Pseudo-Likelihood Inference

Simulation-Based Inference (SBI) is a common name for an emerging family of approaches that infer the model parameters when the likelihood is intractable. Existing SBI methods either approximate the likelihood, such as Approximate Bayesian Computation (ABC) or directly model the posterior, such as Sequential Neural Posterior Estimation (SNPE). While ABC is efficient on low-dimensional problems, on higher-dimensional tasks, it is generally outperformed by SNPE, which leverages function approximation. In this paper, we propose Pseudo-Likelihood Inference (PLI), a new method that brings neural approximation into ABC, making it competitive on challenging Bayesian system identification tasks. By utilizing integral probability metrics, we introduce a smooth likelihood kernel with an adaptive bandwidth that is updated based on information-theoretic trust regions. Thanks to this formulation, our method (i) allows for optimizing neural posteriors via gradient descent, (ii) does not rely on summary statistics, and (iii) enables multiple observations as input. In comparison to SNPE, it leads to improved performance when more data is available. The effectiveness of PLI is evaluated on four classical SBI benchmark tasks and on a highly dynamic physical system, showing particular advantages on stochastic simulations and multi-modal posterior landscapes.

Paper

References (52)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 7dVy6/10 · confidence 4/52023-07-04

Summary

This method perform simulation-based inference with a hybrid of previous methods, in this case integrating a neural likelihood-type step within a SMC-ABC algorithm for performing posterior inference. It considers use of distance metrics such as the Wasserstein and the Maximum Mean Discrepancy to compare simulated and observed data, while performing particle-based updates in a fairly standard way for SMC-type methods. It performs a comparison on five different test examples, with comparisons with different ABC methods and a neural likelihood method.

Strengths

The method is fairly well presented, at least the steps setting up the framework of the algorithm used. The examples used are appropriate, albeit limited to low-dimensional parameter spaces. The Furuta pendulum, which I had not seen before, is particularly nice. You have chosen a few competing methods which, by no means exhaustive, at least make similar assumptions to PLI and therefore the comparison is revealing. “Related Work” is fairly thorough.

Weaknesses

There are a few little typoes in the language that I noticed: "Posteriror” “numerica”. I imagine there are more that a standard spellcheck would help with. The actual algorithm and computational steps used to estimate the relevant divergences could be much more thoroughly explained, i.e. “Compute IPM” without saying how to for MMD vs Wasserstein vs other is unclear. You provide a reference for this estimation but I think it should be better explained in this article itself. I’m not sure that I really buy the analogy with variational inference: the proposal distribution formulation built with weights is fairly standard from the SMC literature, and I don’t think the variational angle is particularly clear. I imagine that you do have a legitimate point to make in this respect, so if you could expand your reasoning on this line then that would be a good thing. A few specific claims that I didn't really agree with: “There might be several parameter configurations yielding the same observation, hence rendering the resulting distribution to be multi-modal.” - Multi-modality is not always a result of non-identifiability, just different parameter values having comparable likelihoods, not necessarily the same observed data. “KL divergence is intractable in SBI because we cannot evaluate the likelihood” - can be evaluated without the likelihood with density ratio estimation and classifiers as in classifier ABC and f-divergences. FIND CITATIONS. The term “likelihood kernel” is introduced without explanation.

Questions

The Furuta pendulum is a nice example. Are there other examples of using chaotic systems rather than stochastic systems as test cases for simulator based inference? Do you treat the variation between two adjacent deterministic chaotic systems as being basically random, and is there a downside to making that assumption? Is “APT” in the results section basically targeting the posterior directly without the SMC-type exploration? Why does it seem to do well up to a given value of N then fall off quite extremely? Is a more concentrated posterior (larger N) more difficult to explore, and the SMC-type method helps with exploration of a complex posterior? Constructing a neural likelihood approximation at every iteration must add quite a large computational overhead that should be acknowledged. What is the relevance of model misspecification here? This is acknowledged in the section describing General Variational Inference, and it would be nice to see at least a small discussion of the possible advantages of using the MMD or Wasserstein compared to the KL for the “true” likelihood. What is the expected scaling with dimensionality of the parameter space? You have constrained yourself to low-dimensional examples in this paper. The density-ratio type methods tend to do quite well in higher-dimensional spaces as neural nets can fit in this spaces with enough data, but SMC suffers quite severely with the curse of dimensionality. Does the use of neural networks in PLI help overcome this curse? If yes, great! If no, is it worth the overhead of training a neural net in a lower-dimensional space when there might exist other methods that do the job adequately? Would the straight APT method do better than any of the SMC-type methods in a higher-dimensional space?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No clear immediate social impact of this work. It really depends on the model on which inference is being performed.

Reviewer dHap6/10 · confidence 3/52023-07-05

Summary

This paper proposes a Pseudo-Likelihood Inference approach that brings neural approximation into ABC. This new approach can condition on a variable number of observations unlike previous neural-based SBI approaches.

Strengths

* The incorporation of Sequential Monte Carlo into SBI is a strength. SMC seems like it should be related to SBI as it follows the structure (as shown in Fig. 1) of annealing to the desired distribution. * The literature review is a strength and highlights the authors' knowledge on the topic. * The paper is well-written.

Weaknesses

* The experiments section could be strengthened to highlight the advantage of the approach compared to SNPE. For example, is the a better way to highlight that $N \not= M$ for PLI? It seems a shame not to highlight this strength of the new approach. Also, does $N=M$ mean that for APT, when $N=1$, only one training point is used to train the normalizing flow? * In Lueckmann et al. (2021) there are actually 10 benchmark experiements. Is there a reason that only four were used here? Additionally, if “benchmarking is not the focus” (line 247) then what is the focus? This motivation could be clearer. * The computational cost is not highlighted in the paper. Is the cost of retraining the density estimator at each step equivalent to APT? * It is a weakness that the code was not provided.

Questions

* The hyperparameters seem to be set without mention of any hyperparameter optimization. What effect does the number of particles have and the number of steps have on the overall performance? What do bins mean on the Supplementary materials?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

* The idea appears novel and is the main reason for the slightly positive score. The experimentation could definitely be improved, as well as the explanation behind various hyperparameters and what the limitation of the approach is when it comes to computational cost.

Reviewer 9BA66/10 · confidence 3/52023-07-06

Summary

This paper proposes an approch to deal with posterior inference when the likelihood is intractable, which is often the case for problems involving simulators of physical processes. The paper investigates Approximate Bayesian Computations (ABC) methods, and it proposes incorporating neural approximations through variational principles. The main strenghts of the proposed approach include the possibility to carry out optimization of neural posteriors via gradient descent, the lack of reliance on summary statistics, and the possibility to handle multiple observations as input.

Strengths

The paper is well written and easy to follow. The methodological developments seem solid and well motivated. Experiments seem to support the main message of the paper.

Weaknesses

Experiments could have included a larger variety of scenarios involving other types of data and simulators.

Questions

As a non-expert on ABC I found the paper interesting and well written, although I find it difficult to assess the novelty of this work. Replacing the KL with some objective that can be estimated through samples is a good idea in general, and this paper does it in eq. 6; not sure how novel the ABC community would consider this. Similarly, moving away from modeling the posterior through particles in favor of neural density estimators yields clear advantages, and again I'm not sure how novel this would be considered in this literature. Having said that, I found the motivation behind the work convincing and the idea well executed with experiments showing improvements over alternative approaches from the state-of-the-art. As a result, I have an overall positive opinion of the paper which combines sensible ideas to produce an approach which seems to work well across a number of experiments.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

I haven't seen any specific text on the limitations of the approach and I think this could be included in the revision.

Reviewer AzJv6/10 · confidence 2/52023-07-06

Summary

This paper introduces Pseudo-Likelihood Inference (PLI) as a novel method inside the simulation-based inference (SBI) family. SBI methods are used to estimate model parameters when the likelihood function is unknown or intractable. Existing SBI methods, such as Approximate Bayesian Computation (ABC) and Sequential Neural Posterior Estimation (SNPE), either approximate the likelihood or directly model the posterior. The contribution by the authors, PLI, combines neural approximation with ABC to achieve competitive performance in challenging Bayesian system identification tasks. The authors make use of integral probability metrics and introduce a smooth likelihood kernel with an adaptive bandwidth that is updated based on information-theoretic trust regions. This formulation allows for optimizing neural posteriors via gradient descent and does not rely on summary statistics. PLI performance improves the more data is available, particularly in scenarios involving stochastic simulations and multi-modal posterior landscapes. The effectiveness of PLI is evaluated through experiments conducted on four classical SBI benchmark tasks as well as a highly dynamic physical system. The results demonstrated particular advantages of PLI in handling stochastic simulations and multi-modal posterior landscapes.

Strengths

* PLI is a novel approach addressing the limitations of existing methods that achieves competitive performance in challenging Bayesian system identification tasks. In fact, the proposed method outperforms SNPEs and other likelihood-free models, particularly when more data is available. * The usage of the adaptive bandwidth for the smooth likelihood kernel allows for better adaptation to different problem settings and improves the accuracy of parameter estimation. * The method enables to directly use gradient descent optimizing neural posteriors using gradient descent directly on raw observations or data points without requiring any intermediate summarization step. * The effectiveness of PLI is evaluated in multiple experiments conducted on four classical Simulation-Based Inference benchmark tasks and in a highly dynamic physical system, demonstrating its applicability across various domains.

Weaknesses

* The changes induced by replacing the KL divergence with the MMD and the Wasserstein distance should further explored. Consequences on changing the KL divergence for other metrics has been studied before in works such as [1], where an adversarial approach is used to evaluate the KL divergence and other divergences using only samples. I strongly suggest discussing this possibility in the article, since it seems that this alternative can be useful here. * To provide more robust experimental results, when comparing different models I would suggest also using other proper scoring rules [2]. * Since efficiency is an important aspect when dealing with high-dimensional tasks, there is limited discussion about computational efficiency throughout the paper. Considering that likelihood-free inference methods can be computationally demanding, it would be valuable to discuss any potential limitations or considerations regarding computational resources required by PLI compared to other SBI approaches * While the paper highlights PLI's advantages over SNPEs, it would be valuable to understand how PLI performs against other popular approaches s.a. Markov Chain Monte Carlo methods. A comparative analysis would provide a more comprehensive understanding of PLI's strengths and weaknesses relative to existing techniques. * (_minor_) Due to the formulation of several function-space Bayesian inference models, if you deemed it fitting it could be interesting discussion on the function-space approaches such as [3-6]. I think most of these models can be encompassed as (S)NPE methods and that could enrich the discussion in the paper. * (_minor_) Please correct typos, s.a. "posteriror" in line 143, "divergnce" in line 156. * (_minor_) The red hyperlink boxes to equations appear with an extra line inside, which difficults reading. Please check this. --- **References**: [1] Santana, S. R., & Hernández-Lobato, D. (2022). Adversarial α-divergence minimization for Bayesian approximate inference. Neurocomputing, 471, 260-274. [2] Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association, 102(477), 359-378. [3] Ma, C., & Hernández-Lobato, J. M. (2021). Functional variational inference based on stochastic process generators. Advances in Neural Information Processing Systems, 34, 21795-21807. [4] Rodrı́guez Santana, S., Zaldivar, B., & Hernandez-Lobato, D. (2022, June). Function-space Inference with Sparse Implicit Processes. In International Conference on Machine Learning (pp. 18723-18740). PMLR. [5] Ma, C., Li, Y., and Hernández-Lobato, J. M. (2019). “Variational implicit processes”. In: International Conference on Machine Learning, pp. 4222–4233. [6] Sun, S., Zhang, G., Shi, J., and Grosse, R. (2019). “Functional variational Bayesian neural networks”. In: International Conference on Learning Representations.

Questions

* What are the limitations and potential challenges associated with implementing PLI? The paper presents PLI as an effective method for challenging Bayesian system identification tasks, but it is important to address potential limitations or practical considerations when applying this approach in real-world scenarios. For example, what computational resources are required by PLI? Are there specific problem domains where its performance may be less favorable?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

* The paper could improve by comparing PLI with more SBI methods. * The evaluation focuses on a limited set of benchmark tasks and one physical system, potentially limiting the generalizability of PLI to diverse problem domains. * There is insufficient discussion about the computational efficiency and resource requirements of implementing PLI.

Reviewer XuP46/10 · confidence 3/52023-07-06

Summary

This work aims at proposing a sequential Monte Carlo scheme to perform approximate Bayesian computation in the context of simulation based inference. Previous works focused on directly modeling posterior distributions whereas this approach presents a novel method to approximate likelihoods utilizing kernel density methods along with an information theoretic kernel bandwidth update scheme. The authors claim that this direction allows for improvement over previous works in the case of having more reference data to learn from.

Strengths

I think that employing variational inference to help make SMC-ABC more tractable is an interesting direction. Further leveraging this setting to derive a trust-region compliant kernel bandwidth is significant and a good direction to explore. A number of different non-trivial steps are performed in order to derive the final pseudo-likelihood inference scheme and I believe this deserves commendation.

Weaknesses

Almost all of the derivations are based on the classic variational inference objective of reducing the KL divergence between a true distribution and the variational distribution being learned; however, the actual method proposed substitutes this for alternative divergences/"integrated probability metrics" such as Wasserstein distance and MMD. I understand the necessity of this due to the likelihood being unavailable in SBI; however, this does undercut a lot of the impact that the various derivations have on the actual problem in my opinion. I do not find the experimental results as convincing as they could be. For the first two benchmarks (Gaussian Location and Gaussian Mixture), it appears that MMD-ABC is either comparable to both versions this paper is proposing or better across all values of N and in both metrics. Additionally, in the appendix it can be seen in Figure C.1 that using sample based metrics (referred to as PPCs in the paper) the PLI methods appear to perform about the same or worse in several cases in all 5 tasks. I wouldn't go as far to say that they are always comparable or worse, but the story does not seem as clear cut as it is made out to be in the main set of results in Figure 3 in the main paper. Note that in line 285 it is justified to use PPCs instead of comparing posterior distributions, so I believe it should be valid to look at these results for the benchmarks as well. Please feel free to correct me if I have misunderstood something as I could have easily missed some key details.

Questions

Is there a clear explanation as for why the APT method starts to perform worse when given more reference data? Naively I would think that it would plateau in performance but it appears to severely degrade. We see somewhat similar behavior in some of the settings with both PLI and ABC methods, although to a much lesser degree (and more infrequent). I would be interested in learning if there is any insight into why this could be happening to all of these methods.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The authors did adequately discuss the limitations in my opinion.

Reviewer AzJv2023-08-12

Brief response

First of all, I would like to thank the authors for addressing in detail the points raised in all reviews. I have read through it all and I must say I am quite satisfied by the responses overall, even more given all the updates and extra information included in the final version of the paper and the fact that the authors plan to release the code as well. I am inclined to maintain my rating (at least for the given moment), although I am in fact more favourable now to the acceptance of the paper than I was earlier. Thanks again for all the work!

Authorsrebuttal2023-08-14

Thank you again for the time and investment in reviewing our proposed method. We are glad that the additional information helped clarify the remaining questions.

Reviewer 7dVy2023-08-14

Thank you for the informative rebuttal. Regarding the APT falling off with larger N: is it really falling off with larger p, and we have a correlation in the examples with number of data points and dimensionality? Are there any results which can make clear that this is really the effect here, i.e. a large N small p example, or small N and large p? Is there good theoretical reasons why to expect a scaling with either N or p for APT, and can you make that explicit in the text?

Authorsrebuttal2023-08-14

Thank you again for the time and investment in reviewing our proposed method. Here we attempt to answer your additional questions: There have been no comprehensive evaluations of the common SBI benchmark methods on problems with high-dimensional input spaces in the literature (i.e., for large N). Similarly, theoretical analysis of the scaling of such methods is also lacking. Regarding increasing p (which we assume you mean the number of training samples), we expect APT to perform well for high N. Nevertheless, it's important to highlight that this is accompanied by a higher sample complexity, while a primary objective of PLI is to enhance sample efficiency. Consequently, we emphasize evaluating its performance within the constraints of a small, fixed simulation budget. Under these conditions, we observe that PLI surpasses APT on tasks with increased N. To better understand the scaling of PLI and APT, we agree that performing ablation on p and N would be interesting, which we will include in the final paper.

Reviewer 7dVy2023-08-16

I actually mean the precise opposite of what you have assumed: using N to mean number of samples and p to mean dimensionality is totally standard within the statistics community. What do you mean in the light of this?

Authorsrebuttal2023-08-18

Thank you for providing additional clarity. We approached the question with respect to our paper's context, where N represents the number of observations utilized for inference. In order to eliminate any potential confusion, we will adopt the suggested notation in our forthcoming response: - p: Dimensionality of the data, indicating the number of observations multiplied by the dimension of each observation - N: Count of training samples Apart from interchanging the symbols p and N, the core message of our original response remains unaltered. Kindly refer to our revised answer below: There have been no comprehensive evaluations of the common SBI benchmark methods on problems with high-dimensional input spaces in the literature (i.e., for large p). Similarly, theoretical analysis of the scaling of such methods is also lacking. Regarding increasing N, we expect APT to perform well for high p. Nevertheless, it's important to highlight that this is accompanied by a higher sample complexity, while a primary objective of PLI is to enhance sample efficiency. Consequently, we emphasize evaluating its performance within the constraints of a small, fixed simulation budget. Under these conditions, we observe that PLI surpasses APT on tasks with increased p. To better understand the scaling of PLI and APT, we agree that performing ablation on p and N would be interesting, which we will include in the final paper.

Reviewer dHap2023-08-14

Response

Thanks for the clarifications. It is still not great that the code was not provided with the submission as that is normally a key part of the work and does not help with assessing reproducibility. However, I will raise my score since the authors did provide a response to every weakness that I highlighted. Thanks for your effort.

Authorsrebuttal2023-08-18

Thank you again for the time and investment in reviewing and acknowledging our proposed method.

Reviewer XuP42023-08-14

Reply to Authors

Thank you for the in depth reply to my questions and concerns. I still think the connections between the theorems derived and the actual proposal with the chosen divergence criteria is a bit tenuous, even with the added empirical results, and could benefit from some added analysis or slightly stronger ties; however, in light of my other points of interested being addressed I have decided to raise my score from a 5 to a 6.

Area Chair ZTmw2023-08-15

Author-reviewer discussion

Dear all, The author-reviewer discussion period has now started. It will continue for one more week, until August 21. @authors: Please respond to the comments or questions reviewers may further have. Remain short and to the point. @reviewers: Please read the author's responses and ask any further questions you may have. To facilitate the decision by the end of the process, please also acknowledge that you have read the responses and indicate whether you want to update your evaluation. - You can update your evaluation positively (if you are satisfied with the responses) or negatively (if you are not satisfied with the responses or share other reviewers' concerns). Please note that major changes are a reason for rejection. - You can also keep your evaluation unchanged. In this case, please indicate that you have read the responses and that you do not have any further comments. Best regards, The AC

Reviewer 9BA62023-08-16

Many thanks for your clarifications and for summarizing the novel aspects charaterizing your work - no need for further clarifications at this stage.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC