acceptodds
Under review as a conference paper at ICLR 2027

Amortized Variational Inference of Latent Intensities from Replicated Point Patterns

Abstract

Bayesian inference for latent point process intensities often relies on per-dataset sampling or optimization, making repeated inference computationally expensive. We introduce an amortized variational framework based on a Neural Process for replicated point patterns, where a variable-size set of realizations shares a latent intensity. Trained from event times using the point process likelihood, the model learns an amortized inference rule that maps these realizations to a variational distribution over their shared intensity, without ground-truth intensity supervision. At test time, it performs inference of the latent distribution in a forward pass. On synthetic data, additional realizations improve intensity recovery and yield narrower credible intervals with broadly stable coverage. On bicycle-sharing data, our method achieves the highest mean held-out log likelihood among the evaluated approaches. It also requires substantially less inference time than the LGCP baseline in our experiments.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
temporal point processes, Cox processes, amortized variational inference, neural processes, uncertainty quantification, replicated point patterns

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_nh35OpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes an amortized variational framework for inferring a latent intensity function shared across replicated temporal point patterns. A permutation-invariant encoder maps a variable number of inhomogeneous Poisson process realizations to a latent distribution, which is decoded into a distribution over continuous-time intensity functions. The model is trained with a variational objective (ELBO) that uses only event times and the point process likelihood, enabling fast inference on new tasks without per-task optimization. Synthetic and real-data experiments evaluated intensity recovery, uncertainty quantification, and adaptation to additional realizations, showing that the proposed method outperformed the baselines compared in intensity recovery and held-out likelihood.

Critical strengths
  • The paper addresses a practically relevant problem: fast inference of a shared latent intensity from replicated point process observations, avoiding costly per-task optimization at test time.
  • The proposed framework is conceptually simple and well structured, combining permutation-invariant pooling with a neural process-based conditional variational objective and point process likelihood training.
  • The effectiveness of the proposed method was demonstrated on synthetic and real-world data.
Critical weaknesses
  • The novelty of treating multiple replicated realizations as a distinct inference setting is unclear. Under the paper’s assumption that the realizations are conditionally independent inhomogeneous Poisson processes sharing the same intensity, the replicated-observation setting appears to reduce directly to standard intensity inference with a known exposure factor KK. The superposition of KK independent realizations with shared intensity λ(⋅)\lambda(\cdot) is a single inhomogeneous Poisson process with intensity Kλ(⋅)K\lambda(\cdot). This suggests that existing amortized intensity-inference methods, such as deep random splines and potentially neural diffusion intensity models, could be adapted to the present setting by incorporating KK into the observation likelihood, without requiring a fundamentally new replicate-aware inference framework. The paper should therefore clarify whether the proposed formulation offers a substantive methodological advantage over such straightforward extensions.
  • The experimental comparison among amortized methods is not sufficiently convincing. The proposed model is trained directly on thousands of tasks generated from the same GP-based inhomogeneous Poisson process family used for evaluation, whereas FIM-PP, the only amortized baseline, is evaluated using its released pretrained checkpoint without any task-specific adaptation. Since FIM-PP is designed to infer history-dependent conditional intensities rather than the intensity functions of inhomogeneous Poisson processes considered here, the substantial performance gap may primarily reflect a mismatch in task formulation and training distribution rather than a clear methodological advantage of the proposed framework. A more informative and fair comparison would either fine-tune FIM-PP on the same training tasks or include an amortized intensity-inference baseline trained on the same data-generating distribution under a comparable computational budget.
  • More broadly, if the goal is to evaluate uncertainty-aware intensity-function estimation, the baseline set should be substantially strengthened. As the authors note, several methods have been proposed that offer faster alternatives to LGCP-based inference (e.g., INLA), and at least some of them should be included for comparison. In addition, PSTH seems too weak a baseline for estimating the smooth intensity functions considered in the experiments, since it is inherently piecewise constant. If a non-Bayesian baseline is to be included, it would be more informative to compare not only with the classical KDE estimator but also with more recent kernel-based intensity estimators such as K2IEK^2IE (Kim et al., ICML, 2025).
Questions for authors

See Critical Weaknesses.

9 comments

Dear Reviewer nh35,

We thank the reviewer for the thoughtful comments, which helped us clarify our contributions relative to existing methods and strengthen the baseline comparisons.

We have carefully considered your comments and address your concerns below.

1. Methodological Novelty

We agree that under our assumption of observing KK independent realisations from a Poisson process, one can combine all the events into a single sequence and divide the estimated intensity, e.g., using some of the cited methods, by KK to obtain the desired intensity. However, our focus is not only on (fast) inference in this setting, but rather on designing a trainable neural-network-based model that, as more realizations become available, refines its estimates of a shared latent intensity and provides calibrated uncertainty quantification, without requiring per-task optimization or a computationally costly hierarchical (fully) Bayesian model. In other words, rather than proposing a fundamentally new inference setting, our main contribution is to introduce an amortized inference framework that processes multiple realizations from a shared intensity and is trained via a context-target objective.

Related to our response to a comment by Reviewer 9Uw9, an alternative approach could entail designing an encoder taking as input the aggregated sequence of events and KK. However, the results reported in the encoder comparison table in our response to Reviewer 9Uw9 show that such an encoder tends to underperform in terms of RMSE relative to ours, when the number of test replicates is larger than the number used during training.

Nonetheless we acknowledge that the set of baselines in our original manuscript was not sufficient, as also suggested by Reviewer F4ea. Therefore we have added 3 baselines: DRS-VAE (Loaiza-Ganem et al., 2019), LGCP-INLA (Rue et al., 2009; Bachl et al., 2019), and K$^2$IE (Kim et al., 2025). We leave the discussion on the latter two to the response to your 3rd comment below and here focus on the Deep Random Splines Variational Autoencoder (DRS-VAE). We adapted the latter to the replicated sequences setting by pooling the events into a single sequence. We trained the released architecture with the default model configuration and compared it to our method. We note that both methods use the same training, validation and test data, with checkpoints selected by validation held-out likelihood. We follow the synthetic and Divvy settings in Section 4 of the paper, using Ktrain=Ktest=4K_{\mathrm{train}}=K_{\mathrm{test}}=4 for synthetic data. We report our results in the table below.

Synthetic and Divvy results (mean ± standard deviation across four seeds).

MethodSynthetic RMSE ↓Synthetic C95 (%)Synthetic IW95Synthetic time (ms) ↓Divvy HLL ↑Divvy time (ms) ↓
DRS-VAE4.32±0.144.32 \pm 0.1430.74±2.1730.74 \pm 2.173.29±0.123.29 \pm 0.1267±0.0867_{\pm 0.08}103.31±2.77103.31 \pm 2.77253±1.04253_{\pm 1.04}
FIM-PP (pretrained)14.19——659.5376
FIM-PP (fine-tuned)2.78±0.062.78 \pm 0.06——5±0.025_{\pm 0.02}106.28±0.18106.28 \pm 0.1859±0.4659_{\pm 0.46}
LGCP–VI1.86±0.011.86 \pm 0.0192.70±0.1092.70 \pm 0.106.67±0.016.67 \pm 0.016,312±36.066{,}312_{\pm 36.06}108.80±0.01108.80 \pm 0.0124,407±76.8324{,}407_{\pm 76.83}
LGCP–INLA1.5992.135.7438,116110.05161,245
PSTH2.45——<1102.73<1
KDE1.91——<1107.1010
K2K^2IE1.60——3107.496
Ours1.55±0.011.55 \pm 0.0193.28±0.1893.28 \pm 0.185.73±0.005.73 \pm 0.008±0.448_{\pm 0.44}110.99±0.03110.99 \pm 0.0327±0.1327_{\pm 0.13}

We observe that our model achieves better performance than the DRS-VAE: on the synthetic data it has lower RMSE and coverage closer to the nominal level 95%, and has higher held-out likelihood on the Divvy dataset. As suggested by Reviewer F4ea, we added 2 prediction performance metrics to our comparison of methods on the Divvy dataset: the next-event RMSE (NE-RMSE) and CRPS. These results are in the next-event prediction table in our response to Reviewer F4ea and confirm that our method performs favourably relative to all baselines.

Besides, we further highlight the differences between our method and DRS-VAE, in terms of encoder, decoder, and training objective. DRS-VAE has a recurrent encoder and a spline decoder while our model only uses fully-connected layers. Moreover, while DRS-VAE uses a standard Gaussian prior VAE ELBO, our conditional objective uses a penalization in term of the divergence between the distribution inferred from a context subset and that inferred from the full set of realizations (context and target).

To assess the effectiveness of our training scheme, we additionally compare our model (trained with the conditional objective) to the same encoder-decoder architecture trained with a standard VAE ELBO. The results of this ablation study are reported in the table below. They show that our conditional objective leads to more accurate estimates and narrower intervals as the number of realizations increases, while maintaining relatively stable coverage, compared to the standard ELBO. These advantages become more pronounced when the number of replicates at test time is larger than the number used during training.

Synthetic results for the objective ablation (mean ± standard deviation across four seeds).

KtestK_{\mathrm{test}}RMSE Ours ↓RMSE Standard ↓C95 Ours (%)C95 Standard (%)IW95 OursIW95 Standard
12.68±0.012.68\pm0.012.64±0.002.64\pm0.0092.21±0.0592.21\pm0.0592.12±0.0892.12\pm0.089.36±0.029.36\pm0.029.38±0.039.38\pm0.03
21.99±0.001.99\pm0.002.09±0.012.09\pm0.0193.18±0.0993.18\pm0.0990.90±0.1590.90\pm0.157.20±0.027.20\pm0.027.10±0.017.10\pm0.01
41.55±0.011.55\pm0.011.75±0.011.75\pm0.0193.28±0.1893.28\pm0.1890.01±0.3590.01\pm0.355.73±0.005.73\pm0.005.90±0.055.90\pm0.05
81.23±0.011.23\pm0.011.50±0.011.50\pm0.0193.78±0.2593.78\pm0.2590.44±0.5390.44\pm0.534.78±0.014.78\pm0.015.27±0.085.27\pm0.08
161.00±0.011.00\pm0.011.37±0.011.37\pm0.0194.32±0.2894.32\pm0.2890.58±0.5490.58\pm0.544.23±0.014.23\pm0.014.97±0.084.97\pm0.08
320.88±0.010.88\pm0.011.36±0.021.36\pm0.0293.84±0.2193.84\pm0.2189.18±0.6989.18\pm0.693.91±0.023.91\pm0.024.84±0.094.84\pm0.09
640.82±0.020.82\pm0.021.44±0.031.44\pm0.0393.16±0.3093.16\pm0.3086.91±0.8586.91\pm0.853.74±0.033.74\pm0.034.79±0.084.79\pm0.08

Finally we underline that one of the methods cited in the related work, Neural Diffusion Intensity Model (NDIM; Du et al., 2026) could not be compared to due to the lack of an available implementation at the time of this rebuttal. We also recall that NDIM places a Markov diffusion prior on the latent intensity through a neural SDE, whereas our method does not impose this structural assumption.

2. Fine-tuning FIM-PP

Firstly, we would like to clarify that FIM-PP (Berghaus et al., 2026) is not restricted to history-dependent processes. In fact, its pre-training data includes realizations of homogeneous Poisson processes, and its data-generation scheme allows for the generation of inhomogeneous Poisson processes. Moreover, their paper also presents inference results for history-independent processes. Nevertheless, we agree with your comment that fine-tuning this model on our data provides a fairer and more informative comparison to our method.

We fine-tuned the released FIM-PP checkpoint on the same training tasks used for our model. The updated results are reported in the “Synthetic and Divvy results” table. We observed that fine-tuning substantially improved FIM-PP, nonetheless our model retains lower intensity RMSE on the synthetic data and higher held-out log-likelihood on the Divvy data.

3. Additional Baselines

We addressed your comment on the lack of strong baselines by adding the approximate Bayesian method LGCP-INLA and the kernel density-based estimator K2K^2IE. However we would like to clarify that our original set of baselines already included a fast inference method for the LGCP (LGCP-VI). The latter uses sparse variational inference with inducing points, which is already a faster alternative to full GP inference.

Nonetheless the updated experimental results, reported in the “Synthetic and Divvy results” table, show that LGCP-INLA is a more accurate baseline than LGCP-VI, at the cost of longer inference time. Moreover compared with LGCP–INLA, our model achieves comparable intensity recovery and credible interval width, but coverage closer to the nominal level 95%, on the synthetic data. On the Divvy data, it achieves slightly higher Divvy held-out log-likelihood, but has substantially shorter inference time. Compared with K2K^2IE, our model achieves slightly lower synthetic intensity RMSE and higher Divvy held-out log-likelihood. We retain PSTH as a simple histogram reference, alongside KDE and K2K^2IE.

Therefore, overall our method achieves favourable performance compared to all tested baselines. We thank the reviewer for their suggestions of adding LGCP-INLA and K2K^2IE which were shown to be the strongest baselines in our comparison.

In the revised manuscript, we will include the additional baseline comparisons and the objective ablation study.

References

Fabian E. Bachl, Finn Lindgren, David L. Borchers, and Janine B. Illian. inlabru: an R package for Bayesian spatial modelling from ecological survey data. Methods in Ecology and Evolution, 10:760–766, 2019.

David Berghaus, Patrick Seifner, Kostadin Cvejoski, César Ali Ojeda Marin, and Ramses Sanchez. In-Context Learning of Temporal Point Processes with Foundation Inference Models. In International Conference on Learning Representations, pp. 29610–29641, 2026.

Xinlong Du, Harsha Honnappa, and Vinayak Rao. Neural Diffusion Intensity Models for Point Process Data. In Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337, pp. 1448–1473, 2026.

Hideaki Kim, Tomoharu Iwata, and Akinori Fujino. K2K^2IE: Kernel Method-based Kernel Intensity Estimators for Inhomogeneous Poisson Processes. Proceedings of the 42nd International Conference on Machine Learning, PMLR 267, pp. 30099–30112, 2025.

Gabriel Loaiza-Ganem, Sean M. Perkins, Karen E. Schroeder, Mark M. Churchland, and John P. Cunningham. Deep Random Splines for Point Process Intensity Estimation of Neural Population Data. In Advances in Neural Information Processing Systems, volume 32, 2019.

Håvard Rue, Sara Martino, and Nicolas Chopin. Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations. Journal of the Royal Statistical Society: Series B, 71(2):319–392, 2009.

Reviewer_F4eaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an amortized variational inference framework based on Neural processes that maps replicated point patterns to a variational posterior distribution over their shared intensity.

The main contribution of the paper is the modelling approach to train Temporal Point Processes under the formulation of NPs. The authors demonstrated the effectiveness of the method on synthetic and real-world data experiments.

Critical strengths
  1. The proposed method is a good combination between accuracy and efficiency as it combines low inference cost with accurate intensity recovery and uncertainty quantification.
Critical weaknesses
  1. It would have been good if authors had included as a baseline one of the methods they mention in the introduction which can process multiple point patterns, at least in the computation of certain metrics. Also, for real-world example, it would be good to compute some other metric in addition to HLL, for instance next-event time RMSE or CRPS for the next-event time distribution. In addition, no stds are provided for FIM-PP, PSTH and KDE, which limits the understanding of how superior the proposed method is, especially in comparison to KDE.
Questions for authors

If authors address the limitations mentioned above, I would be willing to increase my score.

4 comments

Dear Reviewer F4ea,

We thank the reviewer for their suggestions of adding baselines, predictive metrics, and variability across data. We believe that these suggestions contribute to a more comprehensive and informative experimental evaluation. Below we respond point by point to each of your suggestions and detail the modifications we will make in the revised manuscript.

1. Comparison with DRS-VAE (Loaiza-Ganem et al., 2019)

We have implemented a baseline method for amortized intensity inference: the Deep Random Spline Variational Auto-encoder (DRS-VAE) using the released architecture and model configuration. DRS-VAE uses a recurrent encoder, spline decoder and standard VAE objective, therefore both its architecture and training scheme significantly differ from our method. We adapted DRS-VAE to infer a shared intensity by pooling the observed realizations and dividing the inferred intensity by the number of replicates. Both methods use the same training, validation and test data, with checkpoints selected by validation held-out likelihood.

We have evaluated the performance of DRS-VAE in our synthetic experiments and on the Divvy dataset, following the experimental settings described in Section 4 of the paper. Recall that we used Ktrain=Ktest=4K_{\mathrm{train}}=K_{\mathrm{test}}=4 for both synthetic and Divvy data. In addition, following the suggestions of other reviewers, we have included further baseline comparisons with fine-tuned FIM-PP (Berghaus et al., 2026), LGCP-INLA (Rue et al., 2009; Bachl et al., 2019), and K2K^2IE (Kim et al., 2025) to provide a more comprehensive evaluation of our method. The table below summarizes the results of all these comparisons.

Synthetic and Divvy results (mean ± standard deviation across four seeds).

MethodSynthetic RMSE ↓Synthetic C95 (%)Synthetic IW95Synthetic time (ms) ↓Divvy HLL ↑Divvy time (ms) ↓
DRS-VAE4.32±0.144.32 \pm 0.1430.74±2.1730.74 \pm 2.173.29±0.123.29 \pm 0.1267±0.0867_{\pm 0.08}103.31±2.77103.31 \pm 2.77253±1.04253_{\pm 1.04}
FIM-PP (pretrained)14.19——659.5376
FIM-PP (fine-tuned)2.78±0.062.78 \pm 0.06——5±0.025_{\pm 0.02}106.28±0.18106.28 \pm 0.1859±0.4659_{\pm 0.46}
LGCP–VI1.86±0.011.86 \pm 0.0192.70±0.1092.70 \pm 0.106.67±0.016.67 \pm 0.016,312±36.066{,}312_{\pm 36.06}108.80±0.01108.80 \pm 0.0124,407±76.8324{,}407_{\pm 76.83}
LGCP–INLA1.5992.135.7438,116110.05161,245
PSTH2.45——<1102.73<1
KDE1.91——<1107.1010
K2K^2IE1.60——3107.496
Ours1.55±0.011.55 \pm 0.0193.28±0.1893.28 \pm 0.185.73±0.005.73 \pm 0.008±0.448_{\pm 0.44}110.99±0.03110.99 \pm 0.0327±0.1327_{\pm 0.13}

We observe that our model achieves much better performance than DRS-VAE (lower RMSE and coverage closer to the nominal level 95% on the synthetic data and higher held-out likelihood on the Divvy data). This might be explained by the differences in architecture, training schemes, and numbers of parameters, however we did not further investigate this comparison due to the time constraints of the rebuttal.

Finally we note that we could not compare to the Neural Diffusion Intensity Model (Du et al., 2026) because its implementation is not publicly available at the time of this rebuttal.

2. Next-event prediction on Divvy

Following your suggestion, we further compared the performance of our method to the baselines by evaluating the next-event prediction error. Specifically we predict each next-event time in the held-out test sequence (day). At each prediction time, all methods use the other three sequences, while FIM-PP additionally conditions on the observed event history of the held-out day. For Bayesian methods, we sample intensity functions from the inferred posterior and average their corresponding next-event time distributions to obtain the posterior predictive distribution. We evaluate the predictive mean of the next-event time using next-event RMSE (NE-RMSE) and the predictive distribution using CRPS. The table below reports the results.

Next-event prediction on Divvy (×10−2\times 10^{-2}; mean ± standard deviation across four seeds).

MethodNE-RMSE ↓CRPS ↓
DRS-VAE18.47±0.3718.47_{\pm 0.37}6.55±0.176.55_{\pm 0.17}
FIM-PP (pretrained)24.288.40
FIM-PP (fine-tuned)22.39±1.3522.39_{\pm 1.35}6.95±0.116.95_{\pm 0.11}
LGCP–VI18.98±0.0118.98_{\pm 0.01}6.59±0.006.59_{\pm 0.00}
LGCP–INLA18.856.51
PSTH18.606.57
KDE18.496.39
K2K^2IE19.096.59
Ours17.91±0.02\mathbf{17.91}_{\pm 0.02}6.23±0.00\mathbf{6.23}_{\pm 0.00}

We observe that our model achieves the lowest mean RMSE and CRPS among the evaluated methods, therefore this comparison further confirms its advantageous performance.

3. Variability across data splits

We thank the reviewer for their question regarding the variability of the deterministic baselines. In our original experiments, We computed the standard deviations of the performance with respect to the different training seeds for a fixed train-test split. Since deterministic baselines do not exhibit variability across training seeds, we have additionally evaluated them across five random train-validation-test splits, both on the synthetic and Divvy datasets.

The table reports the mean and standard deviation of the test metrics across splits.

Results across five data splits (mean ± standard deviation).

MethodSynthetic RMSE ↓Divvy HLL ↑Divvy NE-RMSE ↓Divvy CRPS ↓
FIM-PP (pretrained)12.93±0.2112.93_{\pm 0.21}57.34±9.2057.34_{\pm 9.20}24.53±1.0724.53_{\pm 1.07}8.48±0.408.48_{\pm 0.40}
LGCP–INLA1.58±0.011.58_{\pm 0.01}103.25±12.41103.25_{\pm 12.41}18.70±1.0418.70_{\pm 1.04}6.58±0.406.58_{\pm 0.40}
PSTH2.44±0.032.44_{\pm 0.03}96.88±11.9596.88_{\pm 11.95}18.64±0.9818.64_{\pm 0.98}6.65±0.406.65_{\pm 0.40}
KDE1.91±0.021.91_{\pm 0.02}100.56±12.31100.56_{\pm 12.31}18.42±0.9918.42_{\pm 0.99}6.48±0.396.48_{\pm 0.39}
K2K^2IE1.64±0.031.64_{\pm 0.03}101.12±12.14101.12_{\pm 12.14}18.97±1.0318.97_{\pm 1.03}6.65±0.396.65_{\pm 0.39}
Ours1.54±0.021.54_{\pm 0.02}104.30±12.42104.30_{\pm 12.42}18.07±1.0718.07_{\pm 1.07}6.36±0.396.36_{\pm 0.39}

Across the five data splits, we now observe that our method still achieves the best mean performance on all four metrics. The standard deviations also provide a measure of how the performance of each method varies across different data splits. Synthetic RMSE varies relatively little. On Divvy, the test tasks correspond to different stations, whose event counts and temporal patterns may differ, potentially leading to greater variability in the metrics across splits.

To further assess statistical significance, the following table reports the mean paired differences (Ours minus baseline) and their 95% confidence intervals across five data splits. The differences are statistically significant in nearly all comparisons at the 5% level. The only exception is next-event RMSE compared with KDE, where our method achieves a lower mean RMSE, but the difference is not statistically significant (95% confidence interval: [−0.71,0.01][-0.71, 0.01]).

Differences across five splits (mean with 95% confidence interval).

BaselineSynthetic RMSEDivvy HLLDivvy NE-RMSEDivvy CRPS
FIM-PP (pretrained)−11.39[−11.65,−11.14]-11.39_{[-11.65, -11.14]}46.96[41.50,52.42]46.96_{[41.50, 52.42]}−6.46[−6.76,−6.17]-6.46_{[-6.76, -6.17]}−2.12[−2.20,−2.04]-2.12_{[-2.20, -2.04]}
LGCP–INLA−0.05[−0.06,−0.03]-0.05_{[-0.06, -0.03]}1.05[0.99,1.11]1.05_{[0.99, 1.11]}−0.63[−0.90,−0.37]-0.63_{[-0.90, -0.37]}−0.22[−0.27,−0.18]-0.22_{[-0.27, -0.18]}
PSTH−0.90[−0.94,−0.87]-0.90_{[-0.94, -0.87]}7.42[6.68,8.16]7.42_{[6.68, 8.16]}−0.57[−0.77,−0.38]-0.57_{[-0.77, -0.38]}−0.29[−0.33,−0.26]-0.29_{[-0.33, -0.26]}
KDE−0.38[−0.40,−0.36]-0.38_{[-0.40, -0.36]}3.74[3.58,3.89]3.74_{[3.58, 3.89]}−0.35[−0.71,0.01]-0.35_{[-0.71, 0.01]}−0.12[−0.16,−0.09]-0.12_{[-0.16, -0.09]}
K2K^2IE−0.10[−0.16,−0.05]-0.10_{[-0.16, -0.05]}3.18[2.75,3.62]3.18_{[2.75, 3.62]}−0.90[−1.17,−0.63]-0.90_{[-1.17, -0.63]}−0.30[−0.35,−0.24]-0.30_{[-0.35, -0.24]}

We will include the additional baseline comparisons, next-event prediction results, and analysis across data splits in the revised manuscript.

References

Fabian E. Bachl, Finn Lindgren, David L. Borchers, and Janine B. Illian. inlabru: an R package for Bayesian spatial modelling from ecological survey data. Methods in Ecology and Evolution, 10:760–766, 2019.

David Berghaus, Patrick Seifner, Kostadin Cvejoski, César Ali Ojeda Marin, and Ramses Sanchez. In-Context Learning of Temporal Point Processes with Foundation Inference Models. In International Conference on Learning Representations, pp. 29610–29641, 2026.

Xinlong Du, Harsha Honnappa, and Vinayak Rao. Neural Diffusion Intensity Models for Point Process Data. In Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337, pp. 1448–1473, 2026.

Hideaki Kim, Tomoharu Iwata, and Akinori Fujino. K2K^2IE: Kernel Method-based Kernel Intensity Estimators for Inhomogeneous Poisson Processes. Proceedings of the 42nd International Conference on Machine Learning, PMLR 267, pp. 30099–30112, 2025.

Gabriel Loaiza-Ganem, Sean M. Perkins, Karen E. Schroeder, Mark M. Churchland, and John P. Cunningham. Deep Random Splines for Point Process Intensity Estimation of Neural Population Data. In Advances in Neural Information Processing Systems, volume 32, 2019.

Håvard Rue, Sara Martino, and Nicolas Chopin. Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations. Journal of the Royal Statistical Society: Series B, 71(2):319–392, 2009.

Thank you for the detailed replies! The authors addressed all my points, so I will increase my score.

Dear Reviewer F4ea,

In response to your comments, we have made the following revisions to the manuscript, with changes highlighted in red.

  1. Tables 2 and 4 add comparisons with DRS-VAE, LGCP–INLA, and K2K^2IE, and replace the original FIM-PP baseline with a fine-tuned version.
  2. Section 4.2 and Table 4 add next-event prediction evaluations using NE-RMSE and CRPS.
  3. Appendix E adds evaluations across five data splits to assess the variability of deterministic baselines and the consistency of our method's performance advantages.

We have also updated Figure 3 to provide a clearer visual comparison across methods.

Once again, we sincerely thank you for your constructive suggestions, which have helped us strengthen the experimental comparisons and provide a more comprehensive evaluation of our method on real-world data.

Reviewer_9Uw9OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposed a variational-inference methodology for estimating shared intensity trajectories for i.i.d. observations of temporal point processes. Under ideal assumptions, a Bayesian posterior should maximize their training objective. This provides some theoretical insight into their methodology, though the paper itself did not provide any theoretical guarantees. They ran experiments on both synthetic datasets and a real-world dataset, showing some competency of the proposed model, but did not provide enough justification for why the real-world dataset fit under their shared intensity assumption.

Critical strengths
  1. The proposed methodology is relatively interesting, and shows some competence on synthetic/real-world datasets over other existing approaches.
Critical weaknesses
  1. The authors provide very few theoretical insights, so the contribution is mainly methodological.
  2. Some architectural choices are not clearly justified.
Questions for authors

I enjoyed reading this paper, thanks. Overall, this paper is well-written and the main ideas come through easily. I have two main concerns:

  1. (Shared Intensity Path Assumption) On the Divvy dataset, is it appropriate to assume the intensity path is fixed for a fixed (station, year, month, and weekday) combination? Usually, this is more commonly modeled as sharing some common parameters for some shared distribution of intensity paths.

  2. (Encoder Architecture) For n i.i.d. observations of inhomogeneous Poisson processes, if we let \mathcal{T} be the collection of all event times, then (\mathcal{T}, n) is sufficient for the dataset (context). Therefore, a simplest encoder would just take (\mathcal{T}, n) as inputs (still need some sorts of pooling). In particular, if you consider two point process observations {(t_1, t_2), (t'_1, t'_2)}, then your encoder would differentiate between {(t_1, t_2), (t'_1, t'_2)}, {(t_1, t'_2), (t'_1, t_2)}, and {(t_1, t_2, t'_1), (t'_2)}. However, regrouping the events should not change the Bayesian posterior since the sufficient statistics (\mathcal{T}, n) are preserved. Please provide valid justifications for your proposed encoder architecture, or at least add the simple encoder above as a baseline for comparison.

6 comments

Dear Reviewer 9Uw9,

We sincerely thank you for your positive feedback and constructive suggestions. We are particularly encouraged that you found our work interesting, appreciated the clarity of our presentation, and enjoyed reading the paper.

We have carefully considered your comments, respond to each of your concerns below, and will revise our manuscript accordingly.

1. Shared Intensity Path Assumption

We thank you for your remark and for pushing us to further justify our modeling assumption on the Divvy dataset. In this data, we assume that some sequences are independent observations of the same Poisson process and thus share the same latent intensity function. Our rationale for assuming that sequences corresponding to the same location (station), day of the week, month and year comes from assuming some stationarity of the data generating process over one month, some weekly periodicity (each day having its own pattern), but some differences between years, months and stations.

In fact, we do not assume that this hypothesis holds for the whole dataset, and we selected a subset of this data so that our assumption is plausible, given the available evidence. Specifically, for each group of 4 sequences corresponding to the same value of (station, day, month, year) we ran hypothesis tests where our modeling assumption is the null hypothesis. The tests and screening procedure are described in Appendix D.1, which we will enrich with more details in the revised manuscript. Briefly, we consider two tests: one compares the total number of events across the four days, and the other compares how events are distributed across 10 equal-width time intervals between the four days. For each test, we reject the null hypothesis if the corresponding p-value is <0.01< 0.01. Groups for which neither null hypothesis is rejected are retained for the subsequent steps. While these tests cannot formally validate our assumption even on the selected sequences, our screening procedure excludes the sequences for which this assumption is less plausible.

Your suggestion of modeling each sequence as a single realization of a point process with its own intensity drawn from a shared distribution is a possible alternative approach. While such a hierarchical approach would be interesting to explore, this is not the focus of our paper. Our work focuses on amortized inference of a shared latent intensity from multiple realizations. Our aim is not only to perform amortized inference, but also to quantify uncertainty about the shared intensity as the number of observed realizations changes. Allowing different intensities within each group is a different inference problem, which we do not address with our methodology. In particular, our training scheme requires at least two replicates per task (intensity), since the context and target sets must contain at least one sequence each. Nonetheless we note that at test time, our model can process a single sequence per task and therefore infer one intensity function per test sequence.

However, your comment led us to test an alternative training scheme that does not require multiple realizations with the same intensity. Replacing our objective in Eq. (19) by the standard ELBO in Eq. (15), we can then train our encoder-decoder model treating each training sequence as having its own latent intensity.

For each test sequence, both models use its events to infer an intensity and compute the log-likelihood as in Eq. (22). Here, the log-likelihood is evaluated on the input sequence rather than on a held-out sequence.

We compared our model trained either with this alternative scheme or our original scheme on the Divvy data. In the table below, we report the test log-likelihood. We observe that our model achieves a slightly higher log-likelihood in this comparison, suggesting that our assumption does not adversely affect performance in this experiment.

Divvy log-likelihood (mean ± standard deviation across four seeds).

MethodLog-likelihood ↑
Single-sequence VAE (standard ELBO)113.09 ± 0.07
Ours113.94 ± 0.04

2. Encoder Architecture

It is absolutely correct that, for independent realizations of an inhomogeneous Poisson process sharing the same latent intensity function, the collection of all event times together with the number of realizations is sufficient for posterior inference.

However we first recall that our encoder-decoder model does not compute the exact posterior, rather an approximation which in fact depends on the architectural choices. A different encoder may thus lead to a different approximation, and below we provide some comparison of these two possible design choices.

Secondly, we also recall that our encoder architecture is motivated by the Neural Process, which models a distribution over functions by encoding and aggregating context observations in a permutation-invariant way. In our setting, each point process realization provides an independent observation of the same latent intensity function, making it a natural choice for a context observation. Furthermore, realization-level features, such as event counts, inter-event intervals, and temporal patterns, may be difficult to represent effectively by directly aggregating event-level embeddings. By encoding each realization separately before aggregation, our encoder explicitly captures both the average features of individual realizations and their variability across realizations. Although these features do not provide additional information for posterior inference beyond the sufficient statistics, explicitly representing them may help the model learn a more effective approximation to the posterior distribution over the shared intensity.

Nonetheless, other encoder designs are indeed possible.

We followed the reviewer's suggestion and evaluated a simpler encoder that pools the event embeddings from all observed sequences. This task embedding also incorporates the total event count and the number of sequences.

Specifically, with KK the number of realizations, NkN_k the number of events in realization kk, N=∑k=1KNkN=\sum_{k=1}^{K}N_k the total event count and hkih_{ki} is the embedding the iith event of the kk-sequence, the task embedding in this simpler encoder is defined as:

\log(1+N),\ \log K\right].$$ To compare our encoder and this simpler version, we train both on the simulated data described in Section 4.1 with $K_{\mathrm{train}}=4$. The synthetic results table below compares their performance on the same test tasks. **Synthetic results (mean ± standard deviation across four seeds).** | $K_{\mathrm{test}}$ | $\mathrm{RMSE}\downarrow$ | | $\mathrm{C95}(\\%)$ | | $\mathrm{IW95}$ | | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | | Ours | Pooled | Ours | Pooled | Ours | Pooled | | $1$ | $2.68_{\pm 0.01}$ | $2.67_{\pm 0.00}$ | $92.21_{\pm 0.05}$ | $92.12_{\pm 0.19}$ | $9.36_{\pm 0.02}$ | $9.32_{\pm 0.05}$ | | $2$ | $1.99_{\pm 0.00}$ | $1.99_{\pm 0.00}$ | $93.18_{\pm 0.09}$ | $93.10_{\pm 0.06}$ | $7.20_{\pm 0.02}$ | $7.18_{\pm 0.04}$ | | $4$ | $1.55_{\pm 0.01}$ | $1.55_{\pm 0.00}$ | $93.28_{\pm 0.18}$ | $93.27_{\pm 0.05}$ | $5.73_{\pm 0.00}$ | $5.72_{\pm 0.01}$ | | $8$ | $1.23_{\pm 0.01}$ | $1.23_{\pm 0.01}$ | $93.78_{\pm 0.25}$ | $94.01_{\pm 0.15}$ | $4.78_{\pm 0.01}$ | $4.80_{\pm 0.03}$ | | $16$ | $1.00_{\pm 0.01}$ | $1.02_{\pm 0.01}$ | $94.32_{\pm 0.28}$ | $94.37_{\pm 0.14}$ | $4.23_{\pm 0.01}$ | $4.28_{\pm 0.05}$ | | $32$ | $0.88_{\pm 0.01}$ | $0.95_{\pm 0.01}$ | $93.84_{\pm 0.21}$ | $93.85_{\pm 0.35}$ | $3.91_{\pm 0.02}$ | $3.94_{\pm 0.06}$ | | $64$ | $0.82_{\pm 0.02}$ | $0.87_{\pm 0.01}$ | $93.16_{\pm 0.30}$ | $93.74_{\pm 0.30}$ | $3.74_{\pm 0.03}$ | $3.80_{\pm 0.07}$ | We observe that both encoders perform similarly for $K_{\mathrm{test}}\leq 8$. For larger realization counts unseen during training, our encoder achieves lower RMSE, particularly at $K_{\mathrm{test}}=32$ and $64$, which may indicate that our encoder generalizes better to unseen numbers of replicates, while C95 and IW95 remain comparable. Moreover on the Divvy datasets, the two encoders achieve comparable held-out log-likelihood (HLL), as shown in the Divvy results table. **Divvy results (mean ± standard deviation across four seeds).** | Method | $\mathrm{HLL}\uparrow$ | $\text{Time (ms)}\downarrow$ | | :--- | ---: | ---: | | Pooled | $110.99 \pm 0.02$ | $23.82 \pm 0.07$ | | **Ours** | $110.99 \pm 0.03$ | $26.82 \pm 0.13$ | We will further clarify the motivation for the shared-intensity assumption in the revised manuscript, provide additional details of the statistical tests and dataset construction in Appendix D.1, and include the pooled encoder comparison and its experimental results in the appendix.

Thank you for the detailed replies to all my questions.

  1. (Shared Intensity Path Assumption) Selecting a subset of the data provides some justification of the model, but still does not provide a convincing rationale. That being said, I understand this is a modeling choice, which is more of a question on the assumptions rather than on the methodology itself. I also greatly appreciate your experiment with the alternative training scheme, which shows the flexibility of the model to relax the single-shared-intensity assumption.

  2. (Encoder Architecture) Thanks for the additional comparisons between your proposed encoder architecture and the baseline architecture. However, I find it hard to convince myself the advantage of using your proposed encoder architecture given i) its marginal performance over the baseline ii) failure of invariance on regroupings of events (as in my original review). Your reply is nevertheless greatly appreciated.

I will wait until all the questions raised by the reviewers have been resolved to adjust my rating. Thanks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.