Learning Robust Statistics for Simulation-based Inference under Model Misspecification

Simulation-based inference (SBI) methods such as approximate Bayesian computation (ABC), synthetic likelihood, and neural posterior estimation (NPE) rely on simulating statistics to infer parameters of intractable likelihood models. However, such methods are known to yield untrustworthy and misleading inference outcomes under model misspecification, thus hindering their widespread applicability. In this work, we propose the first general approach to handle model misspecification that works across different classes of SBI methods. Leveraging the fact that the choice of statistics determines the degree of misspecification in SBI, we introduce a regularized loss function that penalises those statistics that increase the mismatch between the data and the model. Taking NPE and ABC as use cases, we demonstrate the superior performance of our method on high-dimensional time-series models that are artificially misspecified. We also apply our method to real data from the field of radio propagation where the model is known to be misspecified. We show empirically that the method yields robust inference in misspecified scenarios, whilst still being accurate when the model is well-specified.

Paper

References (87)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 1V4M5/10 · confidence 3/52023-06-30

Summary

This paper proposes a general approach to handling model misspecification for SBI. The paper introduces a regularized loss function that penalises mismatches between the observed data and the learned model. The paper focuses on NPE and ABC as the likelihood-free approaches.

Strengths

* The radio propagation example is interesting, and shows what the potential strength of using the MMD regularizer. * The structure of the paper is clear.

Weaknesses

* The novelty of the work is a concern in light of reference [72] equation (9), compared to equation (9) in this paper. Additionally, equation (10) appears similar to the InfoVAE (https://arxiv.org/abs/1706.02262), which also uses an auto-encoder with a MMD loss in the latent space. * A minor weakness is that the paper is difficult to follow. While the structure made sense, the descriptions of the approach and the motivation for it were challenging to understand. For example, it seemed implicit that this approach did not perform amortised NPE until it was highlighted this was the case in the conclusion. This is not in itself a criticism of the approach, but it was very hard to follow what consisted of the training data, and observations. For example, line 255 mentions using 1000 samples for the training data and 100 realisations of both the observed and simulated data for each $\theta$. Under an amortised setting we would not have access to the observed $\theta$. It is not clear what this setup is from the description. Further questions in the next section highlight some of these confusions that likely stem from descriptions of the approach that could be improved. * An additional minor weakness is only comparing to ABC and NPE. Is there any limitation on applying this approach to other estimators such as Neural Likelihood Estimators? Also for ABC, the $\rho$ was only described in the supplementary materials as the Euclidean distance. After defining the MMD as the better regularizer, why would this not also be used/incorporated into the discrepancy? Typo: line 180, the $i$ should be subscript? * In Figure 5, it would be useful to ensure the three plots share the same y axis.

Questions

* Could the authors provide more details about the step in Eq. (4) to go from infimum to upper bound? I found this challenging to understand. * The motivation behind defining a Q and a P that are different is a bit confusing. Why were these definitions highlighted? It seems that Q is equal to P when using the $\theta_{true}$, but is it a requirement to know $\theta_{true}$ prior to the experiments in order to simulate from Q? * The last sentence of section 5 highlights that the MMD is superior for the NPE-RS method. Is this surprising given it is built into the loss function? Are there other statistics that could be used here that are specific to the application domain?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Limitations

This is well captured by the paper. This is appreciated!

Reviewer KhFo5/10 · confidence 3/52023-07-02

Summary

The authors propose a method to make neural posterior estimation robust to misspecification. The method relies on adding a regularizer to NPE loss function. The regularizer is implemented as the MMD between the embedding of the approximate (learned) posterior and the embedded observation. They apply their method to two low-dimensional benchmark tasks and a real-world example from radio propagation.

Strengths

**Originality**: The paper tackles an important issue: how to make neural simulation-based inference robust to misspecification. The method is novel and intuitive. **Quality**: The method is applied to benchmark tasks and to a real-world problem (with iid data samples and high-dimensional data). **Clarity**: The paper is well-written and easy-to-follow. Figures are intuitive and well-designed.

Weaknesses

My main concerns with the paper are that it (1) uses only two very low-dimensional benchmark tasks and (2) overstates it contributions. **Originality**: - the method is a very straightforward extension to Schmitt et al 2021 (which is only cited with a passing reference). The authors should clearly discuss how the works are related. **Quality**: - The authors repeatedly emphasize that their method performs on par with NPE for well-specified models. I do not think that their results support this claim though. NPE-RS performs significantly worse than NPE on well-specified data on one of the (only) two benchmark tasks. In order to make this claim, the authors should use significantly more benchmark tasks, ideally with varying parameter and data-dimensionality, and demonstrate that the behaviour shown in figure 2, left is a rare exception. In addition, the authors would have to show that NPE-RS converges (at least very closely) to the true posterior with many simulations (see below). - Convergence: The method proposed by the authors does no longer converge to the true posterior distribution for well-specified models, which further emphasizes that the method does not replace detection of misspecification. I would appreciate an investigation of how regularization strenght trades-off performance on well-specified data vs robustness to misspecification. - how is the hyperparameter chosen? The authors claim that the regularizer can be selected with a validation set (L353). It is unclear to me how this would work. What would be the loss function used to assess performance on the validation set? In L231 the authors also say that this might be done based on the posterior predictive distribution, but I think this is tricky because it (1) requires (potentially many) more simulations and (2) can easily lead to pathological cases where the posterior is off but the predictive distribution is good. Please elaborate on how the hyperparameter should be set in practice.

Questions

The authors claim that RNPE is not amortized. However, as far as I understand, RNPE does not require retraining for new data (yes, it requires MCMC, but this can be very fast, especially for low-d parameter space). Please clarify if I misunderstood this or clarify in your paper.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Presentation

4 excellent

Contribution

3 good

Limitations

State more explicitly that the method does **not** converge to the true posterior.

Reviewer EXCq6/10 · confidence 3/52023-07-04

Summary

This paper describes a method to perform simulation based inference, focusing on neural posterior estimation and approximate Bayesian computation (ABC), under model miss specification when performing inference using summary statistics of the data by using an MMD loss between the simulated and observed data summaries as a means to regulate miss specification. Experiments are performed to empirically examine the efficacy of the method, and the method is compared to standard neural posterior estimation and ABC, as well as a method called robust neural posterior estimation.

Strengths

The paper is tackling an important issue, the method is interesting, and the description of the method and its motivation are made clear.

Weaknesses

It is not clear how this work differs from ref [72], Schmitt et. al, "Detecting Model Misspecification in Amortized Bayesian Inference with Neural Networks". Much clearer description of the differences is needed to understand the novelty of this work It is not clear how to determine the amount of regularization, i.e. how to set the hyperparameter lambda. This seems to be a key missing piece of information on how to practically use this method. The paper seems to indicate that the summarizer acts on sets of iid samples. But is it not clear why must the summarizer act on the set of x's, rather than summarizing each observation, and using the fact that the summaries will remain iid. This seems to come back as a constraint later in line 179, but does not seem well motivated, or practically how people use summary statistics for iid data. This could also significantly impact the ability to train a conditional normalizing flow due to the massive reduction in information by summarizing over a set of examples rather that summarizing per example. In the experiments, it is not clear what summary statistic is used for RNPE? Summary statistics are not described in the RNPE paper, but rather attempting to model the p(theta|x), so how was this choice made? Does it affect the results? Moreover, in the experiments, why not compare to the method in ref [72]? In terms of reference, there are also methods to learn robust summaries prior to inference, e.g. using pivots, such as in Louppe, et. al, "Learning to pivot with adversarial networks", and similar domain adaptation approaches, some specifically using MMD.

Questions

How does this work differ from ref [72] ? Can you provide a much clearer description of the differences is needed to understand the novelty of this work. Can you compare to this work? On line 257, is is stated that lambda is set using simulations using theta_true. What does it mean that you set lambda using a data set with known theta_true? isn't this something that needs to be estimated? Doesn't this greatly reduce the challenge of inference, and this information is not realistically available? Why does the summarizer act on on sets of iid data, rather than on each iid data example? Acting on each example individually seems to be much closer to practical usage. Do you have experiments in this setting? Would this significantly impact the computation driven by MMD, it seems this may make the method impractical? Please provide more details on how RPNE was used in the experiments, and how the summary statistics were determined.

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors adequately discuss limitations

Reviewer DQRY4/10 · confidence 1/52023-07-05

Summary

The paper addresses the "data selection problem", i.e. identify a low-dimensional statistic (e.g. mean and variance for Gaussian distribution) of a high-dimensional dataset for which the model can replicate even when misspecified. The paper propose using the auto-encoding framework to automatically extract statistics and a penalty on mismatched statistics, i.e. statistics that the model is unable to replicate, as a general approach to handle model misspecification. Empirical results show robust inference in misspecified scenarios whilst still being accurate in well-specified scenarios.

Strengths

1. The paper introduces the "data selection problem". Sorry, I am not a subject matter expert in the field of "data selection" thus I cannot discuss on the originality and significance of this work. However, I learnt about the data selection problem from the paper.

Weaknesses

1. The ideas of auto-encoding, reducing mismatch and tuning the regularizer are not new, but they are perhaps new in the application to "data selection". Please discuss further. 2. From the paper, since it is an auto-encoding framework on the model, the statistics are informative of the model. On the other hand, how to ensure the statistics are informative of the observations? Please discuss further.

Questions

1. Since the extracted statistics are changing/ unknown beforehand, how will these statistics fit into downstream applications for downstream applications that were fixed beforehand? Perhaps the downstream tasks are retrained on the new robust statistics. 2. In the experiment implementation, should the model be the misspecified case while the observations be the true case? Or both can have some degree of misspecification while we try to extract the true parameters?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

1. Authors mentioned the limitation of using the observed statistic during the training procedure and recommended working on it as future work.

Reviewer KhFo2023-08-13

Response to rebuttal

Thanks for the detailed response and the additional simulations. They cleared some of my concerns, but some points are still unclear/unresolved to me: **Low-dimensionality:** I indeed had meant low-d theta. I agree that high-D x are relevant and important, but I still think that the paper would be much stronger if the authors demonstrated that the method scales to parameter spaces with dimensionality >> 2 **NPE-RS performs significantly worse than NPE**: I still think that the paper downplays this limitation. The empirical difference between NPE and NPE-RS is very clearly noticable, also in the new results (in addition to NPE-RS not converging to the true posterior). In my opinion, claims like `It provides accurate inference results even when the model is well-specified` (L 67) are not warranted and have to be removed or significantly trimmed down. **trade-off between performance on well-specified data vs robustness to misspecification**: Which of the three metrics shown in figure 5 should correspond exactly to "robustness to misspecification" and why is this warranted? Wouldn't negative log-likelihood on misspecified x be a better measure? **Hyperparameter choice**: The authors say that it can be set with a `held-out validation dataset (which can be a subset of the observed data)`. Does this mean that using cross-validation requires several datasets of observed data (which are potentially misspecified and can not easily be synthetically generated) in order to choose $\lambda$? What if only one dataset is available at training time?

Authorsrebuttal2023-08-16

Thank you for your response. We really appreciate it. * **"...I still think that the paper would be much stronger if the authors demonstrated that the method scales to parameter spaces with dimensionality $>>$ 2":** Agreed. Having an example with more than 4 parameters (as is the case with the Turin model in Section 5) can further strengthen our paper. To that end, we ran our method on the 10-dimensional Gaussian linear example with fixed covariance matrix $\Sigma$ (parameter of interest is the mean vector) used in the RNPE paper and also available in the SBI benchmark library. To make the model misspecified, we used the same contamination model for the observed data used in the paper, i.e., $\mathbb{Q} = (1-\epsilon)\mathcal{N}(\theta\_{\mathrm{true}}, \Sigma) + \epsilon \mathcal{N}(\theta\_c, \Sigma)$ where $\epsilon = 10$%, $\theta\_{\mathrm{true}} = [0.5, \dots, 0.5]^\top$, $\theta\_c = [2,\dots,2]^\top$, $p(\theta) = \mathcal{U}([-1,1])^{10}$. The average MMD between the posterior predictive distribution and the observed data over 100 runs is shown in the following table (std. deviation is reported in the parenthesis). We will of course include these results in the paper. | | NPE | NPE-RS | NPE-RS | NPE-RS | |:---:|:----:|:--------------:|:------------:|:-------------:| | | - | $\lambda = 20$ | $\lambda=50$ | $\lambda=100$ | | **MMD** | 0.26 (0.02) | 0.19 (0.04) | **0.18** (0.06) | 0.21 (0.08) | * **"In my opinion, claims like 'It provides accurate inference results even when the model is well-specified' (L 67) are not warranted and have to be removed or significantly trimmed down":** Apologies if we weren't clear before. We agree that our results do not support the claim that "It provides accurate inference results even when the model is well-specified, thus circumventing any need to detect model misspecification", and we will remove it. * **"Wouldn't negative log-likelihood on misspecified x be a better measure?":** The log-likelihood is not applicable for simulator-based models due to the intractability of the likelihood function. We did consider fitting a multivariate Gaussian distribution to the predictive distributions from Ricker, OUP, and Turin model, and evaluating the log-likelihood of the data under the respective Gaussians. However, that introduces a new level of misspecification, as none of these models produce data that is jointly Gaussian. Moreover, when the observed data is corrupted by a few outlier points (thus causing model misspecification), the sum of log-likelihoods gets dominated by the few outliers, even if most of the observed data is explained well by the model. This is the reason why people have proposed generalised Bayesian inference (GBI) frameworks, where the likelihood term is replaced by a robust loss such as the MMD (which is known to be robust to outliers) to account for misspecification (see Knoblauch et al. (2019) and ref [24] for more details). We therefore used the MMD between the posterior predictive distributions and the observed data (bottom row of Figure 2) as the measure for robustness to misspecification (apart from RMSE). Knoblauch, J., Jewson, J., \& Damoulas, T. (2019). Generalized variational inference: Three arguments for deriving new posteriors. arXiv preprint arXiv:1904.02063. * **"Which of the three metrics shown in figure 5 should correspond exactly to "robustness to misspecification" and why is this warranted?":** In Figure 5, we are testing the claim that our method indeed converges to the NPE posterior as $\lambda$ goes to zero, and converges to the prior as $\lambda$ goes to infinity. To do that, we varied $\lambda$ and computed the distance (in terms of MMD) between the NPE-RS posterior and the prior (Fig 5(left)), and between the NPE-RS posterior and the NPE posterior (middle and right). * **Does this mean that using cross-validation requires several datasets of observed data (which are potentially misspecified and can not easily be synthetically generated) in order to choose $\lambda$? What if only one dataset is available at training time?:** We do not assume that multiple observed datasets are available, even though that is the case in some fields like Astrophysics where available data is plenty. We meant that for one observed dataset $\mathbf{y}\_{1:n} = \{\mathbf{y}^{(1)}\, \dots, \mathbf{y}^{(n)}\}$ with $n$ iid samples available at training time, we can use a random subset of the data, say $\mathbf{y}\_{1:m}$, where $m<n$, as the validation set to choose $\lambda$, and the rest of the $n-m$ points to run NPE-RS (which is what we did for the radio propagation experiment). In the limiting case that only one observed data point is available, there is not going to be any significant posterior update anyway.

Reviewer 1V4M2023-08-14

Response

Thanks for your detailed response. > line 255 I think my confusion comes from the difference between the $m$ training samples and $n$ realisations and how they both come into the algorithm. I am willing to raise my score, thanks to your response and additional experiments.

Authorsrebuttal2023-08-15

Thank you for your response and for increasing your score. We really appreciate it. Regarding line 255: We have $n$ iid realisations for each dataset $\mathbf{x}\_{1:n} = \{\mathbf{x}^{(1)}\, \dots, \mathbf{x}^{(n)}\} \sim \mathbb{P}\_\theta$ simulated from the model for a given $\theta$. Thus, $(\theta, \mathbf{x}\_{1:n})$ form a pair. In order to train the NPE network (or for ABC), we generate $m$ such pairs by first sampling $\theta_1, \dots, \theta_m \sim p(\theta)$ from the prior, and then generating $\mathbf{x}\_{1:n, i} \sim \mathbb{P}\_{\theta_i}$, $1 \leq i \leq m$. This results in the training data $\{(\theta\_i, \mathbf{x}\_{1:n, i})\}_{i=1}^m$ of size $m$. So to sum up, there are $m$ simulated datasets in the training data, where each dataset has $n$ iid samples.

Reviewer EXCq2023-08-15

Response

Dear Author, thank you for your detailed responses. Some of my concerns have been alleviated, and some of my confusions cleared up. - In general I think would be useful to readers to ensure that the text clearly describes the differences to ref [72], as you have discussed in your responses. I think your response was helpful - Indeed, with a summarizer on iid data points, you only gain in the reduction of dimensionality from the data to summary size. It was not clear why this would necessarily be a problem in ABC if summary is low dimensional. It was also not clear to me how a summarizer over datasets could handle data sets containing millions of iid samples, as is common in mane scientific applications. Thus summarizing over datasets introduces a different challenge of dimensionality. Nonetheless, I recognize that this is not the main goal of this work, but rather the focus is on learning robust statistics, and not about exactly how one summarizes the data. - I suggest changing notation, as theta_true is highly misleading. I am also not quite sure I understand your response. I suggest the authors improve the text here.

Authorsrebuttal2023-08-16

Thank you for your response and for increasing your score. We really appreciate it. * **"In general I think would be useful to readers to ensure that the text clearly describes the differences to ref [72], as you have discussed in your responses. I think your response was helpful":** We are glad. And yes of course we will include that text in the paper. * **I suggest changing notation, as theta_true is highly misleading. I am also not quite sure I understand your response. I suggest the authors improve the text here.":** For simulation experiments, the observed data is generated from some ground-truth parameter values, which we refer to as $\theta_{\mathrm{true}}$. This is done in order to measure the performance of the inference methods in the parameter space (see e.g. Figures 2, 4 and 5 in ref [41] where the ground truth parameters are denoted by red lines). As our method requires a validation dataset to set the hyperparameter $\lambda$, we use the same ground truth parameters to generate it as well. We will edit the notation and the text in the paper to clarify this point.

Reviewer DQRY2023-08-15

Thanks for the response. I have read all the other reviews and responses. My score remains.

Reviewer DQRY2023-08-19

Reduced the confidence level of my rating to 1 since I am not in this area. Hope this helps.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC