acceptodds
Withdrawn from ICLR 2027

Explaining Hyperbolic Neural Networks via Geometry-Aware Relevance Propagation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Hyperbolic neural networks introduce geometric operations that require explicit treatment in relevance propagation. Equivalent geometric realizations can produce different feature attributions, even when local relevance is conserved. We study this problem through Geometric Representation Invariance (GRI), a specialization of Implementation Invariance, and zero-curvature consistency, which requires identity relevance propagation when a geometric module approaches the identity. We propose LRP-radial-all for origin-centered radial modules, treating geometric scaling as modulation and assigning relevance entirely to the signal branch. The rule conserves relevance, is invariant to equivalent radial factorizations, and satisfies zero-curvature consistency, yielding GRI for a specified Poincaré–Lorentz logarithmic-map construction. In contrast, a conservative LRP-half baseline can violate both consistency criteria. Experiments on hyperbolic MNIST, sEEG, and CIFAR-10 classifiers assess attribution fidelity, qualitative explanations, and runtime. LRP-radial-all achieves competitive attribution fidelity across datasets with runtime comparable to GradientInput and substantially lower than Integrated Gradients. These findings motivate geometry-aware propagation rules that distinguish relevance conservation from consistency across equivalent computations.

Primary area:
interpretability and explainable AI
Keywords:
Hyperbolic Neural Networks, Explainable AI, Layer-wise Relevance Propagation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AKv3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies how to explain the predictions of hyperbolic neural networks using layer-wise relevance propagation, which assigns relevance scores to input features. The main difficulty in explaining hyperbolic neural networks is that hyperbolic operations are not isolated coordinate-wise, and different models of hyperbolic spaces can result in different explanations despite making the same predictions. To address this, the paper proposes LRP-radial-all, which separates hyperbolic operations at the origin (e.g. exp and log maps) into a geometric scaling component and an input feature component, and assigns all relevance to the input features instead of the geometric scaling factors. They introduce two criteria, geometric representation invariance and zero-curvature consistency. They then theoretically show that their method satisfies both while conserving relevance, which would not be the case if relevance are also assigned to the geometric scaling factor. They evaluate the method on a hyperbolic MLP for MNIST, a hyperbolic EEGNet variant for sEEG, and a Lorentz CNN for CIFAR-10, using feature removal and retention tests to measure attribution fidelity. The evaluation results for LRP-radial-all is largely mixed. While it achieves better average in some datasets/metrics than the baseline in Table 1, LRP-radial-all underperforms in 2 out of the 6 cases, and the improvements are also usually within confidence intervals (F+ and F- for MNIST, F+ for sEEG and CIFAR-10). In terms of runtime, LRP-radial-all is comparable to Gradient×Input and significantly faster than Integrated Gradients.

Critical strengths
  1. The paper identifies an important issue with explaining hyperbolic neural networks, which is that equivalent geometric representations can produce different relevance scores despite making the same predictions. The authors formalize this through geometric representation invariance and zero-curvature consistency, and provide theoretical results showing that their proposed method satisfies both for exponential and logarithmic maps at the origin. These results provides theoretical motivation to the formulation of the method that are specific to hyperbolic geometry.

  2. The authors evaluate their method across three different datasets and hyperbolic architectures, in both image and time-series classification tasks. While the results are mixed, the method achieves competitive performance and runs substantially faster than Integrated Gradients.

Critical weaknesses
  1. The biggest weakness is that the proposed framework focuses mainly on exponential and logarithmic maps at the origin, but many recent works have proposed fully hyperbolic operations or more complicated operations that do not rely on exp/log maps at the origin. For example, for the Lorentz space, recent works have proposed fully hyperbolic linear layers [a], activation layers [b], residual connections [c], and MLR layers [b]; for the Poincare Ball space, recent works have proposed fully connected layers and MLR layers [d]. It is unclear whether the paper's proposed criteria, such as GRI and zero-curvature consistency, extend to these operations. In fact, it seems to me that these operations do not necessarily converge to their Euclidean counterparts as the curvature approaches zero under the paper's formulation. As recent hyperbolic works have attempted to move away from repeated mappings to and from the tangent space, it is important to consider whether the proposed explainability framework applies to these newer hyperbolic operations. Additionally, I think empirical experiments should also include model architectures that use these operations, such as fully hyperbolic MLR layers for image classification [b], on top of operations on the tangent space. Finally, since the paper's main theoretical guarantees focus on radial operations at the origin, but uses operations that do not rely on exp/log maps in the experiments (e.g. Lorentz linear layers in the CNN experiments), The paper should better justify how it treats different geometric quantities during relevance propagation. For example, the authors assign no relevance to the norm-dependent geometric scaling factors in exp/log maps, but in their Lorentz CNN, they allow relevance to pass through the norm-dependent time coordinate and redistribute it to the spatial features. It's unclear what motivates the discrepancy (why should the time-component receive relevance and are GRI and zero-curvature consistency still satisfied in this case?)

[a] Hypformer: Exploring Efficient Transformer Fully in Hyperbolic Space, Menglin Yang, Harshit Verma, Delvin Ce Zhang, Jiahong Liu, Irwin King, Rex Ying

[b] Fully Hyperbolic Convolutional Neural Networks for Computer Vision, Ahmad Bdeir, Kristian Schwethelm, Niels Landwehr

[c] Lorentzian Residual Neural Networks, Neil He, Menglin Yang, Rex Ying

[d] Hyperbolic Neural Networks++, Ryohei Shimizu, Yusuke Mukuta, Tatsuya Harada

  1. The experimental results show very marginal improvements over existing methods, where all but 1 case show either improvements within confidence intervals or worse performance for LRP-radial-all. These results do not demonstrate that LRP-radial-all gives better explanation than the simple Euclidean counterparts. On CIFAR-10, the proposed method even performs worse than the random baseline on F+. Additionally, the paper only validates GRI through a simple controlled example using equivalent logarithmic map implementations, instead of on full trained hyperbolic networks. Given that geometric consistency is one of the main motivations of the formulation behind LRP-radial-all, demonstrating whether LRP-radial-all actually satisfies GRI on the actual models tested in the experiments is important empirical validation. Currently, stronger empirical evidence is needed to show that these theoretical properties lead to better explanations in practice.
Questions for authors

See weaknesses

Reviewer_nRLvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies feature attribution for origin-centered hyperbolic neural network operations, focusing on Poincaré and Lorentz exponential, logarithmic, and coordinate-conversion maps. It introduces Geometric Representation Invariance (GRI) and zero-curvature consistency as criteria for relevance propagation, proposes the LRP-radial-all rule that routes relevance entirely through the vector signal branch of radial maps, and evaluates it on MNIST, sEEG, and CIFAR-10 classifiers.

Critical strengths
  1. The paper identifies a meaningful issue for propagation-based explainers: local relevance conservation does not by itself imply invariant input explanations under equivalent geometric realizations. The controlled Poincaré-to-Lorentz construction in Figure 1 and Equations (4)-(5) make this issue concrete rather than merely asserting that hyperbolic models are difficult to explain.

  2. The LRP-half counterexample is well developed. Equation (9) and the fuller derivation in Equations (54)-(59) show exactly why applying a half-splitting rule once versus twice changes coordinate-wise relevance while preserving the signed total. Table 3 provides a numerically consistent illustration: radial-all produces matching direct and Lorentz-path relevances, whereas LRP-half does not.

  3. Figure 2 is effective in connecting the theoretical argument to the proposed propagation rule. It explicitly depicts the logarithmic/exponential maps as scalar radial factors times a vector signal, then contrasts the proposed all-to-signal routing with the half allocation. This makes the otherwise abstract distinction in Equations (7)-(9) accessible.

Critical weaknesses
  1. The central prescription is an attribution convention rather than a derivation from the hyperbolic operation, and the paper does not empirically justify this convention beyond the proposed axioms. The proposed rule in Definition 3, Equation (7) sets Rα=0R_α=0 and Rs=RyR_s=R_y, despite α(x)α(x) being a nonconstant function of the same input vector x. This is a defensible convention, analogous to treating a gate as modulation, but it explicitly discards any direct relevance pathway through the radius-dependent geometric factor. In a hyperbolic logarithmic map, the radial term is precisely where curvature changes the computation. Thus, the method can be described as implementation-consistent under the selected signal path, but it is not shown that this is the most faithful explanation of how the geometric operation affects the output.

  2. The GRI guarantee is substantially narrower than the general framing suggests. Definition 1 defines GRI for functionally equivalent Poincaré and Lorentz models, but the proof establishes equality only for the carefully aligned logarithmic-map construction in Equations (31)-(40), with a shared encoder and downstream computation, and with the Lorentz time coordinate treated as a dependent quantity. The analogous exponential-map argument also requires a matched time-coordinate relevance convention, as acknowledged in Section 6.

  3. The practical conservation claim is local, while the reported explanation pipelines need not be globally conservative. The radial rule conserves signed relevance exactly, but the actual experimental pipelines use stabilized LRP rules. Appendix H explicitly derives the nonzero conservation residual for LRP-ϵ in Equations (72)-(73). Appendix K further notes unassigned relevance for zero patches and additional residuals from affine biases and stabilization in the CIFAR-10 model. Thus, “the rule conserves relevance” is correct locally for the radial module, but readers could easily infer end-to-end conservation for the reported explanations.

Questions for authors
  1. Can the authors provide an end-to-end GRI experiment on a nontrivial trained architecture, rather than only the isolated log-map construction of Figure 1 and Table 3? In particular, can a Poincaré model with Möbius biases be converted into a matched Lorentz realization and tested for equality of input explanations under radial-all?

  2. What evidence supports treating α(x) as purely modulatory in Equation (7) when it is itself a function of the explained input x? Can the authors provide controlled examples where the prediction depends predominantly on radius and compare radial-all against finite radial perturbations or an exact differentiation-based analysis?

  3. Please report end-to-end signed conservation residuals for each experimental model and attribution method. How large are residuals caused by the ϵϵ-rules, biases, zero patches, and the time-coordinate redistribution described in Appendix K?

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

The paper proposes LRP-radial-all for propagating relevance through origin-centered hyperbolic operations. It formalizes Geometric Representation Invariance and zero-curvature consistency, establishes guarantees for specified module constructions, and shows that relevance conservation alone does not ensure these properties. Validation includes controlled consistency tests and attribution-fidelity, qualitative, and runtime comparisons on MNIST, sEEG, and CIFAR-10. The method achieves competitive fidelity and lower runtime than Integrated Gradients, although its relative fidelity performance varies across datasets and metrics.

Critical strengths
  • The analysis is sound within its stated scope and clearly distinguishes relevance conservation from consistency across equivalent geometric computations. This distinction provides a principled justification for the proposed propagation rule.
  • Explaining the internal geometric operations of hyperbolic neural networks addresses an important, comparatively underexplored problem in XAI. The precise formulation and methodological derivation constitute a meaningful contribution.
Critical weaknesses
  • MNIST and CIFAR-10 are limited in scale and, in my view, may be no longer sufficiently compelling as the primary benchmarks for this work. Evaluation on more demanding hierarchical classification tasks, where hyperbolic representations are directly motivated, would better establish the practical relevance and credibility of the empirical findings. The sEEG application is comparatively more convincing.
  • The results do not establish a consistent advantage over existing attribution methods. Performance on MNIST is largely comparable to Integrated Gradients, the differences on sEEG are small relative to the reported uncertainty, and the CIFAR-10 results are mixed. These findings leave the practical benefit of the proposed method insufficiently established.
  • Prior work on ROAR, ROAD, and GOAR has identified limitations of perturbation-based fidelity evaluation and proposed approaches to address them. These concerns have already received substantial attention within the XAI community. Complementary evaluation using metrics or protocols that explicitly address these limitations is therefore needed to provide more convincing evidence of explanation quality.

References

  • [ROAR] Hooker, Sara, et al. "A benchmark for interpretability methods in deep neural networks." Advances in neural information processing systems 32 (2019).
  • [ROAD] Rong, Yao, et al. "A consistent and efficient evaluation strategy for attribution methods." arXiv preprint arXiv:2202.00449 (2022).
  • [GOAR] Park, Yong-Hyun, et al. "Geometric remove-and-retrain (GOAR): Coordinate-invariant explainable AI assessment." arXiv preprint arXiv:2407.12401 (2024).
Questions for authors
  • Do paired comparisons support reliable fidelity improvements over Integrated Gradients and Gradient×Input, particularly on sEEG? Please provide confidence intervals for the differences.
  • Can the authors provide results on a more demanding task with an explicit hierarchy that demonstrate the practical value of the proposed attribution method?

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

Related papers

Discussion (0)

Sign in to comment.