Pith. sign in

REVIEW 2 major objections 4 minor 228 references

Score matching isolates EIG double intractability so policy training becomes singly intractable and reusable.

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 · grok-4.5

2026-07-10 09:10 UTC pith:4ZA2KO6P

load-bearing objection Clean isolation of EIG double intractability into policy-independent score matching; additive NLE cost is real and the math holds. the 2 major comments →

arxiv 2607.08335 v1 pith:4ZA2KO6P submitted 2026-07-09 stat.ML cs.LG

Bayesian Experimental Design via Score Matching

Angus Phillips , Gavin Kerrigan , Tom Rainforth This is my paper
classification stat.ML cs.LG
keywords Bayesian experimental designexpected information gainscore matchingpolicy networksamortised inferencereparameterisationmarginal likelihood
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.

Policy-based Bayesian experimental design trains networks that choose the next experiment from past data, but the expected information gain is doubly intractable, so every gradient step usually needs nested sampling or a co-trained variational model. That multiplies cost by the number of policies one wants to try and leaves little budget for restarts or architecture search. This paper shows the information gain itself depends only on the realised designs and data, not on the policy parameters. The double intractability therefore lives in two policy-independent scores of the marginal likelihood. Those scores can be learned once by ordinary regression (marginal score matching) and then plugged into a reparameterised gradient that is only singly intractable. The expensive step becomes additive rather than multiplicative, so many competitive policies can be trained under a fixed likelihood budget and the best one selected.

Core claim

The double intractability of the expected information gain can be separated from policy optimisation: first solve a single policy-independent score-matching problem for the Stein and Fisher scores of the marginal likelihood, then substitute the learned scores into a reparameterised EIG gradient that is only singly intractable. This converts the usual multiplicative likelihood cost into an additive cost and makes multi-policy training cheap.

What carries the argument

Theorem 1 (reparameterised EIG gradient) isolates the intractable terms as the Stein score ∂/∂y log p(y|ξ) and Fisher score ∂/∂ξ log p(y|ξ); marginal score matching then regresses a shared network onto the conditional likelihood scores, recovering both targets without sampling the posterior.

Load-bearing premise

A score network trained only on samples from a hand-chosen design distribution must stay accurate on the trajectories later visited by trained policies, otherwise gradient bias ruins policy optimisation.

What would settle it

Train the score network once, then train many policies under a matched likelihood budget; if the best-of-P ScoreBED policies systematically underperform best-of-P nested or co-trained baselines on the same tasks, or if measured EIG-gradient bias fails to fall with score error, the claim fails.

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

If this is right

  • Under a fixed likelihood budget one can train many policy restarts or architectures and simply keep the best, without re-paying the double-intractability cost.
  • Gradient bias, not variance, becomes the dominant error once outer-sample size is moderate; score-network capacity and training budget therefore control final policy quality.
  • The same two-stage pattern applies to any variational approximation of the marginal or posterior that is independent of the policy.
  • Static-design special cases drop the Fisher-score term, lowering bias and variance even without score matching.

Where Pith is reading between the lines

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

  • Because the score targets live in design and observation space rather than parameter space, the method should scale more gracefully than nested posterior sampling when the latent dimension grows.
  • If the design sampler q(ξ) can be adapted online to the current policy support without destroying amortisation, the coverage assumption becomes far weaker.
  • The same score network could be reused across related models that share the same observation and design spaces, amortising design cost over a family of experiments.

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 / 4 minor

Summary. The paper proposes SCOREBED, a two-stage method for policy-based Bayesian experimental design. It first trains a policy-independent score network via marginal score matching (MSM, Eq. 6) to approximate the Stein and Fisher scores of the marginal p(y|ξ). These scores are then substituted into a reparameterised EIG gradient (Theorem 1 / Eq. 3), yielding a singly intractable estimator for policy SGD. The claim is that this isolates double intractability from policy learning, converting the usual multiplicative nested-sampling NLE cost into an additive cost and thereby enabling cheap multi-policy training (architecture search, restarts) under a fixed likelihood budget. Experiments on source location finding, three dynamical systems, and gravimetry compare against PCE, pre-trained and joint variational bounds, and IO-SMC2 under matched NLE budgets, reporting both lower and upper EIG bounds for P=1 and multi-restart protocols.

Significance. If the isolation argument holds in practice, SCOREBED is a useful engineering contribution for policy-based BED: amortising the hard nested terms once, then reusing them for many policy trainings, is a natural and previously under-exploited separation. Theorem 1 cleanly isolates the intractable scores; the static-design simplification and the Lipschitz error bound (Theorem 3 / Appendix B) are sound; the bias–variance analysis (Appendix C) explains the observed gradient plateaus. Code release, fixed-NLE protocols, and dual EIG bounds strengthen the empirical case. The main practical value is the ability to train multiple competitive policies without repeating the expensive nested work—an advantage that is real under the paper’s own budget accounting and that existing nested or co-trained methods do not share.

major comments (2)
  1. [§3 / App. E.3] §3, Appendix D.4.1 and E.3: the central practical claim—that a single policy-independent score network yields usable gradients for later policies—rests on the support of the hand-chosen design sampler q(ξ). Table 14 shows that an isotropic Gaussian sampler collapses EIG to ~2.86 while scale-aware or low-rank samplers recover ~11. The paper documents the failure mode but does not give a systematic way to choose or validate q(ξ) for a new problem; without that, the additive-cost advantage is conditional on a non-trivial modelling choice that is currently problem-specific.
  2. [§5.2 / App. E.5] §5.2 and Appendix E.5 (Figs. 5–7): on the dynamical systems the score-network stage is unstable (multiple seeds required; Jacobian condition numbers of trained policies correlate with score MSE). The multi-seed P=3/50 protocol mitigates this, but it means part of the “policy” budget is still spent on score restarts. The paper should quantify how much of the claimed additive saving survives when score training itself needs restarts, and whether the Lipschitz constant L in Theorem 3 becomes large enough on non-Markovian trajectories to make the error bound uninformative.
minor comments (4)
  1. [§3] Clarify early that the “marginal” p(y|ξ) is not the policy-induced data distribution p(y;π_ϕ); the distinction is made in §3 but is easy to miss and is load-bearing for the policy-independence claim.
  2. [Table 1] Table 1 / Fig. 1: for d=3,K=10, SCOREBED (P=1) slightly outperforms P=5; a short discussion of when extra restarts help versus when the score budget is still the bottleneck would help readers allocate NLE.
  3. [App. A.2] Appendix A.2: the reparameterisation map g is deferred; a short explicit form for the common additive-noise case (already treated in A.3) would make the gradient expression easier to implement from the main text alone.
  4. [§4] Related work: concurrent Huang et al. (2026a) is noted as complementary; a one-sentence comparison of what is amortised (belief representations vs. scores) would sharpen the positioning.

Circularity Check

0 steps flagged

No circularity: EIG gradient isolation and MSM are derived from first principles; score regression targets known likelihood scores, not fitted EIG quantities.

full rationale

The load-bearing chain is: (i) IG depends on data/designs only, not on the policy that proposed them (Appendix A.1, design proposals cancel in the posterior); (ii) reparameterised differentiation of the mutual-information form of IT yields Theorem 1 / Eq. (3), isolating the Stein and Fisher scores of the design-conditional marginal p(y|ξ); (iii) the MSM objective (Eq. 6) is ordinary supervised regression of a network onto ∇_{y,ξ} log p(y|ξ,θ), which is available from the model likelihood, and whose unique minimiser is the true marginal score by the identity (Eq. 5); (iv) the learned scores are substituted into the gradient estimator, with ℓ2 gradient error bounded by Lipschitz constant times mean score error (Theorem 3 / Appendix B). None of these steps defines the target EIG or its gradient in terms of a fitted parameter, nor renames a fit as a prediction. Self-citations (Foster et al., Rainforth et al., Hyvärinen, Song et al.) supply background BED and score-matching tools; the central gradient expression and MSM construction are derived in-paper and validated against external baselines under fixed NLE budgets. Residual risks (support of q(ξ), Jacobian conditioning) are approximation/support assumptions, not circular reductions.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard measure-theoretic and reparameterisation assumptions of modern BED, plus the modelling restrictions needed for Stein/Fisher scores to exist, and on the practical claim that a score network trained under a fixed design distribution remains accurate on policy-induced trajectories. No new physical entities are postulated; free parameters are ordinary neural-network and sampling hyperparameters.

free parameters (3)
  • design sampling distribution q(ξ) (scale hyperpriors, temporal correlation, mixture weights)
    Chosen by hand to cover regions expected under good policies; ablation shows isotropic Gaussian fails while random-scale or low-rank samplers succeed. Directly affects score accuracy on later policy trajectories.
  • score-network architecture and training budget (transformer depth, λ_y weighting, warmup-cosine schedule)
    Capacity and loss weighting control residual score error, which bounds gradient bias via Theorem 3; budgets are adjusted to meet fixed NLE totals.
  • number of policy restarts P and allocation of NLE between score and policy stages
    Determines how much of the claimed multi-policy advantage is realised; varied systematically but still a free experimental choice.
axioms (4)
  • domain assumption Observations depend on the policy only through the realised designs (ignorability / Rubin 1976), so the posterior and the information gain are policy-independent given (y,ξ).
    Stated and proved in Appendix A.1; required for the separation of intractabilities from the policy.
  • domain assumption Design and observation spaces are continuous and the data-generating process admits a differentiable reparameterisation, so Stein and Fisher scores exist and the reparameterised gradient (Eq. 3) is valid.
    Explicitly listed under Applicability (§3); excludes discrete or non-reparameterisable models.
  • standard math The reparameterisation map ϕ ↦ (y,ξ) is almost-surely L-Lipschitz, allowing the ℓ₂ gradient error to be bounded by L times mean score error (Theorem 3).
    Invoked via Rademacher’s theorem; needed for the formal guarantee that better scores yield better gradients.
  • domain assumption Explicit likelihood evaluations are available (or an implicit score-matching loss can be substituted).
    Required for the MSM objective; SSM is mentioned as a fallback for implicit models.
invented entities (1)
  • Marginal score matching (MSM) objective and the SCOREBED two-stage procedure no independent evidence
    purpose: Amortise the intractable marginal scores once, then reuse them for any number of policy trainings.
    New algorithmic construct; independent evidence is the empirical multi-policy results and the gradient-bias analysis, not an external physical prediction.

pith-pipeline@v1.1.0-grok45 · 44355 in / 3255 out tokens · 34717 ms · 2026-07-10T09:10:52.835992+00:00 · methodology

0 comments
read the original abstract

Policy-based approaches to Bayesian experimental design (BED) allow the learning of deep policy networks that adaptively make intelligent design decisions based on previously collected data. However, the training of such policies is often held back by a fundamental challenge: the double intractability of the expected information gain (EIG). This necessitates expensive or complex approximations that restrict the effort one can invest in optimising the policy itself. To address this, we show that the double intractability of the EIG can be isolated from the policy learning by first solving a score matching problem that is independent of the policy used, then using the learned score approximation to train the policy in a singly intractable manner. This turns the key multiplicative cost into an additive one and reduces the computational burden on the policy training itself, making it far cheaper to train the policy multiple times when needed, e.g. for architecture search, hyperparameter tuning, or avoiding local optima. In our experiments we train multiple competitive policies without inducing a multiplicative cost in likelihood evaluations, which can increase performance by allowing us to select the best policy even without performing hyperparameter or architecture searches.

Figures

Figures reproduced from arXiv: 2607.08335 by Angus Phillips, Gavin Kerrigan, Tom Rainforth.

Figure 1
Figure 1. Figure 1: Left: EIG vs NLE (P = 1) curves for each method on the d = 3, K = 10 location finding setting. Right: Comparing policy training curves for SCOREBED and PCE on the same problem. The exact NLE cost of each method is calculated in Ap￾pendix D.7 and the configurations of each baseline to respect the exact budget are detailed in Appendix D.8. In order to accommodate different numbers of policies P while respect… view at source ↗
Figure 3
Figure 3. Figure 3: EIG lower bound versus EIG gradient squared-bias. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Stability of SCOREBED with respect to score training. Each curve presents the final policy EIG under the P = 1 protocol against number of score training iterations. Left: d = 2, K = 2 location finding. Right: cart-pole. trajectories as low-rank projections of Gaussian noise, ξ = Φz, where Φ ∈ R T ×r is a time-basis matrix and z ∈ R r×d are Gaussian coefficients, zij ∼ N (0, s2 ) with an overall scale drawn… view at source ↗
Figure 5
Figure 5. Figure 5: Scatter plots showing correlation between EIG gradient bias and final policy performance, with points representing [PITH_FULL_IMAGE:figures/full_fig_p033_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Histograms showing the distribution of Jacobian condition number between the score training distribution and the [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: EIG gradient MSE curves for SCOREBED and PCE. PCE uses M = 19 and PCE∗ uses M = 199 inner samples on the location finding tasks, while PCE (P = 50) uses M = 7 and PCE (P = 1) uses M = 399 on the dynamical systems tasks. These values are chosen according to the fixed NLE budgets for each task as per Appendix D.8. From top left to bottom right: location finding (d = 2, K = 2), location finding (d = 3, K = 10… 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

228 extracted references · 228 canonical work pages · 32 internal anchors

  1. [1]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Transformer Neural Processes: Uncertainty-Aware Meta Learning Via Sequence Modeling , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , editor =

    work page 2022

  2. [2]

    1972 , publisher=

    Bayesian statistics: A review , author=. 1972 , publisher=

    work page 1972

  3. [3]

    Telford, W. M. and Geldart, L. P. and Sheriff, R. E. , year=. Gravity Methods , booktitle=

    work page

  4. [4]

    Proceedings of the 35th

    Chen, Zhao and Badrinarayanan, Vijay and Lee, Chen-Yu and Rabinovich, Andrew , month = jul, year =. Proceedings of the 35th

    work page

  5. [5]

    Vincent, Benjamin Thomas and Rainforth, Tom , month = oct, year =. The. doi:10.31234/osf.io/yehjb , abstract =

    work page doi:10.31234/osf.io/yehjb

  6. [6]

    Durkan, Conor and Bekasov, Artur and Murray, Iain and Papamakarios, George , year =. Neural. Advances in

    work page

  7. [7]

    Bradbury, James and Frostig, Roy and Hawkins, Peter and Johnson, Matthew and Leary, Chris and Maclaurin, Dougal and Necula, George and Paszke, Adam and VanderPlas, Jake and Wanderman-Milne, Skye and Zhang, Qiao , year =

    work page

  8. [8]

    and Guan, Cong and Rainforth, Tom , month = oct, year =

    Hedman, Marcel and Ivanova, Desi R. and Guan, Cong and Rainforth, Tom , month = oct, year =. Step-. Proceedings of the 42nd

    work page

  9. [9]

    JADAI: Jointly Amortizing Adaptive Design and Bayesian Inference

    Bracher, Niels and Kühmichel, Lars and Ivanova, Desi R. and Intes, Xavier and Bürkner, Paul-Christian and Radev, Stefan T. , month = dec, year =. doi:10.48550/arXiv.2512.22999 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2512.22999

  10. [10]

    Sutton, Richard S and McAllester, David and Singh, Satinder and Mansour, Yishay , year =. Policy. Advances in

    work page

  11. [11]

    Information

    Barber, David and Agakov, Felix , year =. Information. Advances in

    work page

  12. [12]

    International Journal for Uncertainty Quantification , author =

    Gradient-based stochastic optimization methods in. International Journal for Uncertainty Quantification , author =. doi:10.1615/Int.J.UncertaintyQuantification.2014006730 , abstract =

    work page doi:10.1615/int.j.uncertaintyquantification.2014006730

  13. [15]

    Bayesian Experimental Design via Contrastive Diffusions

    Iollo, Jacopo and Heinkelé, Christophe and Alliez, Pierre and Forbes, Florence , month = mar, year =. Bayesian. doi:10.48550/arXiv.2410.11826 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.11826

  14. [17]

    Statistics & Probability Letters , author =

    Bayesian information in an experiment and the. Statistics & Probability Letters , author =. 2016 , pages =. doi:10.1016/j.spl.2016.01.014 , abstract =

    work page doi:10.1016/j.spl.2016.01.014 2016

  15. [18]

    Advances in Neural Information Processing Systems , volume=

    ALINE: Joint amortization for Bayesian inference and active data acquisition , author=. Advances in Neural Information Processing Systems , volume=

    work page

  16. [19]

    Neural Computation , author =

    Regularized. Neural Computation , author =. 2016 , pages =. doi:10.1162/NECO_a_00844 , abstract =

    work page doi:10.1162/neco_a_00844 2016

  17. [20]

    Sasaki, Hiroaki and Noh, Yung-Kyun and Sugiyama, Masashi , month = feb, year =. Direct. Proceedings of the

    work page

  18. [21]

    Understanding the

    Song, Jiaming and Ermon, Stefano , month = sep, year =. Understanding the

    work page

  19. [22]

    Practical Bayesian Optimization of Machine Learning Algorithms

    Snoek, Jasper and Larochelle, Hugo and Adams, Ryan P. , month = aug, year =. Practical. doi:10.48550/arXiv.1206.2944 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1206.2944

  20. [23]

    and Rainforth, Tom , month = oct, year =

    Kerrigan, Gavin and Naesseth, Christian A. and Rainforth, Tom , month = oct, year =. A. doi:10.48550/arXiv.2510.14848 , abstract =

    work page doi:10.48550/arxiv.2510.14848

  21. [24]

    Certified Monotonic Neural Networks

    Liu, Xingchao and Han, Xing and Zhang, Na and Liu, Qiang , month = dec, year =. Certified. doi:10.48550/arXiv.2011.10219 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2011.10219 2011

  22. [25]

    Constrained Monotonic Neural Networks

    Runje, Davor and Shankaranarayana, Sharath M. , month = may, year =. Constrained. doi:10.48550/arXiv.2205.11775 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2205.11775

  23. [26]

    Learning

    Cho, Kyunghyun and van Merriënboer, Bart and Gulcehre, Caglar and Bahdanau, Dzmitry and Bougares, Fethi and Schwenk, Holger and Bengio, Yoshua , editor =. Learning. Proceedings of the 2014. 2014 , pages =. doi:10.3115/v1/D14-1179 , urldate =

    work page doi:10.3115/v1/d14-1179 2014

  24. [27]

    danielward27/flowjax: v17.1.2 , shorttitle =

    Ward, Daniel and Hickling, Tennessee and Mould, Matthew and Seyboldt, Adrian and Turok, Gilad , month = may, year =. danielward27/flowjax: v17.1.2 , shorttitle =. doi:10.5281/zenodo.15357881 , abstract =

    work page doi:10.5281/zenodo.15357881

  25. [28]

    Papamakarios, George and Pavlakou, Theo and Murray, Iain , year =. Masked. Advances in

    work page

  26. [29]

    Journal of Machine Learning Research , author =

    Normalizing. Journal of Machine Learning Research , author =. 2021 , pages =

    work page 2021

  27. [30]

    and Fischer, Ian and Dillon, Joshua V

    Alemi, Alexander A. and Fischer, Ian and Dillon, Joshua V. and Murphy, Kevin , month = feb, year =. Deep

    work page

  28. [31]

    Yu, Longlin and Zhang, Cheng , month = sep, year =. Semi-

    work page

  29. [32]

    Wen, Liangjian and Zhou, Yiji and He, Lirong and Zhou, Mingyuan and Xu, Zenglin , month = sep, year =. Mutual

    work page

  30. [33]

    What regularized auto-encoders learn from the data-generating distribution , volume =. J. Mach. Learn. Res. , author =. 2014 , pages =

    work page 2014

  31. [34]

    Sticking the

    Roeder, Geoffrey and Wu, Yuhuai and Duvenaud, David K , year =. Sticking the. Advances in

    work page

  32. [35]

    Proceedings of the 37th

    Lim, Jae Hyun and Courville, Aaron and Pal, Christopher and Huang, Chin-Wei , month = nov, year =. Proceedings of the 37th

    work page

  33. [37]

    He, Jiajun and Du, Yuanqi and Vargas, Francisco and Zhang, Dinghuai and Padhy, Shreyas and OuYang, RuiKang and Gomes, Carla and Hernández-Lobato, José Miguel , month = apr, year =. No. doi:10.48550/arXiv.2502.06685 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.06685

  34. [38]

    Denoising

    Vargas, Francisco and Grathwohl, Will Sussman and Doucet, Arnaud , month = sep, year =. Denoising

    work page

  35. [39]

    Zhang, Qinsheng and Chen, Yongxin , month = oct, year =. Path

    work page

  36. [40]

    Salimans, Tim and Ho, Jonathan , month = apr, year =. Should

    work page

  37. [41]

    Noise-contrastive estimation:

    Gutmann, Michael and Hyvärinen, Aapo , month = mar, year =. Noise-contrastive estimation:. Proceedings of the

    work page

  38. [42]

    Poole, Ben and Ozair, Sherjil and Oord, Aaron Van Den and Alemi, Alex and Tucker, George , month = may, year =. On. Proceedings of the 36th

    work page

  39. [43]

    Box, G. E. P. and Lucas, H. L. , year =. Design of. Biometrika , publisher =. doi:10.2307/2332810 , number =

    work page doi:10.2307/2332810

  40. [45]

    and Jaakkola, Tommi S

    Ajay, Anurag and Du, Yilun and Gupta, Abhi and Tenenbaum, Joshua B. and Jaakkola, Tommi S. and Agrawal, Pulkit , month = sep, year =. Is

    work page

  41. [46]

    Spectral

    Phillips, Angus and Seror, Thomas and Hutchinson, Michael John and Bortoli, Valentin De and Doucet, Arnaud and Mathieu, Emile , month = nov, year =. Spectral

    work page

  42. [47]

    Tancik, Matthew and Srinivasan, Pratul and Mildenhall, Ben and Fridovich-Keil, Sara and Raghavan, Nithin and Singhal, Utkarsh and Ramamoorthi, Ravi and Barron, Jonathan and Ng, Ren , year =. Fourier. Advances in

    work page

  43. [48]

    Attention is

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, Ł ukasz and Polosukhin, Illia , year =. Attention is. Advances in

    work page

  44. [49]

    Pope, Phil and Zhu, Chen and Abdelkader, Ahmed and Goldblum, Micah and Goldstein, Tom , month = oct, year =. The

    work page

  45. [50]

    and Ross, Brendan Leigh and Cresswell, Jesse C

    Brown, Bradley CA and Caterini, Anthony L. and Ross, Brendan Leigh and Cresswell, Jesse C. and Loaiza-Ganem, Gabriel , month = sep, year =. Verifying the

    work page

  46. [52]

    Simulation-based optimal. J. Comput. Phys. , author =. 2013 , pages =. doi:10.1016/j.jcp.2012.08.013 , abstract =

    work page doi:10.1016/j.jcp.2012.08.013 2013

  47. [53]

    Monte. J. Mach. Learn. Res. , author =. 2020 , pages =

    work page 2020

  48. [55]

    Diffusion

    Dhariwal, Prafulla and Nichol, Alexander , year =. Diffusion. Advances in

    work page

  49. [56]

    Ajay, Anurag and Du, Yilun and Gupta, Abhi and Tenenbaum, Joshua and Jaakkola, Tommi and Agrawal, Pulkit , month = jul, year =. Is. doi:10.48550/arXiv.2211.15657 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2211.15657

  50. [57]

    The sample size required in importance sampling , volume =

    Chatterjee, Sourav and Diaconis, Persi , month = apr, year =. The sample size required in importance sampling , volume =. The Annals of Applied Probability , publisher =. doi:10.1214/17-AAP1326 , abstract =

    work page doi:10.1214/17-aap1326

  51. [58]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , author =

    Sequential. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , author =. 2006 , note =. doi:10.1111/j.1467-9868.2006.00553.x , abstract =

    work page doi:10.1111/j.1467-9868.2006.00553.x 2006

  52. [59]

    , month = nov, year =

    Kleinegesse, Steven and Gutmann, Michael U. , month = nov, year =. Bayesian. Proceedings of the 37th

    work page

  53. [60]

    , month = apr, year =

    Kleinegesse, Steven and Gutmann, Michael U. , month = apr, year =. Efficient. Proceedings of the

    work page

  54. [61]

    Representation Learning with Contrastive Predictive Coding

    Oord, Aaron van den and Li, Yazhe and Vinyals, Oriol , month = jan, year =. Representation. doi:10.48550/arXiv.1807.03748 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1807.03748

  55. [62]

    Robbins, Herbert and Monro, Sutton , month = sep, year =. A. The Annals of Mathematical Statistics , publisher =. doi:10.1214/aoms/1177729586 , abstract =

    work page doi:10.1214/aoms/1177729586

  56. [63]

    Journal of Mathematical Psychology , author =

    A tutorial on adaptive design optimization , volume =. Journal of Mathematical Psychology , author =. 2013 , keywords =. doi:10.1016/j.jmp.2013.05.005 , abstract =

    work page doi:10.1016/j.jmp.2013.05.005 2013

  57. [64]

    and Drovandi, Christopher C

    Ryan, Elizabeth G. and Drovandi, Christopher C. and McGree, James M. and Pettitt, Anthony N. , year =. A. International Statistical Review / Revue Internationale de Statistique , publisher =

    work page

  58. [65]

    Gal, Yarin and Islam, Riashat and Ghahramani, Zoubin , month = aug, year =. Deep. Proceedings of the 34th

    work page

  59. [66]

    Bioinformatics (Oxford, England) , author =

    A. Bioinformatics (Oxford, England) , author =. 2012 , keywords =. doi:10.1093/bioinformatics/bts092 , abstract =

    work page doi:10.1093/bioinformatics/bts092 2012

  60. [70]

    The Book of Statistical Proofs , author =

    Differential entropy of the multivariate normal distribution , url =. The Book of Statistical Proofs , author =

    work page

  61. [71]

    , editor =

    Bai, Zhidong and Silverstein, Jack W. , editor =. Sample. Spectral. 2010 , pages =. doi:10.1007/978-1-4419-0661-8_3 , abstract =

    work page doi:10.1007/978-1-4419-0661-8_3 2010

  62. [72]

    , year =

    Bai, Zhidong and Silverstein, Jack W. , year =. Spectral. doi:10.1007/978-1-4419-0661-8 , language =

    work page doi:10.1007/978-1-4419-0661-8

  63. [73]

    1967 , pages =

    Mathematics of the USSR-Sbornik , author =. 1967 , pages =. doi:10.1070/SM1967v001n04ABEH001994 , number =

    work page doi:10.1070/sm1967v001n04abeh001994 1967

  64. [79]

    Particle

    Phillips, Angus and Dau, Hai-Dang and Hutchinson, Michael John and Bortoli, Valentin De and Deligiannidis, George and Doucet, Arnaud , month = jul, year =. Particle. Proceedings of the 41st

    work page

  65. [80]

    IEEE Transactions on Information Theory , author =

    The convolution inequality for entropy powers , volume =. IEEE Transactions on Information Theory , author =. 1965 , note =. doi:10.1109/TIT.1965.1053768 , abstract =

    work page doi:10.1109/tit.1965.1053768 1965

  66. [81]

    Robbins, Herbert , month = jan, year =. An. Proceedings of the

    work page

  67. [82]

    Denoising

    Song, Jiaming and Meng, Chenlin and Ermon, Stefano , month = oct, year =. Denoising

    work page

  68. [83]

    , year =

    Beaumont, Mark A. , year =. Approximate. Annual Review of Ecology, Evolution, and Systematics , publisher =

    work page

  69. [84]

    Ecology Letters , author =

    Statistical inference for stochastic simulation models – theory and application , volume =. Ecology Letters , author =. 2011 , note =. doi:10.1111/j.1461-0248.2011.01640.x , abstract =

    work page doi:10.1111/j.1461-0248.2011.01640.x 2011

  70. [85]

    and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , month = oct, year =

    Song, Yang and Sohl-Dickstein, Jascha and Kingma, Diederik P. and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , month = oct, year =. Score-

    work page

  71. [86]

    Denoising diffusion probabilistic models , isbn =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , month = dec, year =. Denoising diffusion probabilistic models , isbn =. Proceedings of the 34th

    work page

  72. [87]

    and Maheswaranathan, Niru and Ganguli, Surya , month = jul, year =

    Sohl-Dickstein, Jascha and Weiss, Eric A. and Maheswaranathan, Niru and Ganguli, Surya , month = jul, year =. Deep unsupervised learning using nonequilibrium thermodynamics , abstract =. Proceedings of the 32nd

    work page

  73. [88]

    Generative

    Song, Yang and Ermon, Stefano , year =. Generative. Advances in

    work page

  74. [89]

    , year =

    Neal, Radford M. , year =. Handbook of

    work page

  75. [90]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) , author =

    Optimal scaling of discrete approximations to. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , author =. 1998 , note =. doi:10.1111/1467-9868.00123 , abstract =

    work page doi:10.1111/1467-9868.00123 1998

  76. [91]

    Coconuts and Islanders: A Statistics-First Guide to the Boltzmann Distribution

    Zhang, Brian , month = apr, year =. Coconuts and. doi:10.48550/arXiv.1904.04669 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1904.04669 1904

  77. [92]

    Auto-Encoding Variational Bayes

    Kingma, Diederik P. and Welling, Max , month = dec, year =. Auto-. doi:10.48550/arXiv.1312.6114 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1312.6114

  78. [93]

    Probabilistic

    Neal, Radford M , year =. Probabilistic

    work page

  79. [94]

    and Stern, Hal S

    Gelman, Andrew and Carlin, John B. and Stern, Hal S. and Rubin, Donald B. , month = jun, year =. Bayesian. doi:10.1201/9780429258411 , abstract =

    work page doi:10.1201/9780429258411

  80. [95]

    Rainforth, Tom and Kosiorek, Adam and Le, Tuan Anh and Maddison, Chris and Igl, Maximilian and Wood, Frank and Teh, Yee Whye , month = jul, year =. Tighter. Proceedings of the 35th

    work page

Showing first 80 references.