Pith. sign in

REVIEW 2 major objections 5 minor 67 references

This paper develops simulation-based empirical Bayes: an iterative loop that learns a prior from many related observations when the likelihood is available only through simulation, and proves the oracle loop converges to the population prio

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 06:32 UTC pith:JYT2YLZN

load-bearing objection A genuinely new EB-in-SBI loop with a solid oracle theorem, but the experiments do not isolate EB from sequential focusing, so the empirical claims need a control. the 2 major comments →

arxiv 2607.21843 v1 pith:JYT2YLZN submitted 2026-07-23 stat.ML cs.LG

Simulation-Based Empirical Bayes

Xinwei Shen , Diana Cai , Cheng Zhang , David M. Blei This is my paper
classification stat.ML cs.LG MSC 62C1262F15
keywords simulation-based empirical Bayessimulation-based inferenceempirical Bayespopulation prioramortized posteriorimplicit likelihoodprior identifiabilitylikelihood-free inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central proposition is that empirical Bayes — learning the prior from many related observations — can be carried out when the likelihood is only a black-box simulator, by replacing classical likelihood-based estimates with an iterative loop of simulation-based inference. The key step is the population-prior identity: the desired prior is the average of the posterior distributions over the observed data. The paper proves that if the inference network is exact and population expectations are available, the oracle update is monotone in the induced marginal and, when the simulator is λ-strongly prior-identifiable, converges geometrically in KL divergence to the population prior. Empirically, the method improves held-out predictive accuracy over fixed-prior SBI on a discrete-choice survey and a measles epidemic simulator. If correct, it widens empirical Bayes to scientific simulators where likelihoods cannot be evaluated.

Core claim

Under the oracle assumption — exact amortized posterior and population expectations — SBEB updates the prior as p_{t+1}(z) = E_{p*(x)}[p_t(z|x)], the population average of posteriors under the current prior. This update is an EM step on the expected log marginal likelihood; Proposition 2 shows each step decreases KL to the population prior by at least the marginal KL, so marginals converge. With λ-strong prior identifiability, Theorem 1 gives geometric contraction at rate (1−λ)^t. The practical algorithm approximates the population expectation with the empirical average of amortized posteriors and resamples simulator pairs from that averaged prior.

What carries the argument

The population-prior identity p(z) = E_{p*(x)}[p(z|x)], a fixed-point equation defining the target prior, and the amortized inference network q(z;x,φ), a conditional density estimator trained on simulated (z,x) pairs, which makes the posterior average sampleable. The convergence analysis relies on λ-strong prior identifiability, an inverse data-processing condition bounding prior KL by marginal KL.

Load-bearing premise

The load-bearing premise is that the amortized inference network can represent the exact posterior under each evolving prior; if it cannot, as with the diagonal-Gaussian family used in the experiments, the averaged posterior is a biased estimate of the population prior and the iteration may converge to a different fixed point.

What would settle it

Run SBEB on a synthetic example with a known population prior and a strongly multimodal or highly correlated true posterior, using a diagonal-Gaussian inference network; if the learned prior stabilizes at a distribution whose induced marginal matches the observed data but whose KL to the true prior remains bounded away from zero, the practical convergence claim is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the paper is right, empirical Bayes no longer requires a tractable likelihood; any forward simulator can supply the needed likelihood information.
  • Simultaneous inference for heterogeneous units can be improved by a learned population prior even when posterior inference itself is amortized.
  • The theoretical convergence gives a target: improving the amortized family and simulator fidelity should bring practical SBEB closer to the population-prior fixed point.
  • Because the method samples from the average posterior rather than reweighting to a fixed prior, it suggests a simple way to focus simulation on observed data while still learning the prior.
  • The learned prior can also serve as a diagnostic: deviations between the average posterior and the initial prior reveal where the fixed prior was misspecified.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the posterior family is expressive enough, the EM interpretation suggests SBEB could be accelerated with stochastic or incremental EM variants, or annealed to escape poor fixed points.
  • The population-prior target depends on the observed empirical distribution; for small n, averaging approximate posteriors over only n units will be noisy, so shrinkage or regularization of the posterior average may be needed — an extension the paper does not explore.
  • A testable extension: on a simulator with known population prior and an intentionally misspecified posterior family, compare the fixed point to the true prior; this would isolate approximation error from algorithmic convergence.
  • The method could be coupled with predictive model checks: the learned prior's induced marginal can be scored against held-out observations to assess simulator adequacy.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes simulation-based empirical Bayes (SBEB), a method that iteratively refines an empirical Bayes prior inside a simulation-based inference (SBI) loop. The prior update is the population-prior identity (Eq. 5): average the posterior under the current prior over the observed data. Under an oracle with exact posteriors and population expectations, the paper proves monotone improvement of the marginal (Prop. 1), prior KL descent (Prop. 2), marginal convergence (Cor. 1), and geometric convergence under a strong identifiability condition (Theorem 1). Empirically, SBEB is compared with fixed-prior SBI and per-unit fits on five synthetic simulators, a discrete-choice dataset, and U.K. measles data, with the claimed outcome of improved held-out predictive accuracy.

Significance. If the results hold, SBEB is a conceptually clean extension of empirical Bayes to implicit-likelihood models, and the oracle convergence theory is a genuine contribution. The proofs in Appendix A are readable and appear mathematically correct, and the connection to EM is insightful. The paper is honest about the oracle assumptions and about the need for an expressive amortized family. However, the empirical evaluation does not separate the effect of learning an EB prior from the effect of sequentially focusing simulations on observed data, and no formal misspecification analysis is provided for the practical algorithm. These gaps undermine the paper's central empirical claim in its current form.

major comments (2)
  1. [§5.2–5.3, Figures 2–3] Algorithm 2 changes two things relative to fixed-prior SBI: it updates the prior, and it adaptively focuses simulations on the observed units by sampling z from the average amortized posterior. The real-data experiments compare only against fixed-prior SBI and per-unit fits. Sequential SBI methods (e.g., SNPE) also focus simulations on observed data while keeping a fixed prior, and such focusing is known to improve amortized posterior accuracy at the observed x's. Without a control that uses the same sequential simulation-adaptation loop but does not update the prior, or an ablation that updates the prior without focusing simulations, the reported gains cannot be attributed to EB prior learning. This is load-bearing because the abstract's central claim is that SBEB improves over SBI with a fixed prior by learning the population prior.
  2. [Section 3 and Algorithm 2, Eq. (9)] All convergence results are for an oracle with exact posteriors and population expectations. In the practical algorithm, Eq. (9) uses amortized posteriors from a diagonal-Gaussian inference network. If that family cannot represent the posterior under the evolving prior, the averaged posterior is a biased estimate of the population prior, and the iterative refinement may converge to a different fixed point. Section 6 acknowledges that SBI needs an expressive family, but this does not address the gap between the theory and the algorithm used in the experiments. Please provide a misspecification analysis (e.g., a stationarity characterization or a bias bound) or alternatively reframe the empirical claims as demonstrating the practical algorithm's utility without claiming it recovers the EB prior. At minimum, report diagnostics on synthetic data with a known population prior to assess whethe
minor comments (5)
  1. [After Corollary 1] The weak-convergence consequence is stated informally ('If the priors are tight and the simulator maps weakly convergent priors...'). These assumptions are not formalized or proved. Please either state them as a lemma with a proof or move the claim to a remark.
  2. [Definition 1] The local version of λ-strong prior identifiability is mentioned ('for some R>0') but not defined precisely. It is unclear whether Theorem 1 requires the global or local condition, and whether the constant λ can be verified for any of the simulators used in the experiments.
  3. [Appendix A, Eq. (13) and Definition 1] The symbol λ is used both for the Lagrange multiplier in the variational derivation (Eq. 13) and for the strong-identifiability constant in Definition 1. This is confusing and should be resolved by renaming one of them.
  4. [Section 5] Hyperparameters are not fully reported: the number of rounds T and the Monte Carlo batch size B in Algorithm 2 are not specified for any experiment. Please report these values, as they affect the runtime and the quality of the approximation.
  5. [Table 1] Table 1 reports a single MSE value per simulator without standard errors or multiple seeds, whereas Figures 2 and 3 include error bars. Adding uncertainty quantification to Table 1 would strengthen the comparison.

Circularity Check

0 steps flagged

No significant circularity: SBEB's fixed-point structure is explicit, convergence proof is independent, and held-out evaluations are not forced.

full rationale

I examined the derivation chain: the population-prior identity (Eq. 5), the oracle update (Eq. 10), Propositions 1–2, Corollary 1, Theorem 1, Algorithms 1–2, and the empirical protocols. The only self-referential structure is Eq. 5, where p(z) appears on both sides, but the paper explicitly identifies this as a fixed-point equation and independently characterizes the target p*(z) by the marginal-matching EB criterion (Eq. 3). The convergence proof is a genuine EM-style argument: under stated assumptions (exact posteriors, population expectations, and λ-strong prior identifiability), the oracle update contracts toward p*(z); it does not assume its conclusion. The practical limitation that a misspecified amortized family biases the averaged posterior is acknowledged in Section 6, not hidden. Empirical claims are evaluated on held-out choice tasks and held-out measles periods, so they are not fitted-input predictions. The absence of a sequential-SBI control in the experiments is a potential alternative-explanation concern (adaptive focusing may drive gains), but that is a correctness/confound issue, not circularity by construction. Self-citations appear only as background or as a practical caveat, and none is load-bearing. Therefore the paper does not exhibit a circular derivation.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The method relies on EB realizability, simulator fidelity, oracle-like posterior approximation for the theory, and identifiability for prior convergence. Experimental controls (initial prior, T, B) are hand-chosen and underreported.

free parameters (3)
  • Initial fixed prior p_0(z) = Measles: η ~ N(m,I) with m=(0.916,-0.134,0,-2.398,-2.996,-0.080,-1.237,-1.735); choice: N(0, diag(τ^2)) with τ=2 (τ=1 fo
    SBEB starts from this prior; the choice affects the trajectory and also defines the fixed-prior SBI baseline. The measles mean is hand-chosen from 'plausible epidemic values' (Appendix C).
  • Number of rounds T and MC samples B = B=400 for simulation studies (Section 5.1); T not reported anywhere
    Algorithm 2's behavior depends on T and B; the paper does not state T for any experiment, and B for real studies is not given. These hand-chosen controls affect the empirical results.
  • Amortized posterior family (diagonal Gaussian via two-layer MLP) = Diagonal Gaussian; two-layer perceptron (simulation studies); unspecified for choice/measles
    The expressiveness of q is a modeling choice on which SBEB's empirical performance depends; the paper acknowledges this in §6.
axioms (5)
  • domain assumption The observed data are i.i.d. from a population distribution p⋆(x) induced by a single population prior p⋆(z).
    The EB setup (Eq. 3 and Section 2.1) assumes the model marginal can match the population; if the population is heterogeneous with no common prior, the EB criterion is undefined.
  • domain assumption The simulator p(x|z) is an accurate conditional model of the data-generating process.
    Section 6: 'SBEB is built on the premise that the simulator gives an accurate conditional distribution of datasets.' If the simulator is wrong, both SBI and SBEB fail.
  • ad hoc to paper In Section 3, the fitted inference network q represents the exact posterior under the current prior, and population expectations are available.
    Stated at the start of Section 3: 'We assume that the fitted inference network can represent the exact posterior under the current prior, and that population expectations are available.' This is the oracle assumption; the practical algorithm relaxes it without a corresponding guarantee.
  • ad hoc to paper For weak prior convergence after Corollary 1, the sequence of priors is tight and the simulator maps weakly convergent priors to weakly convergent marginals.
    Section 3: 'If the priors are tight and the simulator maps weakly convergent priors to weakly convergent marginals...' These conditions are not proven for the update.
  • domain assumption Prior identifiability (Definition 1) holds for the simulators used.
    Theorem 1's geometric convergence requires λ-strong prior identifiability; the paper does not verify this condition for the five simulators or the real-data models.

pith-pipeline@v1.3.0-alltime-deepseek · 13687 in / 15168 out tokens · 145527 ms · 2026-08-01T06:32:01.111108+00:00 · methodology

0 comments
read the original abstract

Empirical Bayes (EB) performs simultaneous inference across many related latent variables. Classical EB assumes that the likelihood p(x | z) is tractable. In many scientific applications, however, the likelihood is available only through a simulator. This paper develops EB for such implicit likelihoods. We introduce simulation-based empirical Bayes (SBEB), which connects nonparametric EB to simulation-based inference (SBI). SBEB computes EB estimates without an explicit density by using the observed data, simulator samples, and an amortized inference network. SBEB iteratively refines the fitted EB prior toward the population prior. With several scientific simulators and real-world data, we demonstrate that SBEB improves accuracy over SBI with a fixed prior.

Figures

Figures reproduced from arXiv: 2607.21843 by Cheng Zhang, David M. Blei, Diana Cai, Xinwei Shen.

Figure 1
Figure 1. Figure 1: Biweekly measles reports from six UK cities. Each panel is one city. The full dataset contains 40 cities observed from 1944 to 1965 (Dalziel et al., 2016). p(xi | zi). Second, the problem is simultaneous. We have many related cities and many related inference problems. Their epidemic parameters differ, but they come from a common population. These two features of the problem suggest two statistical ideas. … view at source ↗
Figure 2
Figure 2. Figure 2: Heterogeneous choice study. Held-out negative log predictive density on tasks 13–16 as a function of the number of observed training tasks per person. Each point averages three independent runs; bands show one standard error. Lower is better. tasks. For tasks after R, the choice entry is masked. The network maps s (R) i to a diagonal Gaussian approximation q(zi ; s (R) i , φ). We compare fixed-prior SBI, S… view at source ↗
Figure 3
Figure 3. Figure 3: Measles outbreak study. Held-out mean log-incidence absolute error as a function of the number of observed training biweeks per city. Training uses the R biweekly observations immediately before observation 200. Evaluation is always on observations 200 through 300. Each point averages three independent runs; bands show one standard error. Lower is better. 6 Discussion Simulation-based empirical Bayes (SBEB… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references

  1. [1]

    Journal of the American Statistical Association , volume=

    Strictly proper scoring rules, prediction, and estimation , author=. Journal of the American Statistical Association , volume=. 2007 , publisher=

    2007

  2. [2]

    T. M. Cover and J. A. Thomas , title =

  3. [3]

    Robbins , title =

    H. Robbins , title =. Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability , year =

  4. [4]

    B. P. Carlin and T. A. Louis , title =

  5. [5]

    Efron , journal =

    B. Efron , journal =. Bayes, oracle

  6. [6]

    Laird , title =

    N. Laird , title =. Journal of the American Statistical Association , year =

  7. [7]

    Goyal and S

    A. Goyal and S. Jain and A. Lahlou and T. Pham and A. Courville and Y. Bengio , title =. Conference on Uncertainty in Artificial Intelligence , year =

  8. [8]

    Ravi and J

    S. Ravi and J. Ba , title =. Neural Information Processing Systems Workshop on Meta--Learning , year =

  9. [9]

    Yoon and T

    J. Yoon and T. Kim and O. Dia and S. Kim and Y. Bengio and S. Ahn , title =. Neural Information Processing Systems , year =

  10. [10]

    Kucukelbir and D

    A. Kucukelbir and D. Tran and R. Ranganath and A. Gelman and D. M. Blei , title =. Journal of Machine Learning Research , year =

  11. [11]

    Marino and Y

    J. Marino and Y. Yue and S. Mandt , title =. International Conference on Machine Learning , year =

  12. [12]

    C. C. Margossian and D. M. Blei , title =. Conference on Uncertainty in Artificial Intelligence , year =

  13. [13]

    Tomczak and M

    J. Tomczak and M. Welling , title =. International Conference on Artificial Intelligence and Statistics , year =

  14. [14]

    McInerney and R

    J. McInerney and R. Ranganath and D. M. Blei , title =. Neural Information Processing Systems , year =

  15. [15]

    D. P. Kingma and T. Salimans and R. Jozefowicz and X. Chen and I. Sutskever and M. Welling , title =. Neural Information Processing Systems , year =

  16. [16]

    M. A. Beaumont and W. Zhang and D. J. Balding , title =. Genetics , year =

  17. [17]

    S. A. Sisson and Y. Fan and M. Beaumont , title =. WIREs Computational Statistics , year =

  18. [18]

    Papamakarios and D

    G. Papamakarios and D. Sterratt and I. Murray , title =. International Conference on Artificial Intelligence and Statistics , year =

  19. [19]

    Cranmer and J

    K. Cranmer and J. Brehmer and G. Louppe , title =. Proceedings of the National Academy of Sciences , year =

  20. [20]

    Deistler and J

    M. Deistler and J. Boelts and P. Steinbach and G. Moss and T. Moreau and M. Gl. Simulation-Based Inference: A Practical Guide , year =

  21. [21]

    J. M. Lueckmann and J. Boelts and D. S. Greenberg and P. J. Gon. Benchmarking Simulation-Based Inference , journal =. 2021 , volume =

    2021

  22. [22]

    Perrin and C

    C. Perrin and C. Michel and V. Andr. Improvement of a Parsimonious Model for Streamflow Simulation , journal =. 2003 , volume =

    2003

  23. [23]

    A. J. Newman and M. P. Clark and K. Sampson and A. Wood and L. E. Hay and A. Bock and R. J. Viger and D. Blodgett and L. Brekke and J. R. Arnold and T. Hopson and Q. Duan , title =. Hydrology and Earth System Sciences , year =

  24. [24]

    Addor and A

    N. Addor and A. J. Newman and N. Mizukami and M. P. Clark , title =. Hydrology and Earth System Sciences , year =

  25. [25]

    Tejero--Cantero and M

    A. Tejero--Cantero and M. Boelts and J. M. Lueckmann and P. Greenberg and A. Gonc. sbi: A Toolkit for Simulation--Based Inference , journal =. 2020 , volume =

    2020

  26. [26]

    Efron , journal =

    B. Efron , journal =. Two Modeling Strategies for Empirical

  27. [27]

    H. B. Experience Rating and Credibility , journal =. 1967 , doi =

    1967

  28. [28]

    Efron and C

    B. Efron and C. Morris , title =. Journal of the American Statistical Association , volume =. 1973 , doi =

    1973

  29. [29]

    Stein , title =

    C. Stein , title =. Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability , editor =

  30. [30]

    Biometrika , volume =

    Empirical Bayes deconvolution estimates , author =. Biometrika , volume =. 2016 , publisher =

    2016

  31. [31]

    Peter Rossi , year =

  32. [32]

    Quantitative Marketing and Economics , year =

    Greg Allenby and Jeff Brazell and John Howell and Peter Rossi , title =. Quantitative Marketing and Economics , year =

  33. [33]

    Journal of Law and Economics , year =

    Greg Allenby and Jeff Brazell and John Howell and Peter Rossi , title =. Journal of Law and Economics , year =

  34. [34]

    Rossi and Greg M

    Peter E. Rossi and Greg M. Allenby and Robert McCulloch , title =

  35. [35]

    Frontiers in Econometrics , editor =

    Daniel McFadden , title =. Frontiers in Econometrics , editor =

  36. [36]

    Train , title =

    Kenneth E. Train , title =

  37. [37]

    Strogatz , title =

    Steven H. Strogatz , title =

  38. [38]

    Lotka , title =

    Alfred J. Lotka , title =

  39. [39]

    Nature , year =

    Vito Volterra , title =. Nature , year =

  40. [40]

    R. A. Fisher , title =

  41. [41]

    Genetics , year =

    Sewall Wright , title =. Genetics , year =

  42. [42]

    Psychological Review , year =

    Roger Ratcliff , title =. Psychological Review , year =

  43. [43]

    Neural Computation , year =

    Roger Ratcliff and Gail McKoon , title =. Neural Computation , year =

  44. [44]

    Navarro and Ian G

    Daniel J. Navarro and Ian G. Fuss , title =. Journal of Mathematical Psychology , year =

  45. [45]

    Wiecki and Imri Sofer and Michael J

    Thomas V. Wiecki and Imri Sofer and Michael J. Frank , title =. Frontiers in Neuroinformatics , year =

  46. [46]

    Cavanagh and Thomas V

    James F. Cavanagh and Thomas V. Wiecki and Michael X. Cohen and Christina M. Figueroa and Johan Samanta and Scott J. Sherman and Michael J. Frank , title =. Nature Neuroscience , year =

  47. [47]

    Hauser and Michael Moutoussis and Rani Moran and Mehdi Keramati and NSPN Consortium and Raymond J

    Nitzan Shahar and Tobias U. Hauser and Michael Moutoussis and Rani Moran and Mehdi Keramati and NSPN Consortium and Raymond J. Dolan , title =. PLOS Computational Biology , year =

  48. [48]

    Time Series Modelling of Childhood Diseases: A Dynamical Systems Approach , journal =

    B. Time Series Modelling of Childhood Diseases: A Dynamical Systems Approach , journal =. 2000 , volume =

    2000

  49. [49]

    Ottar N. Bj. Dynamics of Measles Epidemics: Estimating Scaling of Transmission Rates Using a Time Series. Ecological Monographs , year =

  50. [50]

    King and Dao Nguyen and Edward L

    Aaron A. King and Dao Nguyen and Edward L. Ionides , title =. Journal of Statistical Software , year =

  51. [51]

    Korevaar and C

    Hannah M. Korevaar and C. Jessica E. Metcalf and Bryan T. Grenfell , title =. Journal of the Royal Society Interface , year =

  52. [52]

    Dalziel and Ottar N

    Benjamin D. Dalziel and Ottar N. Bj. Persistent Chaos of Measles Epidemics in the Prevaccination United States Caused by a Small Change in Seasonal Transmission Patterns , journal =. 2016 , volume =

    2016

  53. [53]

    Reference posterior distributions for

    Bernardo, Jose M , journal =. Reference posterior distributions for. 1979 , publisher =

    1979

  54. [54]

    Noninformative priors and

    Berger, James O and Yang, Ruo-Yong , journal =. Noninformative priors and. 1994 , publisher =

    1994

  55. [55]

    International Conference on Machine Learning , year =

    Learning Likelihood-Free Reference Priors , author =. International Conference on Machine Learning , year =

  56. [56]

    Neural Information Processing Systems , year =

    Flow Matching for Scalable Simulation-Based Inference , author =. Neural Information Processing Systems , year =

  57. [57]

    Neural Information Processing Systems , fpages =

    Truncated proposals for scalable and hassle-free simulation-based inference , author =. Neural Information Processing Systems , fpages =

  58. [58]

    International Conference on Machine Learning , year =

    Automatic Posterior Transformation for Likelihood-Free Inference , author =. International Conference on Machine Learning , year =

  59. [59]

    Neural Information Processing Systems , year =

    Multilevel neural simulation-based inference , author =. Neural Information Processing Systems , year =

  60. [60]

    Neural Information Processing Systems , year =

    Learning Robust Statistics for Simulation-based Inference under Model Misspecification , author =. Neural Information Processing Systems , year =

  61. [61]

    Simulation-based

    Kelly, Ryan P and Warne, David J and Frazier, David T and Nott, David J and Gutmann, Michael U and Drovandi, Christopher , journal =. Simulation-based

  62. [62]

    Population empirical

    Kucukelbir, Alp and Blei, David M , booktitle =. Population empirical

  63. [63]

    Neural Information Processing Systems , year =

    Flexible statistical inference for mechanistic models of neural dynamics , author =. Neural Information Processing Systems , year =

  64. [64]

    Fast -Free Inference of Simulation Models with

    Papamakarios, George and Murray, Iain , booktitle =. Fast -Free Inference of Simulation Models with

  65. [65]

    Neural Empirical

    Vandegar, Maxime and Kagan, Michael and Wehenkel, Antoine and Louppe, Gilles , booktitle =. Neural Empirical

  66. [66]

    Neural Information Processing Systems , year =

    Sourcerer: Sample-based maximum entropy source distribution estimation , author =. Neural Information Processing Systems , year =

  67. [67]

    Neural Information Processing Systems , pages =

    Robust Neural Posterior Estimation and Statistical Model Criticism , author =. Neural Information Processing Systems , pages =