acceptodds
Under review as a conference paper at ICLR 2027

Learning Directional Randers Geometry for Latent-Space Exploration

14 views

Abstract

Deep neural networks extract powerful latent representations from images, enabling fast and efficient downstream analysis. However, these latent spaces lack explicit geometric structure, leading to possibly poor interpolation fidelity and misleading extrapolation. Latent interpolation techniques assume symmetric transition costs which are often not the case for real world data. Their symmetric nature cannot capture directional processes, such as disease progression, where certain transitions are impossible (*e.g.* tumor growth dynamics or cognitive improvement in Alzheimer's disease). Although Riemannian metrics extend Euclidean geometry by incorporating curvature, they remain symmetric. We show that understanding asymmetry and not curvature alone directly addresses these problems. To do so we propose a new Finsler-Randers metric learning framework. It incorporates direction-dependent distances, thus enabling the computation of trajectories respecting monotonic sequential constraints. We demonstrate its potential for interpolation and extrapolation on controlled cross-sectional datasets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Finsler geometry, Randers metric, Riemannian geometry, metric learning, latent space, interpolation, extrapolation, monotonicity, sequential constraints, geodesic shooting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UQcvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper augments a Riemannian latent metric with a learned directional term to construct a Randers geometry for interpolation and extrapolation. The direction is learned from an ordered scalar attribute, and the method is evaluated on synthetic phantoms and plant images. The main concern is that the experiments do not isolate the contribution of the geometry or sufficiently establish the quality of the resulting extrapolations.

Critical strengths
  1. The construction provides a modular way to incorporate local ordering information into existing latent space metrics.
  2. Comparisons across several base metrics and ablations of label noise and sampling choices help characterize the behavior of the framework.
Critical weaknesses
  1. The main comparison changes both the metric and the directional sampling procedure. Randers extrapolation repeatedly samples around the learned direction, while the symmetric baselines do not receive the same guidance. The Wasserstein comparison uses a different trajectory construction, leaving the benefit of Randers geometry over an equally guided symmetric method unresolved.
  2. The extrapolation metrics measure monotonicity and change in the supervised attribute, while quantitative image quality is evaluated only for interpolation. The qualitative examples provide some support, but the plausibility of extrapolated images has not been independently quantified at comparable amounts of scalar progression. The practical significance of the reported gains therefore remains uncertain.
  3. The conversion from the learned direction to the asymmetric cost is not consistently specified: the ordering loss uses the base metric in its pairing, while the asymmetric cost does not. The dual expression also needs reconciliation with the stated pairing convention. These discrepancies should be resolved against the implementation so that the dynamics used to produce the reported results are unambiguous.
  4. The main tables do not clearly define what the reported variability summarizes. Several extrapolation results show substantial dispersion in the magnitude of scalar change, but the analysis does not establish how representative the mean changes are of typical trajectories. This limits conclusions about how consistently the method provides useful exploration.
  5. The results tables extend beyond the text width, and their numbered citations are difficult to reconcile with the unnumbered bibliography. These inconsistencies reduce readability and make the references for the compared methods harder to trace.
Questions for authors
  1. How does Randers extrapolation compare with a symmetric method using the same learned direction and repeated sampling procedure?
  2. How does the plausibility of extrapolated images compare when methods achieve similar amounts of scalar progression?
  3. Which formulations of the ordering loss, asymmetric cost, and dual norm are implemented, and how are they made consistent?
  4. What does each reported ± value summarize, and what accounts for the large variation in extrapolation magnitude?
4 comments
  1. We added an additional setup where the sampling is done using the learned ω\omega from the Randers metric, similarly to the Finsler framework, but the Hamiltonian dynamic integrated is the Riemannian one (Appendix I). With refresh at every step, results are on par with the Finslerian setups: the directional sampler seems to account for most of the gain over the unguided symmetric baselines. Without refresh, i.e. ζ=1\zeta=1, we observe that the trajectories similar if initialized with the same speed, as expected from Proposition 1. The dataset presented in this work appears to be in this potential-like regime. The advantage of the Finsler metric is expected to appear for non-gradient progression which we leave to future work. Still, our work's aim is exactly to propose to use Finsler geometry to encompass these ideas of directional sampling on Riemannian manifolds. We argue that the Finsler metric encodes the directional information in the geometry rather than in an external sampling rule and hence allows to work with a more self-contained geometrical framework.
  2. As pointed out by another reviewer, we added image quality metrics to the extrapolated results. They show that, as expected per the few extrapolation examples initially displayed in Appendix K, all the methods produce meaningful images with FWD in the same range as in the interpolation setup.
  3. These issues have also been pointed out by another reviewer. We fixed the inconsistencies in the paper and clarified the conventions used. The code already used the correct conventions. The only significant change is that we removed the metric GG in the hinge loss, making it both correctly typed and coordinate invariant (proof in Appendix C). Retraining ω\omega was required and all Randers results were rerun (extrapolation hyper-parameters ablation is still running).
  4. Tables now report the median and IQR (25th–75th percentile) of the Kendall tau and of the relative variation over N trajectories, plus image-quality metrics (SheppLogan : 250 interpolation/ 500 extrapolation ; KOMATSUNA : 90 interpolation / 180 extrapolation) with 25 time-points. Relative variation depends strongly on the initial value of the sequential factor, which explains much of the dispersion (Fig. 3). Table layout now fits the text width.

We hope these additions and corrections resolve the reviewer’s concerns about the sampling controls and the reporting, and we remain available for any further clarification.

Thank you for your response.

My main concerns regarding the distinct contribution of the asymmetric geometry and the consistency of the mathematical formulation still require verification through the revised experiments and manuscript.

Unfortunately, I am also unable to open the revised PDF, as noted by another reviewer. Could you please re-upload the manuscript so that I can properly assess the additional results and revisions?

I will reconsider my assessment once the revised version is accessible.

Dear reviewer,

Sorry for the previously corrupted file. We have reupdated the file. It seems to load fine now.

Note that the experiments of appendix H.3 are still running as we had to retrain ω\omega. They will be ready for the camera ready version. We expect the conclusions that we drew previously to still hold.

Best,

I thank the authors for their thoughtful rebuttal and clarifications. After considering the points raised in the response, my final assessment remains unchanged, so I will maintain my original score.

Reviewer_Z2oqOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates interpolation and extrapolation tasks in the latent space of deep neural networks. Existing Euclidean and Riemannian metrics for these tasks remain symmetric, thus not adequate to capture the asymmetry of directional processes. The Authors provide a Finsler-Randers metric learning framework that, in addition to a Riemannian metric GG, considers a direction-dependent term ω\omega. The proposed method is tested on both the interpolation and extrapolation tasks on the KOMATSUNA and Shepp-Logan datasets.

Critical strengths

The research question is clear, the proposed method is sound and the provided results on the extrapolation task are convincing in terms of monotonicity. Furthermore, although I am not familiar with the topic, the paper remains accessible. These reasons lead me to make a positive recommendation.

Critical weaknesses

While the gain on the extrapolation task seems well supported in terms of monotonicity; the same does not hold for the interpolation task. The Authors acknowledge this result in lines 416-426. Hence, the claimed potential for the interpolation task (e.g., lines 22-23) does not appear to me convincingly supported by Tabs. 1 and 2. I would ask the Authors either to temper their claims, providing a more accurate description of their findings, or to provide an experiment where the interpolation gain is statistically significant. At the current stage, the potential impact of the framework prevents me from recommending clear acceptance.

Update after the discussion

Regarding the critical weakness above, the Authors acknowledged it and tempered their claims accordingly in the revised version. The FWD metric I requested for the extrapolation task raised concerns about the potential benefits of the proposed method in this setting as well, thus weakening the critical strength pointed out in my initial assessment. This led me to reconsider my initial score and to lower it from 3 to 2.

Questions for authors

Please see the “Critical Weaknesses” box. In addition, below are minor concerns/comments I would ask the Authors to address.

  • Could the Authors provide a cheap validation of the approximation used in Eq.(6)? For instance, just using pairs of encoded test images (which provide the endpoints of the trajectories), return the ratio between the pullback length of the PBCentroids geodesic and that of the linear interpolation between the same endpoints. Both are computable with forward passes only as the length of the decoded curve. Ideally, the same measure should be applied to the geodesics of LAND, LAND RBF, and CovCentroids. Relatedly, the pullback metrics WarpedPB, RBF PB, and PBIG described in App.C1 are not reported in the empirical assessment as they do not fit in memory. Could the Authors include them in a reduced setting? This would also help in validating PBCentroids against the exact pullback.
  • “we consider the region orientable if it is above 0.5” (line 264): What is the rationale behind this choice? Are the Authors implicitly assuming that the neighbors of each point are balanced between preceding and subsequent ones, so that 0.5 corresponds to chance level? In this case, have the Authors considered a measure that accounts for label imbalance?
  • Proposition 2 is an application of a known result, Eq.(4.4) in Banerjee et al (2005), I would ask the Authors to clearly state it in the main body as well, not only in App.G.
  • “Then we can recover the unknown sequential factors stγs_t^\gamma ” (line 337): How are these factors recovered?
  • Tabs. 1 and 2: Kendall-tau is bounded in [−1,1][-1,1], since several entries extend beyond that range, I guess the ±\pm indicates the standard deviation. Please clarify and indicate the number of trajectories over which it is computed. Relatedly, since the distributions of the Kendall Tau may be highly skewed or even bimodal (as suggested by the non-Finsler extrapolation results, e.g., 0±0.80 \pm 0.8 for LAND RBF), mean ±\pm std seems to me to be a poor summary. Median with IQR would be more informative. Furthermore, it seems to me that no method dominates on both Kendall Tau and FWD. A scatterplot Kendall Tau vs FWD would help the reader in assessing this apparent trade-off between monotonicity and realism of the generated images, and the impact of the Finsler component. Could the Authors provide FWD within a similar scatter plot also for the extrapolation task? Finally, it would be useful to explain why interpolation results are missing for the Wasserstein baseline.
  • Typos: in line 339, sγs^\gamma has length TT, the range T+1T+1; in line 244 “.” right after “sequential dynamic”.
4 comments

We thank the reviewer for the comments.

Regarding the interpolation performances, Proposition 1 yields the expected behavior: when the learned ω\omega is (close to) exact, ω≈ds\omega\approx\mathrm{d}s, Randers and Riemannian geodesics between fixed endpoints follow the same paths and differ only in length, by ≈s1−s0\approx s_1 - s_0. Tables 1–2 are consistent with this, since we see no systematic monotonicity gain for interpolation. We agree that our original wording overstated the interpolation potential, so we revised the abstract and the discussion (lines 416–426). Interpolation is now presented as a consistency check, showing that the directional term does not degrade the base geometry. The asymmetric distance remains informative because it encodes the directional bias, but we no longer claim better interpolation paths. The main empirical benefit of the framework is for extrapolation, where there is no fixed endpoint to constrain the path.

We compare the lengths on 10 geodesics computed using the PBCentroids metric with the true vanilla pullback metric. The MAPE ( (dst_pb - dst_pbcentroids).abs()/dst_pb ) is of 0.96+-0.01 (mean/std). This massive discrepancy is both expected and desirable. Indeed the vanilla pullback metric is very badly calibrated to respect the data manifold support. This finding led to the introduction of the method that we reintroduce in Appendix D. Therefore the actual geometry induced by our approximation, that tailors to the manifold support, is quite different from the original pullback one : hence the differences in distance. As to compare the length with other metrics, it will not provide any sensible information. The geometry induced by each metric is different and consequently there is no reliable way to compare the scale of the lengths in a way that provides useful information. Regarding the other pullback-based metrics presented in the appendix, we argue that our experiments are already a reduced setup with small image, dataset and network size. These other metrics are simply not easily usable at scale.

The threshold of 0.5 for orientability indeed assumes a balanced neighbourhood. It could be interesting to experiment with a more balanced measure. In practice, we found that the accuracy of learned hyperplane is very high (>90% on average) and that the framework is not really sensitive to moderate noise level in the labelled sequential factor, see Appendix H.1 and H.2. Hence, we did not explore this further.

We have added the correct attribution to Banerjee et al. (2005) in the main text.

In section 4.5, we provide the general evaluation pipeline. The retrieval method is then dataset-dependent. For the SheppLogan pipeline, we use cheap morphological operations that fallback to a ResNet18 when needed. For the KOMATSUNA dataset, we use a nnU-Net to segment the leaves and compute the area. We added more precise details in the dataset description and in the experimental details of the Appendix. On the SheppLogan dataset we report a Mean Absolute Percentage Error of 8% on the test set, and 3% on the KOMATSUNA dataset.

The reporting issues of the uncertainty in the tables has been raised by another reviewer. We modified the tables to report median and IQR instead of mean and std. We also added a scatter plot of the relative variation vs initial value of the sequential factor and the requested FWD vs. monotonicity. This allows to demonstrate that the initial value plays a big role in the relative variation.

The Wasserstein baseline does not provide interpolation results because it is not a geodesic method, but rather a transport method that does not provide a continuous path between two points given the endpoints. We have clarified this in the text.

We have also corrected the caught typos, as well as other typos in the paper.

We hope that with these additional results, corrections, and clarifications, the reviewer will feel more comfortable with the paper and its conclusions. We are happy to provide any additional information or clarification if needed.

Reviewer_Z2oqOpenReview ↗

Revised version cannot be opened

Thank you for your response to my concerns. Many answers refer to changes in the revised paper, but unfortunately the file cannot be opened. On my end, it fails both on openreview and when downloaded and opened locally, with different browsers and PDF readers.

Could you please verify and upload again the revised version? I will review the changes and provide feedback once it is available.

Dear reviewer,

Sorry for the previously corrupted file. We have reupdated the file. It seems to load fine now.

Note that the experiments of appendix H.3 are still running as we had to retrain ω\omega. They will be ready for the camera ready version. We expect the conclusions that we drew previously to still hold.

Best,

Reviewer_Z2oqOpenReview ↗

Comment on the Authors response

I thank the Authors for their response. After checking the revised manuscript, I confirm that my minor concerns have been addressed.

As far as the critical weakness I pointed out (i.e., relevance for the interpolation task) is concerned, I thank the Authors for acknowledging it and tempering their claims in the text. However, the newly provided results raise a concern about the extrapolation setting, which I had identified as the main strength of the paper. While I appreciate the Authors’ transparency, the FWD columns added to Tabs. 1 and 2, as well as the scatterplots provided in the appendix, reveal an important aspect that was not apparent in the original submission. Specifically, although the Kendall Tau improvement over the non-Finsler case is statistically significant, the FWD (which, according to the text, quantifies the extent to which generated images lie within the data distribution) is worse (i.e., higher) in all cases. The Authors comment on this “trade-off” in lines 447-452; however, from my perspective, the disagreement between the monitored metrics does little to support the proposed framework in the extrapolation setting as well, leaving its potential benefits unclear.

Therefore, after careful consideration, I have decided to lower my score from 3 to 2.

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

The paper learns a directional geometry on a pretrained VAE latent space from cross-sectional samples carrying an ordered scalar factor. A Randers norm adds a learned one-form to a base Riemannian norm. A local pairwise hinge objective learns a separating direction, and a second network predicts its amplitude from local ordering accuracy. The construction is applied to Euclidean, LAND, LAND-RBF, CovCentroids, and a centroid approximation of decoder pullback geometry.

For interpolation, the authors compute endpoint-conditioned geodesics. For extrapolation, they repeatedly sample velocities concentrated around the learned progression direction, apply a Legendre transform, and advance a dual-Hamiltonian system with momentum smoothing. Experiments use Shepp–Logan phantoms with a varying ellipse area and KOMATSUNA plant images with leaf area as the sequential factor. Evaluation emphasizes Kendall monotonicity, relative factor change, and image-distribution distance. The contribution is a method for incorporating local ordering into latent exploration; the experiments do not identify individual longitudinal dynamics from cross-sectional data.

Critical strengths
  1. A modular way to exploit local ordering without longitudinal training pairs. The proposed direction/amplitude factorization can be added to several existing latent metrics without retraining the VAE. Learning local before/after separation, while reducing the asymmetry where ordering is less reliable, is a useful inductive bias for generating candidate monotone traversals. The distinction between the base metric's support geometry and the directional component is conceptually helpful, even though the value of the learned amplitude is not established in the tested regimes.

  2. Controlled experiments make the desired behavior measurable and expose limitations. The scalar factors are interpretable in both datasets, allowing direct measurement of trajectory monotonicity. Tables 1–2 report Kendall scores close to one for several directed configurations, while also exposing poor performance of PBCentroids and cases where interpolation degrades. The label-noise and sampling-parameter analyses show that monotonicity and the amount of exploration can behave differently. Proposition 1 provides a useful conditional explanation for unchanged interpolation paths rather than asserting that asymmetry must improve every operation.

Critical weaknesses
  1. The extrapolation comparison does not isolate a benefit of asymmetric geometry. Section 4.4 gives the Randers method a learned, highly concentrated progression-direction sampler (eta=0.999 in the main tables), resampled throughout the trajectory, while symmetric baselines draw centered Gaussian velocities. A near-zero average directional score is unsurprising for the latter: it is not evidence that a Riemannian metric cannot extrapolate when provided the same direction information. The necessary control is a Riemannian sampler using the same learned progression field, speed, concentration, and refresh schedule, plus a direct learned-field flow. This is particularly important because Proposition 1 says an exact one-form changes fixed-endpoint lengths without changing unparameterized geodesic paths. Algorithm 6 is also a repeatedly refreshed stochastic process, rather than one uninterrupted geodesic. Finally, Kendall tau and large relative factor change establish ordered traversal, not image fidelity or a meaningful future trajectory. The FWD columns in Tables 1–2 are placed under interpolation, leaving extrapolated-image quality unquantified there. Appendix H switches from morphological measurement to a ResNet18 score predictor when decoding becomes unreliable, but does not validate that fallback on extrapolated images. Thus the main claim about directionality beyond curvature is not distinguished from directed sampling and potential score-estimator bias.

  2. The mathematical specification contains consequential inconsistencies in precisely the components used for directional shooting. For an increasing-score neighbor, Eq. (9) assigns label −1, but Algorithms 1–2 assign +1. Moreover, the hinge uses phi_tildeᵀ G delta, while Eqs. (4), (7–8), and Algorithm 3 insert the same normalized coordinates directly into the covector pairing omegaᵀ delta. These signs need not agree when G is not identity: with G=diag(10,1), phi_tilde=(1,1), and delta=(−1,2), the hinge pairing is −8 while the Randers linear pairing is positive. A consistent tangent-to-cotangent conversion is missing or unstated. Appendix A.3 Eq. (28) additionally inserts G_star inverse into the dual linear term. Under Eq. (27)'s natural norm convention, the correct term is −G inverse omega/delta, without that extra factor; for the one-dimensional example G=1, omega=1/2, p=1, the printed formula gives 23/24 instead of the defining dual value 2/3. The standard dual expression is given in Weber et al., CVPR 2024, Eqs. (22–24). Algorithm 5 also uses full-step momentum kicks where the stated generalized leapfrog requires half steps, and Eq. (14) averages velocities whereas Algorithm 6 averages momenta, which are not interchangeable under a nonlinear Legendre transform. Some issues may be typographical, but collectively they prevent determining whether the experiments implement the claimed Randers Hamiltonian and learned ordering.

Questions for authors
  1. Matched directional controls: With the same learned field and ordering supervision, how do Riemannian shooting and direct vector-field integration perform under matched initial directions, speed, concentration, and refresh/smoothing schedules? Can you separately show a single Hamiltonian trajectory without refresh, and the refreshed stochastic sampler? This would identify the contribution of the asymmetric metric rather than the sampling policy.

  2. Signs and geometric types: Which label convention generated the results? Is the learned network output a tangent vector or a one-form, and where is the metric conversion applied between the hinge, norm constraint, Randers pairing, and sampled velocity? Please reconcile the non-Euclidean counterexample above and state the exactness condition in Proposition 1 as omega=d psi. Local ordering accuracy does not by itself establish exactness or monotonicity along a finite trajectory.

  3. Dual Hamiltonian and numerical solver: Please resolve Definition 1, Eq. (28), Algorithm 5's momentum coefficients, and the velocity-versus-momentum smoothing discrepancy. Can you verify the Legendre/dual identity on simple non-Euclidean metrics and report Hamiltonian drift or step-size convergence for unrefreshed shooting? Please indicate which printed corrections, if any, require rerunning the tables.

  4. Extrapolation validity and statistical reporting: Please report extrapolation-specific image-quality/distribution metrics and independently validate the sequential-factor estimator, including the frequency and error of the ResNet fallback. How is leaf area recovered for decoded KOMATSUNA images? How many independent starting points, sampled paths, and training seeds underlie each uncertainty, and are KOMATSUNA splits separated by plant/sequence? Please also explain the substantially different zero-noise CovCentroids extrapolation result in Appendix F.2 (tau=0.48) versus Table 1 (approximately 1.0), including any changed settings. These checks can be performed on the existing controlled datasets.

2 comments

We thank the reviewer for the detailed review that will significantly help us improve the paper. We address the points below.

Typos and inconsistencies.

We initially used in the paper the expression of the dual Randers metric given in Yang et al. (2018) and Mirebeau et al. (2013) that contained some errors in their derivation. We now use the correct expression derived in Weber et al. (2024). The code in the supplementary already used the correct expression nevertheless. On random metrics GG and ω\omega, we verified the dual identity. We also fixed

  • The label in Eq.9 and Alg.2 are now coherent with the code, i.e. yij=2×1{(si>sj)}−1y^j_i = 2\times \mathrm{1}_{\{(s_i > s_j)\}} - 1, ensuring increasing dynamics.
  • Alg.5 matches the code and theory and correctly uses half-step momentum updates.
  • Alg.6 correctly averages the velocities instead of the momenta, as done in the code and as stated in the main body.
  • Prop.1 and its proof correctly state the exactness condition as ω=dψ\omega = \mathrm{d}\psi. Note that we do not prove that the learned ω\omega satisfies it but it provides relevant intuition to interpret the results.

Loss, pairing and sampling conventions

Also, as pointed out by R3, we (i) correctly specify that ϕ\phi is a covector, and (ii) removed the metric GG in the hinge loss, making it both well defined and coordinate invariant (proof in the appendix). This resolves the sign issue and the pairing issue as the loss now correctly associates a covector with a vector (⟨ϕ~θ(zi),zj−zi⟩\langle\tilde{\phi}_\theta(z_i) ,z_j-z_i\rangle). This warrants a retraining of ω\omega. After rerunning of all the tables using Rander geometry, we updated the results accordingly. Note also that in the extrapolation setting, we use the rescaled covector ω0\omega_0 as the vector mean of the vMF distribution, i.e. a Euclidean identification in latent coordinates. This choice is not coordinate invariant: a tractable and coordinate-invariant tangent Finsler distribution adapted to the Randers metric is left for future work.

Hamiltonian Convergence

We report the relative variation of the Hamiltonian along unrefreshed trajectories (Finslerian and Riemannian, LAND base metric) for an increasing number of time points. The variation stays below 10% even with few points and quickly decreases as the step size decreases (plot: imageupload.io/i/gVciVvSv). We integrate the dual Hamiltonian with the generalized leapfrog scheme, solving implicit steps with 5 fixed-point iterations, and clarified this in the appendix.

Experiments reporting

We now report median and IQR instead of mean and std, and add a scatter plot of relative variation versus initial value of the sequential factor (Fig. 3), showing the large influence of the initial value. Trajectories use 25 time points. For Shepp-Logan we compute 250 interpolated and 500 extrapolated trajectories (test samples as start/end points); for KOMATSUNA, 90 and 180. Results use a single training seed. The KOMATSUNA dataset is split by plant, so no plant is shared between train and test. Leaf area is obtained by nnU-Net segmentation. We report the regression error of the sequential-factor estimator in the appendix: MAPE of 8% (Shepp-Logan) and 3% (KOMATSUNA) on the test set. Note that extrapolated images have no ground truth, so the estimator cannot be validated on them directly. We acknowledge this as a limitation. The Appendix F.2 (now H.2) value at zero noise was stale from before ω\omega was retrained and will be updated in the final version.

Sampling discrepancies

As requested, we added a Riemannian baseline using the same learned ω\omega, concentration η\eta, initial speed magnitude α\alpha and refresh schedule, integrating the Riemannian Hamiltonian (Appendix I). With velocity refreshed at every step, it performs on par with the Randers framework. In this regime the directional sampler provides most of the observed benefit. We observe similar Kendall tau and relative variation for both methods. Note that we observe that unrefreshed trajectories are also more prone to exploding numerical instabilities.

In short, our datasets show monotone, potential-like progression, which is exactly the regime of Proposition 1, so they cannot separate the two formulations. We expect the Randers formulation to matter when progression is not a gradient field (non-closed ω\omega) and leave this to future work. Our work's aim is exactly to propose to use Finsler geometry to encompass these ideas of directionnal sampling on Riemannian manifolds. We argue that the Finsler metric encodes the directional information in the geometry rather than in an external sampling rule and hence allows to work with a more self-contained geometrical framework.

We hope that with these additionnal results, corrections, and clarifications, the reviewer might consider increasing their score. We are happy to provide any additional information or clarification if needed.

Thanks for the replies and extra experiments, they clear up several points, but my main concern remains: the matched Riemannian baseline performs similarly to Randers, suggesting most gains come from directional sampling rather than asymmetric geometry. Extrapolated-image quality is also still unclear. Seeing the full retrained results and clarified sign conventions would help, but for now, I’m not fully convinced that the experiments support the paper’s central claim.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.