represent?
  4. In the introduction, topology is used instead of geometry, although their definitions are not the same. For example, in the Introduction, relative spatial relationships are described as “topological,” while Section 3.2.2 models metric distances in hyperbolic space. The claim of preserving spatial topology in the Conclusion lacks explicit justification or evaluation.
  5. The released code raises reproducibility concerns. In particular, the DSMIL pipeline assigns all concatenated multi-resolution tokens to the finest resolution, eliminating the cross-resolution directional component of H2PE. Also, the relational projection layer seems to be instantiated after optimizer initialization, potentially preventing its parameters from being trained. Could the authors clarify these discrepancies?
Questions for authors
  1. Could you add quantitative values to Figure 3b? The current visualization and axes are not clear enough to compare H2PE with SinPE, TNAM, and PPEG.
  2. Although H2PE is formulated to support arbitrary magnifications, the experiments are limited to dyadically spaced MPP levels (2, 1, 0.5). Have the authors evaluated non-dyadic resolution combinations to demonstrate the benefit of continuous scale modeling over discrete hierarchical positional encodings?
  3. Pls refer to weaknesses
7 comments
AuthorsOpenReview ↗

wxFH#W1: Statistical Significance

wxFH#W1: Statistical Significance

We sincerely appreciate your valuable comment on statistical significance. Before conducting the statistical analysis, we identified an implementation error thanks to your feedback in Weakness 5 and reran the affected experiments. Your thoughtful observation helped us refine our implementation and report more reliable and objective performance results.

  • We employed two-sided paired tests and added †^\dagger p<0.05 and ‡^\ddagger p<0.01 to Table 1 of the revised manuscript, with corresponding explanations in the caption. We also report the results for the two MIL methods suggested in your Weakness 2 in the Appendix, following the same evaluation protocol.

  • We clarify that the statistically significant improvements over the second-best methods in Table 1 are observed for CAMELYON16 with HIPT and BRACS-C with ZoomMIL. We interpret these results in light of the nature of the diagnostic tasks. In these settings, MIL primarily operates as an OR-gate, determining whether a bag contains instances associated with specific pathological features. Thus, rather than introducing an entirely new source of diagnostic information, H2PE enables models to better exploit structural information that conventional approaches do not fully utilize, as further supported by our subsequent experiments. Nevertheless, its consistent superiority or comparable performance across other settings indicates that H2PE provides useful information for diagnostic prediction.

  • More importantly, while predictive performance remains an important consideration, our central research question concerns preserving structural relationships in spatial hierarchies that expand exponentially with increasing magnification. This property is extensively investigated through multiple experiments in Sec. 4.3 of the revised manuscript. We sincerely ask the reviewer to consider these findings when evaluating the contribution of our work.

We have explicitly acknowledged these points in the Limitations section.

We sincerely thank you once again for your insightful feedback. All revisions are highlighted in red in the revised manuscript.

AuthorsOpenReview ↗

wxFH#W2: Adding Recent MIL Architectures

wxFH#W2: Adding Recent MIL Architectures

  • We acknowledge that the MIL backbones used in our experiments are not the most recent architectures, and we agree with your suggestion. We conducted additional experiments using DAS-MIL and CS-MIL (2024), which adopt a Cross-Scale Attention mechanism.

  • We would like to clarify our rationale for selecting the three MIL backbones (i.e., ZoomMIL, H2MIL, and DSMIL) in the original experiments. Our research focuses on developing PE methods suitable for multi-resolution WSIs, with the goal of comparing competing PE methods using well-established MIL architectures. Accordingly, for each MRMIL variant, we selected a representative architecture with a distinct aggregation mechanism that has been validated across multiple studies.

  • At the same time, we acknowledge that our original experiments did not sufficiently investigate Cross-Scale Attention as an aggregation mechanism. We sincerely appreciate you bringing this aspect to our attention. Due to space constraints, we have included the corresponding quantitative results in the Appendix.

AuthorsOpenReview ↗

wxFH#W3: Improvements to Figure 4

wxFH#W3: Improvements to Figure 4

We fully agree with your comment and have revised Figure 4 to make the visualization more informative and easier to interpret.

  1. Legend to Explain the Colors
  • We have added a legend to the visualization in Figure 4 and incorporated it into the revised manuscript. Specifically, the colors from blue to red represent a continuous gradient on a jet colormap, with normalized values ranging from [0, 1].

  • In Figure 4, we use an arbitary instance at 2 MPP as the reference point. For each PE method, instances that are closer to the reference point in the corresponding PE are shown in red, whereas those farther away are shown in blue. For SinPE and RoPE, cosine similarity is used to measure the similarity to the reference instance.

  • Figure 4 visualizes how well different PE methods represent the exponentially expanding zoom-in space of a WSI. However, we apologize that our previous description was difficult to follow and did not adequately facilitate understanding of the figure. Following your suggestion, we have restructured the explanation and incorporated the revised description into the manuscript.

  1. Missing Red Points in (2, 0.5)
  • This observation demonstrates that H2PE models not only hop-based relationships but also the geometric relationships among instances.

  • (d) TreePE and (e) GPSE represent hierarchical relationships among instances using graph structures. However, they treat the relationships between (2, 1) MPP and (2, 0.5) MPP identically, despite their distinct pathological relationships. This illustrates a limitation of graph-based PEs that rely solely on hop-based relationships. In contrast, H2PE preserves not only hop information but also the spatial information encoded in the hyperbolic space, allowing it to better capture relationships between instances that are farther apart in the embedding space, even when an intermediate MPP is removed. Specifically, the absence of red points at (2, 0.5) indicates that H2PE considers the instances at 2 MPP and 0.5 MPP to be relatively distant, despite their one-hop relationship. Meanwhile, the red points at (2, 1) indicate that H2PE considers those instances relatively close.

  1. Last Column above the Label (f) H2PE
  • We acknowledge that the incomplete layout made the figure difficult to interpret. Our original intention was to examine multiple MPP combinations for (f) H2PE, just as we did for (d) and (e). To make this clearer, we have rearranged the panels of (f) in the same layout as (d) and (e) to improve readability.

We sincerely appreciate the opportunity to further improve our manuscript through your valuable feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

We sincerely thank all reviewers, chairs, and everyone involved

We sincerely thank all three reviewers for taking the time to provide constructive feedback on our work. We are also grateful to the Chairs and everyone involved for giving us the opportunity to further improve our research and manuscript through this review process. We particularly appreciate wxFH for the well-motivated geometric formulation, PaCM for highlighting the clear motivation, and cbXU for identifying an unexplored research gap. We have made every effort to address the reviewers' suggestions and concerns with clear and thorough responses. Detailed replies to all individual comments are provided, and we will upload the revised manuscript before the rebuttal deadline.

Once again, we sincerely appreciate your constructive feedback.

Authors of Submission 13792

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.