Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Does Unsupervised Domain Adaptation Improve the Robustness of Amortized Bayesian Inference? A Systematic Evaluation

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Domain alignment helps noisy data but fails when priors shift

desk verdict The core finding—UDA helps under likelihood misspecification but hurts under prior shifts—is real and well supported in synthetic benchmarks, but the real-world experiment is too cleanly composed to support the practical claim. read the letter →

arxiv 2502.04949 v2 pith:IM72NOCM submitted 2025-02-07 stat.ML cs.LGstat.ME

Lasse Elsemüller , Valentin Pratz , Mischa von Krause , Andreas Voss , Paul-Christian Bürkner , Stefan T. Radev This is my paper · ORCID
classification stat.MLcs.LGstat.ME
keywords unsuperviseddomainadaptationamortizedBayesianinferenceneuralposteriorestimationmodelmisspecificationmaximummeandiscrepancydomain-adversarialnetworkssimulation-basedrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether unsupervised domain adaptation (UDA) can make amortized Bayesian inference reliable when the simulator does not match the observed data. It systematically evaluates two UDA-based neural posterior estimators, one built on maximum mean discrepancy and one new domain-adversarial variant, across likelihood and prior misspecification scenarios. The central finding is that aligning summary spaces between simulated and observed data effectively removes the impact of unmodeled noise or contamination, but the same alignment degrades inference when the prior location or scale is shifted. The authors explain this by showing that the UDA loss moves the inference target from the analytic posterior to a 'denoised' posterior, which is desirable under likelihood shifts but discards information needed to correct a bad prior. These results imply that the type of misspecification must be diagnosed before UDA is applied to neural posterior estimation.

What carries the argument

The load-bearing object is the joint NPE-UDA objective $\mathcal{L}_{\mathrm{NPE-UDA}} = \mathcal{L}_{\mathrm{NPE}} + \lambda \cdot \mathcal{L}_{\mathrm{UDA}}$, in which the standard negative log-posterior loss is augmented by a divergence between the summary statistics $\phi(x)$ of simulated data and those of unlabeled observed data. Two instantiations are studied: NPE-MMD, which uses maximum mean discrepancy as the divergence, and NPE-DANN, which uses a domain classifier with a gradient reversal layer to make the summary network produce domain-invariant embeddings. The regularization weight $\lambda$ controls how much the inference target shifts toward the 'denoised' posterior, and the paper shows that finding a good $\lambda$ is application-specific and that large values can destroy the summary space or destabilize training.

What would settle it

Run NPE-MMD or NPE-DANN on a benchmark with a known prior location or scale shift and a well-specified likelihood, and observe the posterior approximation error improve over the standard NPE baseline; the paper predicts the opposite, so any such improvement across multiple seeds would falsify the central claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that the source of misspecification determines whether domain adaptation helps or hurts neural posterior estimation. In experiments with a Ricker model, a two-dimensional Gaussian model, a Bayesian image denoising task, and large-scale human decision-making data, NPE-MMD and NPE-DANN both reduce summary-space domain distance and improve posterior estimates under likelihood misspecification such as added noise, contamination, or blur. Under prior location or scale misspecification, the same methods perform worse than standard NPE, even though the summary spaces are aligned; the alignment removes or distorts information that the posterior needs to counteract the wrong prior. The authors introduce NPE-DANN as a new variant and find it more stable over a wider range of regularization weights than NPE-MMD, which can over-align until the summary space becomes uninformative. Their framing is that the UDA loss shifts the inference target from the analytic posterior $p(\theta|x_{\mathrm{obs}})$ to a posterior based on adjusted data $p(\theta|\tilde{x}_{\mathrm{obs}})$, and the empirical real-data study shows that this shift does not transfer from the majority of well-specified data sets to a minority of misspecified ones.

Load-bearing premise

The unlabeled observed data used to compute the domain-alignment loss come from the same target distribution as the observations the network will face at deployment, and the real-data experiment shows this premise fails when the misspecified cases are a minority.

Editorial extensions

If this is right

  • Under likelihood misspecification, NPE-MMD and NPE-DANN can match or beat standard NPE and NNPE on contamination and noise benchmarks, with the biggest gains in the contamination scenario.
  • Under prior location or scale shifts, both UDA methods degrade posterior calibration and error relative to standard NPE, so UDA should not be used when the prior is suspect.
  • The regularization weight $\lambda$ is the key control: too little alignment leaves the domain shift unhandled, while too much alignment makes the summary space uninformative or destabilizes training.
  • In real-world deployment, aligning to the bulk of observed data does not guarantee reliability for a misspecified minority; NNPE's explicit noise model proved more useful for that minority.
  • Because UDA shifts the inference target, downstream quantities such as posterior predictive distances to raw observed data can punish successful noise removal; evaluation should use a denoised reference.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A practical corollary the authors leave implicit is that UDA should be paired with a misspecification-type diagnostic: detect whether the gap is in the likelihood or the prior before choosing alignment-based robustness.
  • The target-shift framing suggests a testable extension: combining UDA with a prior-adjustment method could recover the losses under prior misspecification while keeping the likelihood-robustness gains.
  • The observed sweet spot between summary-space domain distance and prediction error implies that summary-space alignment alone is not a reliable model-selection metric; practitioners should not tune $\lambda$ to minimize domain distance without checking parameter recovery.
  • For high-stakes applications with heterogeneous populations, the real-data experiment points to a need for UDA variants that adapt to clusters within the observed domain rather than to one global aligned summary space.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper evaluates whether unsupervised domain adaptation (UDA) improves the robustness of amortized Bayesian inference. The authors introduce NPE-DANN and compare it with NPE-MMD, standard NPE, and NNPE on four benchmark settings: a Ricker model with contamination (Experiment 1), a 2D Gaussian means model with prior location, prior scale, likelihood scale, and contamination misspecifications (Experiment 2), a high-dimensional Bayesian denoising task with prior and likelihood shifts (Experiment 3), and a real-world decision-making (IAT) data set (Experiment 4). The central claim is that aligning summary spaces between simulated and observed domains mitigates likelihood misspecification but degrades performance under prior misspecification. The experiments support this pattern for several likelihood misspecification scenarios (notably contamination) and for prior shifts, but the likelihood-scale scenario in Experiment 2 and the real-world Experiment 4 exhibit notable deviations, which the paper partially acknowledges. The work contributes a systematic benchmark, a Target 1/2/3 taxonomy of robustness goals, and a candid account of the severe sensitivity of UDA methods to the regularization weight λ.

Significance. If the claimed pattern holds, this is a useful contribution to the growing literature on robust simulation-based inference: it provides the first broad comparison of NPE-MMD and NPE-DANN, proposes a clear conceptual distinction between targeting the observed-data posterior, a denoised-data posterior, and an adjusted-prior posterior, and it systematically documents when UDA helps and when it fails. The synthetic benchmarks are carefully designed with multiple metrics (parameter-space, data-space, and network-space), and the paper is transparent about instabilities, λ-dependence, and the loss-of-information danger of overly aggressive alignment. However, the blanket statement in the abstract that UDA 'effectively mitigates' likelihood misspecification is not fully supported by the likelihood-scale results in Experiment 2, and the real-world significance of the null result in Experiment 4 is limited by the composition of the unlabeled target pool and by the NPE-based definition of 'misspecified' test sets. These issues require revision but do not invalidate the core synthetic findings.

major comments (4)
  1. [5.2 (Experiment 2, Likelihood Scale)] In the likelihood-scale scenario (τ=10 and τ=20), NPE-DANN fails to align the summary space (high SSDD) and performs poorly on NRMSE, ECE, and PPD, while NPE-MMD yields only a slight improvement in posterior contraction (Figure 5 and accompanying text). This contradicts the abstract's claim that aligning summary spaces 'effectively mitigates' likelihood misspecification and the Discussion's statement that the pattern was 'consistently found throughout our systematic evaluations' (Section 6). The claim should be qualified to specify which types of likelihood misspecification (e.g., contamination) are mitigated, or the authors should provide a mechanistic explanation for why likelihood-scale shifts behave differently.
  2. [5.4 (Experiment 4, Target-Pool Composition)] The NPE-UDA methods are trained on 32,000 unlabeled empirical data sets drawn from the general population, which is dominated by well-specified participants, while the evaluation of 'misspecified' data uses only 730 atypical participants. The UDA alignment loss in Eqs. (3)–(4) therefore matches the summary space to the majority well-specified distribution. The observed failure to help the misspecified minority is an expected consequence of the target-pool composition and does not test whether UDA would help when the unlabeled target pool represents the deployment subpopulation of interest. This weakens the real-world message that 'the general domain adaptation ... did not carry over' (Section 5.4). Please add an experiment where the target pool is enriched for misspecified participants, or explicitly restrict the conclusion to the well-specified-dominated target regime.
  3. [5.4 / Appendix B.6 (Circularity of Misspecification Labels)] The well-specified vs. misspecified split is created by thresholding Mahalanobis distances in the summary space of three standard NPE networks—the same model family (and, for NPE-MMD and NPE-DANN, the same base summary network architecture) that is being evaluated. The test-set labels are therefore relative to the baseline NPE embedding, not to an independent notion of model validity. Since the UDA methods deliberately deform the summary space, their view of 'misspecified' versus 'well-specified' may differ, and the comparison on these fixed labels is circular. The real-world conclusion would be more robust if the labels were anchored to external behavioral criteria (e.g., attention checks, RT outliers) or if a sensitivity analysis re-classified the test sets using each method's own summary space.
  4. [Abstract and Section 6 (Scope of Claims)] The abstract's first contribution statement says the paper 'demonstrate[s] that aligning summary spaces between domains effectively mitigates the impact of unmodeled phenomena or noise.' Given the Experiment 2 likelihood-scale failure of NPE-DANN and the Experiment 4 null result, this wording overgeneralizes. The authors should either narrow the claim (e.g., 'in several likelihood-misspecification settings') or add a caveat in the abstract about the dependence on misspecification type, target-pool composition, and λ. The Discussion's 'consistently found these patterns' (Section 6) should be softened accordingly.
minor comments (6)
  1. [Eq. (4)] The statement that d(·,·) 'attains its global minimum if and only if ϕ(x)=ϕ(x_obs)' is not correct for MMD or adversarial divergence: the minimum is attained when the distributions of the summary statistics are equal, not pointwise equality of individual summaries. Please rephrase to 'if and only if the distributions of ϕ(x) and ϕ(x_obs) coincide'.
  2. [Eq. (7)] The discriminator loss notation is confusing: it writes p(ψ(ϕ(x))) as a probability but does not define ψ's output range. Consider writing L_D = -E_{x~p(x)}[log σ(ψ(ϕ(x)))] - E_{x_obs~p(x_obs)}[log(1-σ(ψ(ϕ(x_obs))))] or clarify the notation.
  3. [Section 5, metric list] There is a typo: 'fidelity of credible intervalst' should be 'credible intervals'.
  4. [Section 5.4, Figure 7] The main text leaves out the λ=10 'failure setting' from the visualizations, and although the Appendix (B.6, Figure B12) discusses the corresponding prior-posterior collapse, the main-text figure does not show the severity of the instability. Adding a sentence or a supplementary panel summarizing the λ=10 behavior would make the main narrative more self-contained.
  5. [Section 3.3] For NPE-DANN, Eq. (6) multiplies the discriminator loss L_D by λ, while the standard DANN uses a gradient reversal weight λ_grl on the reversed gradients; the relation between λ and λ_grl (both optimized in Experiment 1, Table B.1) should be clarified, since they play different roles in the optimization.
  6. [Tables 1 and B.4] The SSDD values of 0.000 for NPE-MMD at λ=1.00 are reported without precision; since exact zero is unlikely for a finite-sample MMD estimate, please state the actual magnitude or specify that they are below a numerical threshold.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are supported by independent synthetic benchmarks, and the Experiment 4 design limitation is not a circular reduction.

full rationale

The paper contains no derivation that could reduce to its own inputs. The central claims are empirical: NPE-UDA methods are trained on simulated source data plus unlabeled target data and evaluated on held-out target data under controlled misspecification (Experiments 1-3), so the reported improvements under likelihood misspecification and degradations under prior misspecification are genuine out-of-sample findings rather than construction artifacts. The only methodological concern is in Experiment 4, where misspecified test sets are labeled by atypicality in standard NPE summary space and the UDA training pool is dominated by well-specified participants; this limits the generalizability of the real-world null result, but it is a validity limitation, not a circularity, because the labels are not the quantities being predicted and the main conclusions do not depend on Experiment 4 alone. Self-citations (e.g., BayesFlow, von Krause & Radev 2025) are implementation or modeling choices that are not load-bearing for the paper's conclusions.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the UDA framework (Eq. 3-4), which inherits standard domain-adaptation assumptions: a low-error shared hypothesis must exist, the observed target data must be representative, and aligning summary statistics must not destroy posterior information. The paper's own experiments show these assumptions can fail, especially in Experiment 4 where alignment to predominantly well-specified data does not transfer to misspecified cases.

free parameters (2)
  • UDA alignment weight lambda = varied per experiment (0.01-150 in Exp. 1; 0.1, 1, 10 in Exp. 2-4)
    Controls the trade-off between NPE loss and domain alignment. The paper shows results flip between improvement and failure depending on lambda, so the central claim is conditional on this parameter.
  • MMD kernel choice = sum of inverse multiquadric kernels
    Selected for stable training; the paper notes that sums of Gaussian kernels destabilized training, so the kernel choice affects the reported behavior of NPE-MMD.
assumptions (4)
  • standard math Ben-David domain adaptation bound (Eq. 1) holds and motivates minimizing representation divergence.
    Used in Section 2 to justify the UDA approach; standard result from Ben-David et al. (2010).
  • domain assumption There exists a low-error hypothesis for both source and target domains.
    Invoked in Section 3.1 via Redko et al. (2022); if no shared hypothesis works, alignment cannot succeed.
  • domain assumption The observed target data x_obs used for alignment are drawn from the deployment distribution.
    Required for the UDA loss in Eq. 3-4; Experiment 4 shows this can fail when training observations are mostly well-specified while test data contain a misspecified minority.
  • domain assumption Aligned summary statistics retain sufficient information about parameters for posterior inference.
    The UDA loss in Eq. 4 trades off information extraction against domain alignment; if alignment discards posterior-relevant information, inference fails, as seen with NPE-MMD at lambda=1.0 in Experiment 3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Does Unsupervised Domain Adaptation Improve the Robustness of Amortized Bayesian Inference? A Systematic Evaluation." pith.science (2026). https://pith.science/paper/IM72NOCM

@misc{pith2026250204949,
  author       = {Pith},
  title        = {Pith review of: Does Unsupervised Domain Adaptation Improve the Robustness of Amortized Bayesian Inference? A Systematic Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IM72NOCM}},
  note         = {Machine review of arXiv:2502.04949}
}
read the original abstract

Neural networks are fragile when confronted with data that significantly deviates from their training distribution. This is true in particular for simulation-based inference methods, such as neural amortized Bayesian inference (ABI), where models trained on simulated data are deployed on noisy real-world observations. Recent robust approaches employ unsupervised domain adaptation (UDA) to match the embedding spaces of simulated and observed data. However, the lack of comprehensive evaluations across different domain mismatches raises concerns about the reliability in high-stakes applications. We address this gap by systematically testing UDA approaches across a wide range of misspecification scenarios in silico and practice. We demonstrate that aligning summary spaces between domains effectively mitigates the impact of unmodeled phenomena or noise. However, the same alignment mechanism can lead to failures under prior misspecifications - a critical finding with practical consequences. Our results underscore the need for careful consideration of misspecification types when using UDA to increase the robustness of ABI.

Figures

Figures reproduced from arXiv: 2502.04949 by the authors.

Figure 1
Figure 1. Schematic overview of NPE-UDA methods that combine neural posterior estimation (NPE) with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Experiment 3: Sum￾mary space domain distance (SSDD; MMD) vs. normal￾ized root mean squared error (NRMSE) for row deletions. We observe a sweet spot of domain alignment without losing impor￾tant information. From Simulated to Real Domains The preceding discussion makes the connection between UDA and ABI immediately apparent: When the distance between the data distribution p(xobs) and the model-implied distribution p(… view at source ↗
Figure 3
Figure 3. Experiment 1. Parameter space performance metrics resulting from 50 separate Bayesian hyper￾parameter optimization runs per method. The solid trend lines represent the predictive mean of a Gaussian process regression fitted to the individual run results, with the shaded areas representing 95% confidence intervals of the predictive distribution. If a parameter was not optimized, the methods average performance is dep… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Experiment 1. Further metrics resulting from 50 separate Bayesian hyperparameter optimization runs per method. The solid trend lines represent the predictive mean of a Gaussian process regression fitted to the individual run results, with the shaded areas representing …
Figure 5
Figure 5. Figure 5: Experiment 2. Performance metrics and summary space domain distance (SSDD) of the methods in all misspecification scenarios (columns), aggregated via the median of 10 runs. The first row shows the well-specified setting, with misspecification increasing from top to bot…
Figure 6
Figure 6. Figure 6: Experiment 3 – Contamination (Rows): t-SNE representation of the summary spaces from the best run (lowest NRMSE) of each method. The t-SNE map is calculated jointly using summary statistics of data from both the simulated and the observed domain. For NPE, the two domai…
Figure 7
Figure 7. Figure 7: Experiment 4 – Metric results for all methods and different λ weights for the NPE-UDA methods. All runs display the averaged results across three runs per method, with across-run standard deviations shown as shaded areas (for NPE and NNPE) or error bars (for NPE-DANN a…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Diffusion Models in Simulation-Based Inference: A Tutorial Review

    stat.ML 2025-12 conditional novelty 5.0 of 10

    Design choices — noise schedule, parameterization, sampler, and model family — measurably change posterior accuracy in diffusion-based SBI; variance-preserving EDM diffusion with adaptive solvers leads on low-dimensio...

Reference graph

Works this paper leans on

71 extracted references · 45 canonical work pages · cited by 1 Pith paper

  1. [1]

    Analyzing inverse problems with invertible neural networks

    Lynton Ardizzone, Jakob Kruse, Sebastian Wirkert, Daniel Rahner, Eric W Pellegrini, Ralf S Klessen, Lena Maier-Hein, Carsten Rother, and Ullrich K \"o the. Analyzing inverse problems with invertible neural networks. arXiv preprint arXiv:1808.04730, 2018

    arXiv 2018

  2. [2]

    u th, Niels Bracher, Carsten Rother, and Ullrich K \

    Lynton Ardizzone, Jakob Kruse, Carsten L \"u th, Niels Bracher, Carsten Rother, and Ullrich K \"o the. Conditional invertible neural networks for diverse image-to-image translation. In Pattern Recognition: 42nd DAGM German Conference, DAGM GCPR 2020, T \"u bingen, Germany, September 28--October 1, 2020, Proceedings 42 , pp.\ 373--387. Springer, 2021

    work page 2020

  3. [3]

    Neural networks enable efficient and accurate simulation-based inference of evolutionary parameters from adaptation dynamics

    Grace Avecilla, Julie N Chuong, Fangfei Li, Gavin Sherlock, David Gresham, and Yoav Ram. Neural networks enable efficient and accurate simulation-based inference of evolutionary parameters from adaptation dynamics. PLoS Biology, 20 0 (5): 0 e3001633, 2022

    work page 2022

  4. [4]

    Analysis of representations for domain adaptation

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006

    2006

  5. [5]

    A theory of learning from different domains

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79: 0 151--175, 2010

    2010

  6. [6]

    Combined statistical-mechanistic modeling links ion channel genes to physiology of cortical neuron types

    Yves Bernaerts, Michael Deistler, Pedro J Gon c alves, Jonas Beck, Marcel Stimberg, Federico Scala, Andreas S Tolias, Jakob Macke, Dmitry Kobak, and Philipp Berens. Combined statistical-mechanistic modeling links ion channel genes to physiology of cortical neuron types. bioRxiv, pp.\ 2023--03, 2023

    work page 2023

  7. [7]

    Calibrating model-based inferences and decisions

    Michael Betancourt. Calibrating model-based inferences and decisions. arXiv preprint arXiv:1803.08393, 2018

    arXiv 2018

  8. [8]

    A practical guide to statistical distances for evaluating generative models in science

    Sebastian Bischoff, Alana Darcher, Michael Deistler, Richard Gao, Franziska Gerken, Manuel Gloeckler, Lisa Haxel, Jaivardhan Kapoor, Janne K Lappalainen, Jakob H Macke, et al. A practical guide to statistical distances for evaluating generative models in science. arXiv preprint arXiv:2403.12636, 2024

    arXiv 2024

Show all 71 references
  1. [9]

    A general framework for updating belief distributions

    Pier Giovanni Bissiri, Chris C Holmes, and Stephen G Walker. A general framework for updating belief distributions. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 78 0 (5): 0 1103--1130, 2016

    2016

  2. [10]

    Some models are useful, but how do we know which ones? towards a unified bayesian model taxonomy

    Paul-Christian B \"u rkner, Maximilian Scholz, and Stefan T Radev. Some models are useful, but how do we know which ones? towards a unified bayesian model taxonomy. Statistic Surveys, 17: 0 216--310, 2023

    2023

  3. [11]

    Investigating the impact of model misspecification in neural simulation-based inference

    Patrick Cannon, Daniel Ward, and Sebastian M Schmon. Investigating the impact of model misspecification in neural simulation-based inference. arXiv preprint arXiv:2209.01845, 2022

    2022 arXiv

  4. [12]

    Stan: A probabilistic programming language

    Bob Carpenter, Andrew Gelman, Matthew D Hoffman, Daniel Lee, Ben Goodrich, Michael Betancourt, Marcus Brubaker, Jiqiang Guo, Peter Li, and Allen Riddell. Stan: A probabilistic programming language. Journal of statistical software, 76 0 (1), 2017

    2017

  5. [13]

    A likelihood-free inference framework for population genetic data using exchangeable neural networks

    Jeffrey Chan, Valerio Perrone, Jeffrey Spence, Paul Jenkins, Sara Mathieson, and Yun Song. A likelihood-free inference framework for population genetic data using exchangeable neural networks. Advances in neural information processing systems, 31, 2018

    2018

  6. [14]

    Domain adaptation techniques for improved cross-domain study of galaxy mergers

    A \'C iprijanovi \'c , Diana Kafkes, S Jenkins, K Downey, Gabriel N Perdue, Sandeep Madireddy, T Johnston, and Brian Nord. Domain adaptation techniques for improved cross-domain study of galaxy mergers. arXiv preprint arXiv:2011.03591, 2020

    2011 arXiv

  7. [15]

    Domain adaptation and sample bias correction theory and algorithm for regression

    Corinna Cortes and Mehryar Mohri. Domain adaptation and sample bias correction theory and algorithm for regression. Theoretical Computer Science, 519: 0 103--126, January 2014. ISSN 0304-3975. doi:10.1016/j.tcs.2013.09.027

    2014 doi

  8. [16]

    Elements of information theory

    Thomas M Cover and Joy A Thomas. Elements of information theory. John Wiley & Sons, 2012

    2012

  9. [17]

    The frontier of simulation-based inference

    Kyle Cranmer, Johann Brehmer, and Gilles Louppe. The frontier of simulation-based inference. Proceedings of the National Academy of Sciences, 117 0 (48): 0 30055--30062, 2020

    2020

  10. [18]

    Simulation-based inference of single-molecule experiments, 2024

    Lars Dingeldein, Pilar Cossio, and Roberto Covino. Simulation-based inference of single-molecule experiments, 2024. URL https://arxiv.org/abs/2410.15896

    2024 arXiv

  11. [19]

    u ller, Hans Olischl \

    Lasse Elsem \"u ller, Hans Olischl \"a ger, Marvin Schmitt, Paul-Christian B \"u rkner, Ullrich K \"o the, and Stefan T Radev. Sensitivity-aware amortized bayesian inference. Transactions on Machine Learning Research (TMLR), 2024

    2024

  12. [20]

    Frazier, Ryan Kelly, Christopher Drovandi, and David J

    David T. Frazier, Ryan Kelly, Christopher Drovandi, and David J. Warne. The statistical accuracy of neural posterior and likelihood estimation, 2024. URL https://arxiv.org/abs/2411.12068

    2024 arXiv

  13. [21]

    Domain-adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Fran c ois Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17 0 (59): 0 1--35, 2016

    2016

  14. [22]

    Generalized bayesian inference for scientific simulators via amortized cost estimation

    Richard Gao, Michael Deistler, and Jakob H Macke. Generalized bayesian inference for scientific simulators via amortized cost estimation. Advances in Neural Information Processing Systems, 36: 0 80191--80219, 2023

    2023

  15. [23]

    Amortized inference in probabilistic reasoning

    Samuel Gershman and Noah Goodman. Amortized inference in probabilistic reasoning. In Proceedings of the annual meeting of the cognitive science society, volume 36, 2014

    2014

  16. [24]

    Variational methods for simulation-based inference

    Manuel Gl \"o ckler, Michael Deistler, and Jakob H Macke. Variational methods for simulation-based inference. arXiv preprint arXiv:2203.04176, 2022

    2022 arXiv

  17. [25]

    Adversarial robustness of amortized bayesian inference

    Manuel Gl \"o ckler, Michael Deistler, and Jakob H Macke. Adversarial robustness of amortized bayesian inference. arXiv preprint arXiv:2305.14984, 2023

    2023 arXiv

  18. [26]

    All-in-one simulation-based inference

    Manuel Gloeckler, Michael Deistler, Christian Weilbach, Frank Wood, and Jakob H Macke. All-in-one simulation-based inference. arXiv preprint arXiv:2404.09636, 2024

    2024 arXiv

  19. [27]

    Borgwardt, Malte Rasch, Bernhard Schölkopf, and AJ Smola

    A Gretton, K. Borgwardt, Malte Rasch, Bernhard Schölkopf, and AJ Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13: 0 723--773, 03 2012

    2012

  20. [28]

    Likelihood-free mcmc with amortized approximate ratio estimators

    Joeri Hermans, Volodimir Begy, and Gilles Louppe. Likelihood-free mcmc with amortized approximate ratio estimators. In International conference on machine learning, pp.\ 4239--4248. PMLR, 2020

    2020

  21. [29]

    Learning robust statistics for simulation-based inference under model misspecification

    Daolang Huang, Ayush Bharti, Amauri Souza, Luigi Acerbi, and Samuel Kaski. Learning robust statistics for simulation-based inference under model misspecification. Advances in Neural Information Processing Systems, 36, 2024

    2024

  22. [30]

    Robust statistics

    Peter J Huber. Robust statistics. John Wiley & Sons, 1981

    1981

  23. [31]

    J.J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16 0 (5): 0 550--554, 1994. doi:10.1109/34.291440

    1994 doi

  24. [32]

    Support and invertibility in domain-invariant representations

    Fredrik D Johansson, David Sontag, and Rajesh Ranganath. Support and invertibility in domain-invariant representations. In The 22nd International Conference on Artificial Intelligence and Statistics, pp.\ 527--536. PMLR, 2019

    2019

  25. [33]

    Misspecification-robust sequential neural likelihood

    Ryan P Kelly, David J Nott, David T Frazier, David J Warne, and Chris Drovandi. Misspecification-robust sequential neural likelihood. arXiv preprint arXiv:2301.13368, 2023

    2023 arXiv

  26. [34]

    Glow: Generative flow with invertible 1x1 convolutions

    Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018

    2018

  27. [35]

    Lecun, L

    Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86 0 (11): 0 2278--2324, 1998. doi:10.1109/5.726791

    1998 doi

  28. [36]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  29. [37]

    Deep unsupervised domain adaptation: A review of recent advances and perspectives

    Xiaofeng Liu, Chaehwa Yoo, Fangxu Xing, Hyejin Oh, Georges El Fakhri, Je-Won Kang, Jonghye Woo, et al. Deep unsupervised domain adaptation: A review of recent advances and perspectives. APSIPA Transactions on Signal and Information Processing, 11 0 (1), 2022

    2022

  30. [38]

    Learning transferable features with deep adaptation networks

    Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pp.\ 97--105. PMLR, 2015

    2015

  31. [39]

    Domain adaptation: Learning bounds and algorithms

    Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430, 2009

    2009 arXiv

  32. [40]

    Unsupervised domain adaptation: A reality check

    Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. Unsupervised domain adaptation: A reality check. arXiv preprint arXiv:2111.15672, 2021

    2021 arXiv

  33. [41]

    Simulation-based inference of black hole ringdowns in the time domain

    Costantino Pacilio, Swetha Bhagwat, and Roberto Cotesta. Simulation-based inference of black hole ringdowns in the time domain. Physical Review D, 110 0 (8): 0 083010, 2024

    2024

  34. [42]

    Domain adaptation via transfer component analysis

    Sinno Jialin Pan, Ivor W Tsang, James T Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. IEEE transactions on neural networks, 22 0 (2): 0 199--210, 2010

    2010

  35. [43]

    Fast -free inference of simulation models with bayesian conditional density estimation

    George Papamakarios and Iain Murray. Fast -free inference of simulation models with bayesian conditional density estimation. Advances in neural information processing systems, 29, 2016

    2016

  36. [44]

    Unsupervised domain adaptation for medical imaging segmentation with self-ensembling

    Christian S Perone, Pedro Ballester, Rodrigo C Barros, and Julien Cohen-Adad. Unsupervised domain adaptation for medical imaging segmentation with self-ensembling. NeuroImage, 194: 0 1--11, 2019

    2019

  37. [45]

    Bayesflow: Learning complex stochastic models with invertible neural networks

    Stefan T Radev, Ulf K Mertens, Andreas Voss, Lynton Ardizzone, and Ullrich K \"o the. Bayesflow: Learning complex stochastic models with invertible neural networks. IEEE transactions on neural networks and learning systems, 2020

    2020

  38. [46]

    u ller, Valentin Pratz, Yannik Sch \

    Stefan T Radev, Marvin Schmitt, Lukas Schumacher, Lasse Elsem \"u ller, Valentin Pratz, Yannik Sch \"a lte, Ullrich K \"o the, and Paul-Christian B \"u rkner. Bayesflow: Amortized bayesian workflows with neural networks. arXiv preprint arXiv:2306.16015, 2023

    2023 arXiv

  39. [47]

    Modern bayesian experimental design

    Tom Rainforth, Adam Foster, Desi R Ivanova, and Freddie Bickford Smith. Modern bayesian experimental design. Statistical Science, 39 0 (1): 0 100--114, 2024

    2024

  40. [48]

    Greenberg, Pedro J

    Poornima Ramesh, Jan-Matthis Lueckmann, Jan Boelts, \'A lvaro Tejero-Cantero, David S. Greenberg, Pedro J. Goncalves, and Jakob H. Macke. GATSBI : Generative adversarial training for simulation-based inference. In International Conference on Learning Representations, 2022. URL...

    2022

  41. [49]

    A survey on domain adaptation theory: Learning bounds and theoretical guarantees, July 2022

    Ievgen Redko, Emilie Morvant, Amaury Habrard, Marc Sebban, and Youn \`e s Bennani. A survey on domain adaptation theory: Learning bounds and theoretical guarantees, July 2022

    2022

  42. [50]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part...

    2015

  43. [51]

    Synthetic data could be better than real data

    Neil Savage. Synthetic data could be better than real data. Nature, 2023

    2023

  44. [52]

    u rkner, and K \

    Marvin Schmitt, Paul-Christian B \"u rkner, and K \"o the. Detecting model misspecification in amortized bayesian inference with neural networks. Proceedings of the German Conference on Pattern Recognition (GCPR), 2023

    2023

  45. [53]

    Learning transferrable representations for unsupervised domain adaptation

    Ozan Sener, Hyun Oh Song, Ashutosh Saxena, and Silvio Savarese. Learning transferrable representations for unsupervised domain adaptation. Advances in neural information processing systems, 29, 2016

    2016

  46. [54]

    Reliable amortized variational inference with physics-based latent distribution correction

    Ali Siahkoohi, Gabrio Rizzuti, Rafael Orozco, and Felix J Herrmann. Reliable amortized variational inference with physics-based latent distribution correction. Geophysics, 88 0 (3): 0 R297--R322, 2023

    2023

  47. [55]

    Unsupervised domain adaptation through self-supervision

    Yu Sun, Eric Tzeng, Trevor Darrell, and Alexei A Efros. Unsupervised domain adaptation through self-supervision. arXiv preprint arXiv:1909.11825, 2019

    1909 arXiv

  48. [56]

    Domain adaptation for measurements of strong gravitational lenses

    Paxson Swierc, Megan Zhao, Aleksandra \'C iprijanovi \'c , and Brian Nord. Domain adaptation for measurements of strong gravitational lenses. arXiv preprint arXiv:2311.17238, 2023

    2023 arXiv

  49. [57]

    Domain-adaptive neural posterior estimation for strong gravitational lens analysis

    Paxson Swierc, Marcos Tamargo-Arizmendi, Aleksandra \'C iprijanovi \'c , and Brian D Nord. Domain-adaptive neural posterior estimation for strong gravitational lens analysis. arXiv preprint arXiv:2410.16347, 2024

    2024 arXiv

  50. [58]

    Deep domain confusion: Maximizing for domain invariance

    Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014

    2014 arXiv

  51. [59]

    Rheeya Uppaal, Yixuan Li, and Junjie Hu. How useful is continued pre-training for generative unsupervised domain adaptation? In Proceedings of the 9th Workshop on Representation Learning for NLP (RepL4NLP-2024), pp.\ 99--117, 2024

    2024

  52. [60]

    S ch\"onberger, J uan Nunez-Iglesias , F ran c ois B oulogne, J oshua D

    S t\'efan van der Walt, J ohannes L . S ch\"onberger, J uan Nunez-Iglesias , F ran c ois B oulogne, J oshua D . W arner, N eil Y ager, E mmanuelle G ouillart, T ony Y u, and the scikit-image contributors. scikit-image: image processing in P ython. PeerJ, 2: 0 e453, 6 2014. ISS...

    2014 doi

  53. [61]

    Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St \'e fan J

    Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St \'e fan J. van der Walt , Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nels...

    2020

  54. [62]

    Kelly, David J Warne, and Christopher Drovandi

    Xiaoyu Wang, Ryan P. Kelly, David J Warne, and Christopher Drovandi. Preconditioned neural posterior estimation for likelihood-free inference. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=vgIBAOkIhY

    2024

  55. [63]

    Robust neural posterior estimation and statistical model criticism

    Daniel Ward, Patrick Cannon, Mark Beaumont, Matteo Fasiolo, and Sebastian Schmon. Robust neural posterior estimation and statistical model criticism. Advances in Neural Information Processing Systems, 35: 0 33845--33859, 2022

    2022

  56. [64]

    Addressing misspecification in simulation-based inference through data-driven calibration

    Antoine Wehenkel, Juan L Gamella, Ozan Sener, Jens Behrmann, Guillermo Sapiro, Marco Cuturi, and J \"o rn-Henrik Jacobsen. Addressing misspecification in simulation-based inference through data-driven calibration. arXiv preprint arXiv:2405.08719, 2024

    2024 arXiv

  57. [65]

    Macke, and Bernhard Sch \"o lkopf

    Jonas Bernhard Wildberger, Maximilian Dax, Simon Buchholz, Stephen R Green, Jakob H. Macke, and Bernhard Sch \"o lkopf. Flow matching for scalable simulation-based inference. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.ne...

    2023

  58. [66]

    Deep sets

    Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems, 30, 2017

    2017

  59. [67]

    Neural methods for amortized inference

    Andrew Zammit-Mangion, Matthew Sainsbury-Dale, and Rapha \"e l Huser. Neural methods for amortized inference. Annual Review of Statistics and Its Application, 12, 2024

    2024

  60. [68]

    The balancing principle for parameter choice in distance-regularized domain adaptation

    Werner Zellinger, Natalia Shepeleva, Marius-Constantin Dinu, Hamid Eghbal-zadeh, Hoan Duc Nguyen, Bernhard Nessler, Sergei Pereverzyev, and Bernhard A Moser. The balancing principle for parameter choice in distance-regularized domain adaptation. Advances in Neural Information ...

    2021

  61. [69]

    Free lunch for domain adversarial training: Environment label smoothing

    YiFan Zhang, Xue Wang, Jian Liang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Free lunch for domain adversarial training: Environment label smoothing. arXiv preprint arXiv:2302.00194, 2023

    2023 arXiv

  62. [70]

    Domain generalization: A survey

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (4): 0 4396--4415, 2022

    2022

  63. [71]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.