acceptodds
Under review as a conference paper at ICLR 2027

Revisiting Log-Transformation in Deep Regression: Rectifying Metric-Warped Interpolation Bias via Transient Variance Guidance

Abstract

Log-transformation is a standard paradigm for deep regression on heavily skewed data. We find that, for data whose underlying structure is additive-dominant in the original space, the network's smoothness preference in log space drives its predictions away from the data's intrinsic regularity, inducing a systematic interpolation bias during generalization. We define this issue as metric-warped interpolation bias, which arises from the joint effect of the smooth inductive bias of neural networks and the strict convexity of the exponential mapping. Unlike the widely studied pointwise inverse-transform error, this issue has remained unexplored in previous work. Although directly introducing linear intervention to mitigate this bias appears intuitive, it easily destabilizes optimization. To address this dilemma, we propose Transient Variance Guidance (TVG), which repurposes the transient variance–error coupling of Negative Log-Likelihood (NLL), typically regarded as undesirable, as an endogenous proxy for fitting lag. TVG uses this transient variance-based fitting-lag signal to safely inject linear intervention into the learned representation, incurs zero computational overhead during inference, and consistently outperforms mainstream log-regression baselines across five cross-domain skewed datasets.

Primary area:
optimization
Keywords:
Deep Regression, Log-Transformation, Inductive Bias, Skewed Distributions, Optimization Dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WjZtOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3(+1)
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies interpolation bias induced by log transformations in deep regression. The authors argue that when the underlying target function is locally additive in the original space, while DNNs approximately interpolate linearly in log space, the inverse-transformed predictions may exhibit systematic underestimation. The authors then propose Transient Variance Guidance (TVG), which uses the predictive variance learned through Gaussian NLL to construct auxiliary supervision in the original target space, with stop-gradient controlling gradient propagation. Various experiments validated the proposed method.

Critical strengths
  1. The paper analyzes prediction bias induced by log transformations from the perspective of interpolation geometry, providing an interesting perspective on this problem. The design of TVG is well motivated and tailored to the proposed issue.

  2. The empirical evaluation is relatively comprehensive. The proposed method demonstrates improvements across multiple real-world datasets and network architectures, while ablation studies support the effectiveness of its main design choices.

  3. Appendix A explicitly investigates the applicability and limitations of the method. Through synthetic experiments with additive and multiplicative data-generating mechanisms, the authors demonstrate how TVG behaves under different target function structures, including cases where it may substantially underperform the baseline under strongly multiplicative relationships.

Critical weaknesses

My main concerns relate to the applicability of the proposed method and the conditions under which it is expected to be effective.

  1. The proposed method mainly targets the mismatch between locally additive structures in the original target space and approximately linear interpolation in log space. To what extent is TVG limited to regression problems exhibiting this type of local structure? Could the underlying idea be extended to other commonly used nonlinear target transformations, such as the logit transformation?

  2. The effectiveness of the method is primarily supported by empirical evidence, and the synthetic experiments also show that TVG can substantially underperform the baseline in certain settings. Are there simple empirical criteria or diagnostic procedures that could help practitioners determine in advance whether TVG is appropriate for a given task? Clearer guidance on its applicability would improve the practical usefulness of the method.

Questions for authors

Questions For Authors:

  1. In the existing 1D synthetic experiments, could the authors plot the learned predictions expm1⁡(μ(x))\operatorname{expm1}(\mu(x)) for both the baseline and TVG, together with the ground truth Y(x)Y(x) and the prediction obtained by linearly interpolating Z=log⁡(1+Y)Z=\log(1+Y) between training points? This would help determine whether the baseline exhibits the interpolation bias described in the paper and whether TVG reduces it.

  2. Could the authors examine how the learned variance σt2(x)\sigma_t^2(x) relates to the squared fitting error (E[Z∣x]−μt(x))2(\mathbb{E}[Z\mid x]-\mu_t(x))^2 during training? The paper motivates TVG using the optimal variance under Gaussian NLL, but it remains unclear whether the learned σt2(x)\sigma_t^2(x) tracks this error in practice.

  3. The synthetic experiments in Appendix A use only five training points. Could the authors vary the number of training points and examine how the performance gap between TVG and the baseline changes under Y(x)=a+bxY(x)=a+bx and Y(x)=exp⁡(a+bx)Y(x)=\exp(a+bx)? This would help clarify whether the observed effects depend on the sparsity of the training data.

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer WjZt

We thank the reviewer for recognizing the method design, experimental results, and applicability analysis. Following the reviewer’s suggestions, we added the following experiments and discussion.

1. Actual prediction curves and bias reduction

Our work concerns the potential mismatch between the inductive bias induced by target transformation and the original-space target structure. Equations (1)–(2) use two points to illustrate how this mismatch manifests as trajectory deviation; the difference between the arithmetic and inverse-transformed geometric paths is not an assumption about the full generalization behavior of an actual network. For networks jointly constrained by multiple training points, the deviation depends on the architecture and optimization and must be examined through actual predictions.

We therefore added the ground-truth, Log-MSE, and TVG prediction curves for the additive task (b=10,50,100b=10,50,100) alongside 500-seed signed-error statistics (Appendix A.3). The learned predictions exhibit both local overestimation and underestimation, but the baseline shows an overall negative bias that TVG reduces while lowering MAE: for b=10b=10, 5050, and 100100, mean signed errors improve from −0.0272-0.0272, −0.2374-0.2374, and −0.7107-0.7107 for Log-MSE to −0.0122-0.0122, −0.1553-0.1553, and −0.4703-0.4703 for TVG. For visual clarity, deviations from the ground truth are magnified 10×10\times and labeled in the prediction figure, while the signed-error plot and all metrics remain on their original scales. Thus, under multi-point training, TVG effectively alleviates the overall negative bias and reduces original-space error.

2. Relationship between predicted variance and fitting lag

We recorded the evolution of the predicted variance σt2(x)\sigma _ t ^ 2(x) and squared fitting lag (E[Z∣x]−μt(x))2(\mathbb{E}[Z\mid x]-\mu _ t(x)) ^ 2 under the joint loss in Equation (6) over 500 random seeds. In early training (Epochs 50 and 100), they show clear positive correlations of (0.8243,0.5584)(0.8243,0.5584) for b=10b=10, (0.9119,0.9656)(0.9119,0.9656) for b=50b=50, and (0.9090,0.9929)(0.9090,0.9929) for b=100b=100. By Epoch 2000, both means decrease substantially (Appendix A.4), showing that predicted variance retains information associated with fitting lag under joint training.

3. Effect of the number of training points

We varied the number of training points from 55 to 300300 in both additive and multiplicative settings (Appendix A.5). Representative endpoint results (mean MAE ±\pm std) are:

Data-generating mechanismTraining pointsLog-MSETVG
Additive50.7233±0.38930.7233\pm0.38930.6711±0.29990.6711\pm0.2999
Additive3000.4515±0.43860.4515\pm0.43860.3689±0.38720.3689\pm0.3872
Multiplicative50.3621±0.32900.3621\pm0.329018.8447±21.168818.8447\pm21.1688
Multiplicative3000.2648±0.29100.2648\pm0.29100.5883±4.29840.5883\pm4.2984

In the additive setting, TVG achieves lower average MAE and RMSE across all tested training-set sizes, showing its benefit is not restricted to the sparse five-point setting. In the multiplicative setting, TVG’s average disadvantage decreases substantially as training points increase—from −18.48-18.48 to −0.32-0.32—but remains present throughout the tested range, clarifying the applicability boundary.

4. Applicability criteria and other target transformations

TVG targets settings where the original-space response is locally additive, while logarithmic preprocessing is introduced to compress dynamic range or improve optimization. In practice, domain knowledge and mechanistic models provide prior evidence for applicability, while task-specific benefits should be confirmed on the validation set. The multiplicative experiments (Y=exp⁡(a+bu)Y=\exp(a+bu) in Point 3 and Appendices A.2 and A.5) explicitly demonstrate the failure mode under predominantly multiplicative relationships. Thus, a locally additive original-space mechanism is the central criterion for applying TVG.

Conceptually, this perspective may be extended to other nonlinear target transformations such as the logit transformation. Because the present work focuses on logarithmic transformation, we leave extensions and evaluations under other nonlinear transformations to future work.

I thank the authors for their detailed response and for addressing all of my concerns. I will maintain my original positive assessment of this work and have accordingly increased my individual scores and confidence rating.

We thank the reviewer for carefully reconsidering our work and for confirming that the response and revisions address the concerns.

Reviewer_975tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies prediction bias in the original target space that may arise from log-transformation in deep regression. From a local interpolation perspective, the authors analyze the difference between affine interpolation in log space and arithmetic interpolation in the original space, and propose Transient Variance Guidance (TVG). The method trains the primary branch with a Gaussian negative log-likelihood in log space and uses the predicted variance to construct an auxiliary prediction. An original-space loss provides training guidance, with the direct gradient from the auxiliary loss to the mean output blocked; the auxiliary branch is removed at inference. The authors report improvements over standard log regression across five datasets and multiple backbone architectures, and use synthetic experiments to examine performance under additive and multiplicative data structures. The main contributions are a conditional explanation of interpolation bias and a training method supported by empirical results across tasks.

Critical strengths
  1. The method shows practical value across multiple tasks, with limited additional computation at inference: In Table 1, TVG reduces MAE on CIKM from 6.8930 to 6.2059 and RMSE from 10.2653 to 9.4859 relative to LogMSE. Tables 4-5 report improvements on other datasets and with different backbone architectures, and the main experiments report results from five independent runs. In the configuration compared in Table 3, TVG and LogMSE have the same inference FLOPs. These results support the practical value of TVG as a training method and are the main reason for my positive recommendation. (pp. 7-9, Tables 1 and 3-5)

  2. The geometric analysis provides a checkable explanation, and the synthetic experiments reveal limits to its applicability: Assuming an arithmetic reference path in the original space and an affine prediction path in log space, the paper derives the difference between the two paths after the inverse transformation, together with a second-order approximation. Appendix A examines both additive and multiplicative constructions and reports degradation in the stronger multiplicative settings. This helps identify circumstances in which the method may be beneficial and makes clear that its empirical benefits are not unconditional. (pp. 2-3, Section 2.1; Appendix A, Tables 6-9; Appendix B.1)

Critical weaknesses

The experiments support the effectiveness of the method, but do not sufficiently distinguish the proposed mechanism from other possible sources of improvement.

The optimal variance decomposition in Eq. (4) applies when the mean is fixed and only the Gaussian negative log-likelihood is optimized, whereas the auxiliary loss in the actual training objective, Eq. (6), also affects the variance. More direct evidence is therefore needed to establish how well the variance under joint training continues to represent fitting lag. The direct-residual comparison in Table 2 also changes the parameterization and thus does not isolate the role of the variance term. Figure 5 shows the primary and auxiliary predictions becoming closer, but this alone does not establish that the auxiliary gradients also diminish, or that this transient behavior is necessary for the performance gains.

The authors appropriately acknowledge in Section 5 that the empirical gains provide only indirect support for the mechanism. The wording in Section 4.2 about ruling out a generic regularization explanation should be similarly cautious. This concern mainly limits the explanation of where the gains come from, rather than directly undermining the reported improvements. (pp. 4-5, Eqs. (4)-(6); p. 7, Table 2 and Section 4.2; p. 8, Figure 5; p. 9, Section 5)

I also suggest the following minor corrections:

  • Please proofread repeated equation-reference prefixes such as "Equation equation", remove the duplicated "prediction prediction" on p. 8, change "liner space" to "linear space" on p. 24, and check capitalization within sentences and punctuation.

  • Please check the derivative on p. 23, line 1225. Given the definition on the same page, Yaux=expm1⁡(sg⁡[μ]+γ)Y_{\mathrm{aux}} = \operatorname{expm1}(\operatorname{sg}[\mu] + \gamma), the derivative should be ∂Yaux/∂γ=Yaux+1\partial Y_{\mathrm{aux}}/\partial\gamma = Y_{\mathrm{aux}} + 1, rather than YauxY_{\mathrm{aux}}, when clipping is inactive. Please correct the expression and clarify whether this discrepancy is confined to the written derivation.

Questions for authors
  1. Under the actual joint loss, is there direct evidence that the predicted variance still reflects fitting lag and that this relationship contributes to the performance gains? For example, a control that replaces variance modulation in the auxiliary prediction with a constant scale, while retaining the other architectural choices, gradient handling, and tuning budget, could help distinguish the use of variance information from a general auxiliary-training effect.

  2. Which data split is used in Figure 5, and does the figure aggregate multiple independent runs? As the primary and auxiliary outputs become closer, do the gradients of the auxiliary loss with respect to the shared parameters also decrease? This would help establish whether the training dynamics directly support the "Transient" characterization and under what conditions the explanation in Appendix B.4.3 applies.

  3. Please point to the locations in the supplementary code or configurations that specify the data splits and prediction horizons, the training and tuning budgets for each baseline, and the data used for the Duan correction and for selecting λ\lambda. These details would help verify the comparison settings and reproduce the results.

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer 975t

We thank the reviewer for the careful and constructive comments.

1. Variance–fitting-lag association and matched ablation

To examine whether the predicted variance continues to reflect fitting lag under the joint objective, we measured the correlation between σt2(X)\sigma _ t ^ 2(\mathbf{X}) and (Z∗(X)−μt(X))2(Z ^ {*}(\mathbf{X})-\mu _ t(\mathbf{X})) ^ 2 on the noise-free additive task. At Epochs 50 and 100, the correlations are (0.8243,0.5584)(0.8243,0.5584) for b=10b=10, (0.9119,0.9656)(0.9119,0.9656) for b=50b=50, and (0.9090,0.9929)(0.9090,0.9929) for b=100b=100. Thus, the variance branch remains informative of fitting lag under joint training (Appendix A.4).

Following the reviewer’s suggestion, we also introduced a matched constant control that retains the NLL mean and variance heads, stop-gradient, and original-space MAE loss, while replacing β(X)σ2(X)\beta(\mathbf{X})\sigma ^ 2(\mathbf{X}) with cβ(X)c\beta(\mathbf{X}). Both variants use identical data splits, training budgets, and model-selection rules.

On CIKM2017, the constant control obtains an MAE/RMSE of 6.762±0.2056.762\pm0.205 / 10.006±0.35510.006\pm0.355, compared with 6.8930±0.07806.8930\pm0.0780 / 10.2653±0.171910.2653\pm0.1719 for Log-MSE. TVG obtains 6.2059±0.09926.2059\pm0.0992 / 9.4859±0.16239.4859\pm0.1623, further reducing MAE and RMSE by 8.22% and 5.20% relative to the constant control (Table 2).

On the noise-free additive task, TVG achieves a lower mean MAE than the constant control for all evaluated slopes (b=10,50,70,100b=10,50,70,100), with significant paired differences at b=50b=50 (p=0.0001p=0.0001), b=70b=70 (p<0.0001p<0.0001), and b=100b=100 (p<0.0001p<0.0001). Thus, original-space MAE supervision and stop-gradient alone do not reproduce TVG’s performance; transient variance modulation is an indispensable component (Appendix A.6).

2. Figure 5 and transient auxiliary dynamics

Figure 5 reports the validation-set mean absolute discrepancy E[∣Y^−Y^aux∣]\mathbb{E}[\lvert \hat{Y}-\hat{Y} _ {\mathrm{aux}} \rvert] during one representative CIKM2017 run, illustrating why the auxiliary branch can be removed at inference.

Equation (46) gives the single-sample auxiliary gradient on the shared backbone as proportional to sign⁡(Y^aux−Y)Y^aux∇θ[β(X)σ2(X)]\operatorname{sign}(\hat{Y} _ {\mathrm{aux}}-Y)\hat{Y} _ {\mathrm{aux}}\nabla _ \theta[\beta(\mathbf{X})\sigma ^ 2(\mathbf{X})]; under the shifted inverse transformation, the multiplier is Y^aux+1\hat{Y} _ {\mathrm{aux}}+1. In the noise-free task, the auxiliary-gradient norm should decay as fitting converges and σ2(X)→0\sigma ^ 2(\mathbf{X})\to0. In noisy real-world data, because the MAE derivative has unit magnitude for any nonzero residual, nonzero variance and finite-batch sampling may retain gradient fluctuations even when the modulation becomes small.

We therefore examined both settings. From Epoch 50 to Epoch 2000 on the noise-free additive task, the mean auxiliary-gradient norms decrease from 1.921.92, 5.135.13, and 8.068.06 to 3.73×10−53.73\times10 ^ {-5}, 6.24×10−46.24\times10 ^ {-4}, and 1.3×10−21.3\times10 ^ {-2} for b=10b=10, 5050, and 100100, respectively (Appendix A.7).

On CIKM2017, we additionally plot the validation-set auxiliary-modulation magnitude in Figure 5. Both the modulation magnitude and the primary–auxiliary prediction discrepancy decrease smoothly to very low levels, demonstrating the transient disappearance of the auxiliary correction.

3. Wording and minor corrections

We revised Section 4.2’s statement about excluding generic regularization to match the qualified wording in Section 5.

Equation (46) uses the unshifted exp⁡(⋅)\exp(\cdot) formulation, whose derivative is Y^aux\hat{Y} _ {\mathrm{aux}}. Page 23 subsequently introduces the shifted inverse expm1⁡(⋅)\operatorname{expm1}(\cdot) but inadvertently retains the unshifted derivative notation. We corrected it to Y^aux+1\hat{Y} _ {\mathrm{aux}}+1. This was only a written-derivation issue; training uses PyTorch automatic differentiation and is unaffected.

We also corrected the repeated equation prefixes, the duplicated “prediction” on page 8, and the “linear space” typo on page 24.

4. Reproducibility details

We expanded Appendix C with the data splits and prediction horizons for all datasets and confirmed that all baselines and TVG use identical training and tuning budgets (Table 10). We further clarify that Duan’s smearing factor is computed from training-set residuals, while λ\lambda and cc are selected strictly using validation performance.

Reviewer_pDd3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies interpolation bias in log-transformed deep regression and proposes Transient Variance Guidance (TVG). TVG uses the predictive variance learned through Gaussian NLL to modulate an auxiliary original-scale loss, with stop-gradient applied to the mean prediction. The auxiliary branches are removed at inference. Experiments report improvements across five datasets and several backbone architectures.

Critical strengths

The paper addresses an important practical issue, as target transformations are widely used in regression and can substantially affect both optimization and interpolation behavior.

The proposed TVG is straightforward to implement and introduces few additional training parameters. Removing the auxiliary branches preserves the original inference architecture.

Critical weaknesses

(1) The paper should more clearly identify the specific conceptual and methodological improvements of TVG over TranSUN. The interpolation-bias formula follows from convexity and a Taylor expansion under an assumed affine interpolation path, but there is insufficient evidence that networks trained in practice follow such a path.

(2) The paper does not adequately explain whether the evaluated tasks satisfy the theoretical assumptions, particularly local additivity on the original scale. Skewness or zero inflation alone does not imply this property.

(3) The current experiments do not clearly separate the benefits of transient variance guidance from those of original-scale auxiliary supervision. A matched control that retains the NLL variance head, stop-gradient, and original-scale auxiliary loss, while removing variance modulation from the auxiliary prediction, would help isolate the contribution of this central mechanism.

Questions for authors

What are the specific conceptual and methodological advances of TVG over TranSUN? Under what conditions should a trained network exhibit approximately affine interpolation in log space? Is this assumption supported by theoretical analysis or measurements of actual network interpolation paths?

Beyond target skewness or zero inflation, what task-specific properties or data evidence support local additivity on the original scale? Could you provide practical criteria for determining when TVG is appropriate and when it might degrade performance?

Could you add a matched ablation on controlled synthetic data and one representative real-world dataset, retaining the NLL variance head, stop-gradient, and original-scale auxiliary loss, but replacing β(X)σ2(X)\beta(X)\sigma^2(X) with β(X)\beta(X) in the auxiliary prediction? How does this control compare with full TVG under identical training and model-selection settings?

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer pDd3

We thank the reviewer for the suggestions.

1. Difference from TranSUN and the interpolation assumption

TranSUN uses an explicit correction factor at inference to estimate the conditional mean in the original space. TVG instead focuses on how target transformation affects the prediction trajectories between samples. Through variance modulation, TVG introduces original-space supervision into the shared backbone during training, while retaining only the primary branch at inference (namely, the log-mean prediction head Y^=expm1(μ(X))\hat{Y}=\mathrm{expm1}(\mu(\mathbf{X}))) and removing the auxiliary variance and modulation branches. Thus, TVG and TranSUN address different sources of error.

The distinction is particularly clear in the noise-free setting: when random noise is absent, the pointwise retransformation bias targeted by TranSUN vanishes, whereas finite-sample prediction bias between observed samples may still remain. Appendix D provides a detailed discussion of these two types of error.

Our work concerns the potential mismatch between the inductive bias introduced by target transformation and the target structure in the original space. Equations (1)–(2) use two points to illustrate how this mismatch may manifest as a deviation in the prediction trajectory. The difference between the arithmetic path and the inverse-transformed geometric path in this illustration is not an assumption about the full generalization behavior of an actual network. For a network jointly constrained by multiple training points, the exact form of the deviation depends on the model architecture and optimization process and must therefore be examined through the actual predictions.

To this end, we added the actual prediction curves of the baseline and TVG on the synthetic task, together with signed-error statistics over 500 random seeds (Appendix A.3). The learned predictions exhibit both local overestimation and local underestimation. Nevertheless, in the additive settings examined, the baseline exhibits an overall negative bias, while TVG reduces this bias and lowers the MAE. These results show that, under multi-point training, TVG effectively alleviates the aggregate bias and improves prediction accuracy in the original space.

2. Local additivity and applicability

We agree that skewness and zero inflation alone are insufficient evidence of local additivity. Our work targets settings in which the target response has a locally additive structure in the original space, while logarithmic preprocessing is introduced to compress the target dynamic range or improve optimization. The resulting concern is a potential mismatch between the transformation-induced inductive bias and the original-space target structure.

For example, in rainfall–runoff forecasting as represented by CAMELS, the classical unit-hydrograph formulation models direct runoff as a linear superposition of effective-rainfall responses, providing a clear domain-modeling basis for additive structure in the original space. This motivates retaining original-space supervision in such tasks.

In practice, domain knowledge and existing mechanistic models can inform the applicability of TVG. TVG has a clear motivation when the log transformation is used primarily for optimization and the original-space response has a locally additive structure. Conversely, TVG may be unsuitable when the target relationship is predominantly multiplicative, as demonstrated by the multiplicative synthetic experiments in Appendix A.2. For any particular task, its empirical benefit should still be assessed on the validation set.

3. Matched ablation

Following the reviewer’s suggestion, we performed a matched ablation that retains the NLL mean and variance heads, stop-gradient, and original-space MAE loss, while replacing β(X)σ2(X)\beta(\mathbf{X})\sigma ^ 2(\mathbf{X}) in the auxiliary prediction with cβ(X)c\beta(\mathbf{X}). We refer to this variant as the constant control. The two variants use identical data splits, training budgets, and model-selection rules. The constant cc is searched over the range [0.1,1][0.1,1] and selected according to validation performance.

On CIKM2017, the constant control achieves an MAE of 6.762±0.2056.762\pm0.205, representing only a limited improvement over the original Log-MSE baseline (6.8930±0.07806.8930\pm0.0780). In contrast, TVG achieves an MAE of 6.2059±0.09926.2059\pm0.0992 and further reduces MAE and RMSE by 8.22% and 5.20%, respectively, relative to the constant control.

On the noise-free additive task, TVG obtains a lower mean MAE than the constant control under all evaluated slope parameters (b=10,50,70,100b=10,50,70,100). The paired differences are highly significant at b=50b=50 (p=0.0001p=0.0001), b=70b=70 (p<0.0001p<0.0001), and b=100b=100 (p<0.0001p<0.0001). These results show that the original-space MAE loss and stop-gradient alone are insufficient to reproduce the performance of TVG, and that transient variance modulation is an indispensable component of the method (Appendix A.6).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.