Pith. sign in

REVIEW 3 major objections 3 minor 79 references

RAMP: Recognition parametrisation by Amortised Message Passing

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read RAMP claims that latent structure can be defined implicitly by learning a flexible, nonlinear, amortised message-passing network on a tree, so that optimisation of a summed free energy recovers true latent beliefs even in expressive nonline

desk verdict RAMP is a promising extension of RPM to tree-structured latent graphs with learned message passing, but its main theorem is an existence result the authors concede doesn't apply to their exponential-family implementation. read the letter →

arxiv 2607.18883 v1 pith:2SLVT3ZH submitted 2026-07-21 cs.LG cs.AI

Lior Fox , Kai Biegun , James Heald , Samo Hromadka , Arielle Rosinski , Maneesh Sahani This is my paper · ORCID
classification cs.LGcs.AI MSC 68T0762F15
keywords recognition-parametrisedmodelsamortisedmessagepassingbeliefpropagationtree-structuredgraphicalunsupervisedrepresentationlearningfreeenergy/ELBOnonlinearlatentvariableexponentialfamily
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

RAMP is a method for unsupervised learning that lets the inference procedure define the model itself. Instead of specifying tractable generative conditionals, it learns a single shared amortised message-passing network on a tree-structured graphical model, where each message is a normalised belief about a latent variable. The paper claims that optimising the sum of nodewise recognition-parametrised free energies—one RPM for each latent node—forces the learned beliefs to capture the conditional-independence partitions the latent variables induce on the observations. The central theorem says that when the data are compatible with the tree and the message functions optimise this objective, every nodewise model attains its optimal value and each learned belief is a function of the true minimal sufficient posterior. A reader should care because this promises efficient likelihood-based representation learning in expressive nonlinear models, with demonstrations on hierarchical trees, nonlinear dynamical systems, and human pose from local image patches.

What carries the argument

The central mechanism is the amortised message-passing network defined by Eq. (6), together with the summed nodewise free energy of Eq. (10). A message from latent $z_j$ to $z_k$ transforms a belief about $z_j$ (formed from all other incident messages) into a belief about $z_k$; in exponential-family form this becomes $\eta_{j\to k} = g_{j\to k}(\sum_{i \in \partial j \setminus k} \eta_{i\to j})$, where $g_{j\to k}$ is a learned neural network and $\eta$ are natural parameters. Each latent node anchors its own RPM whose recognition factors are these messages; multiplying them recovers the local posterior. The summed objective is what couples all the nodewise models and encodes the tree structure: only a latent whose partition of observations is the o

What would settle it

Train RAMP on a known tree-structured generative model whose exact posterior beliefs are multimodal (for example, a switching nonlinear dynamical system) while using Gaussian messages. If the learned Gaussian messages fail to make the inferred posterior means reproducible from the true latent state, yet a non-Gaussian message family on the same objective recovers them, the claim that RAMP enables likelihood-based recovery of latent-variable distributions in expressive nonlinear models is falsified in that setting. A simpler direct test: in a linear-Gaussian tree with known exact beliefs, if co

Watch

Extended reading notes

Core claim

The paper's central claim is that latent variables need not be defined by a generative model at all: RAMP implicitly defines them by learning how beliefs should be passed along a tree. Formally, for each latent node $z_k$ it constructs a nodewise Recognition-Parametrised Model (RPM) whose recognition factors are amortised messages: $f^k_j(z_k|X^k_j) = G_{j\to k}(\ldots)$, with a shared learned functional $G$ for every edge. In the exponential-family implementation, messages are natural parameters and each edge is a neural network $g_{j\to k}$ mapping the sum of incoming natural parameters to an outgoing belief. Training maximises the average of the nodewise free energies $F = \frac{1}{K} \sum_k F_k$. Lemma 2 shows that

Load-bearing premise

The load-bearing premise is that there exists a single shared amortised message-passing network that exactly computes the true posterior beliefs and that optimising the summed free energy reaches it, whereas the practical exponential-family parameterisation cannot represent exact non-Gaussian beliefs, making Theorem 3 inapplicable to the implemented algorithm.

Editorial extensions

If this is right

  • If Theorem 3 holds, optimising the summed nodewise free energies simultaneously makes every node's RPM optimal, so a single training pass yields likelihood-based posterior beliefs across an entire tree-structured model.
  • Because the model likelihood is defined through recognition factors and normalised by empirical marginals, RAMP gives a proper probabilistic objective that can be optimised by gradient ascent on the free energy, without tractable generative conditionals or Monte-Carlo integration over latent paths.
  • In time-series settings with tied edge functions, RAMP becomes a bidirectional recurrent network whose hidden states are probabilistic beliefs; the pendulum experiment shows that this lets the model infer dynamical state components (such as angular velocity) that are not visible in any single observation.
  • The structural-misspecification experiment suggests that correctness of a latent's induced observation partition, not of the whole tree, is what determines whether that latent is recovered; a latent whose partition is intact keeps its accuracy even inside a misspecified tree.
  • The exponential-family implementation makes belief multiplication tractable but approximate; the paper argues that what is optimised is then the sum of nodewise KL divergences from the ideal model, giving a principled approximate objective.

Reading between the lines

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

  • The theoretical guarantee in Theorem 3 applies to an idealised regime where a shared amortised network exactly computing true beliefs exists and the optimiser reaches it; the exponential-family parameterisation used in experiments cannot represent exact non-Gaussian beliefs, so the experiments support the method's practical objective rather than the theorem's literal conclusion.
  • The dependency on the learned partition suggests a testable robustness claim: RAMP should be able to discover which latent tree is correct from data, since misspecified latents should show systematically lower nodewise free energies; this could be turned into a model-selection procedure over tree structures.
  • The recognition-parametrised view implies the message family is an inductive bias: any tractable exponential family (Gaussian, Beta, etc.) can be plugged in, and non-Gaussian families should help when true posteriors are bounded or multimodal; this is testable on the pendulum model with Beta messages already demonstrated.
  • If the amortised-message-passing logic carries over to factor graphs and eventually loopy graphs—which the paper leaves as future work—the same objective could provide practical approximate inference on general graphical models, where exact belief propagation is not guaranteed to converge.
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

3 major / 3 minor

Summary. The paper proposes RAMP (Recognition parametrisation by Amortised Message Passing), a method for unsupervised learning of latent variable models on tree-structured graphs. RAMP defines a set of nodewise recognition-parametrised models (RPMs), one per latent node, whose recognition factors are computed by a shared amortised message-passing network. The parameters are learned by optimising the sum of nodewise free-energy lower bounds (Eq. 10). The central theoretical claim, Theorem 3, states that if the data are compatible with the tree and the message-passing functions globally optimise this objective in the population limit, then each nodewise model attains its optimal value and the learned recognition factors are functions of the true posterior beliefs (minimal sufficient statistics). The implementation uses exponential-family beliefs, with message passing realised as neural network maps between natural parameters (Eq. 11). Experiments on hierarchical linear-Gaussian and nonlinear tree models, a nonlinear pendulum dynamical system, and human pose estimation demonstrate that RAMP can recover latent representations with high correlation to ground-truth latents, often outperforming baselines.

Significance. The RAMP framework is a conceptually interesting contribution: it extends the recognition-parametrised modelling paradigm to structured latent models by amortising the belief-propagation transformation itself, thereby avoiding explicit specification of generative conditionals. If the theoretical guarantees could be made to apply to the actual algorithm, the paper would provide a principled way to learn expressive nonlinear latent-variable models with a well-defined likelihood-based objective. The paper is clearly written, the experiments are diverse, and the appendices provide substantial detail. Strengths include the pseudocode for the free-energy computation (Listing 1), the explicit treatment of structural and distributional misspecification in Section 5.1, and the careful description of baselines. The main weakness, discussed below, is the gap between the idealised assumptions of Theorem 3 and the exponential-family implementation used in all experiments.

major comments (3)
  1. [§3.3, Theorem 3] The theorem's hypothesis requires that 'there exist functions {f0^k}, {fp^k} and {G_{j→k}} defining accurate beliefs' and that the optimiser finds them. This existence step is not established for the function classes actually used. Section 3.4 explicitly concedes that for the exponential-family parametrisation 'Theorem 3 no longer applies directly'. Since all experiments use this parametrisation, the formal guarantee covers an idealised regime distinct from the evaluated algorithm. The paper should either prove an approximate version (e.g., a misspecification bound that degrades gracefully with the exponential-family projection error) or clearly state that Theorem 3 is illustrative and not a guarantee for the implemented method.
  2. [§3.3/§3.4, global optimality] Theorem 3 assumes that the message-passing functions 'all optimise the objective of Eq. (10)'. No convergence result is provided that stochastic gradient descent (or any other optimiser) reaches this global optimum, even when it exists. The empirical evaluations measure correlation between inferred posterior means and true latents, which is not equivalent to verifying that the learned messages coincide with the true sufficient statistics. The paper should at least discuss the nonconvexity of the optimisation problem and the conditions under which the objective can be globally optimised, or weaken the claim accordingly.
  3. [§3.2/§3.3, the summed objective] Eq. (10) optimises the average of K nodewise free energies, each derived from a different RPM. These are lower bounds on different likelihood functions, not on a single joint model. Although the authors argue that the node-wise conditional independencies collectively identify the tree (Lemma 2), it is not shown that optimising the unweighted sum is consistent—for example, whether the same shared parameters can simultaneously make all F_k tight, or whether conflicting gradients can arise. The claim in Section 3.3 that 'RAMP instead optimises the summed objective' needs a more formal justification that the sum is a principled surrogate, not just a heuristic.
minor comments (3)
  1. [§3.4] The statement 'we expect F_k^j → f_0^k as N → ∞ in a well-specified model' is plausible but not proved. Since the interior variational bound (Appendix A.4) relies on this heuristic, a brief formal justification or a reference would be helpful.
  2. [§3.1, Eq. (6)] The notation G_{j→k}(·) is described as a functional mapping a belief on z_j to a belief on z_k, but the argument in Eq. (6) includes division by f_0^j, which is not explicitly defined as a functional argument. Clarify the domain and codomain of G_{j→k}.
  3. [§5.2, Table 1] The R2 values for SVAE on cosθ and sinθ are high but the R2 for ω is negative on test. The discussion mentions this, but it would be useful to state whether SVAE was limited to a 2D latent or also allowed higher-dimensional latents, as the text says 'When limited to the true two-dimensional latent space'. Clarify the settings for both baselines.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 3 is a consistency tautology: it assumes an exact amortised message-passing network exists and is found, then concludes the learned beliefs determine the true beliefs; Section 3.4 removes the theorem from the evaluated exponential-family algorithm.

  1. self definitional [Section 3.3, Theorem 3 and its proof (objective Eq. 10); representability concession in Section 3.4.]
    "the set of possible amortised message-passing networks includes one that computes the true beliefs ... That is, there exist functions {f k 0 (zk)}, {f k p (zk|xp)} and {Gj→k } defining accurate beliefs ... As each nodewise RPM is then an accurate model of the joint over {X k j }j∈∂k , the amortised belief f k j (zk|X k j ) is a sufficient statistic ... Claim (2) then follows as p(z∗ k |X k j ) is a minimal sufficient statistic by assumption, and so a function of any other sufficient statistic."

    The proof begins by assuming that an amortised network computing the true beliefs exists; the theorem conclusion (2) is that the true beliefs are functions of the learned f_j^k. If the learned message is the accurate belief, that function is the identity, so the conclusion restates the existence assumption. The sufficiency/minimal-sufficiency step only says any sufficient statistic determines the minimal sufficient statistic; it does not show that optimising Eq. (10) selects the accurate message set, nor that a joint optimum of the sum makes each F_k individually optimal. Section 3.4 concedes the implemented exponential-family parameterisation cannot represent exact non-Gaussian beliefs and that "Theorem 3 no longer applies directly", so the formal recovery claim is not established for the

full rationale

RAMP's central formal recovery claim is Theorem 3. The proof does not derive that optimising Eq. (10) yields exact beliefs; it first asserts that an amortised network computing the true beliefs exists (from compatibility), then treats the optimising network as that accurate model, so the conclusion that true beliefs are functions of the learned f_j^k is the assumption restated. This is a consistency tautology rather than a derivation of recovery from the objective. The implemented exponential-family message parameterisation cannot represent exact non-Gaussian posteriors; the paper itself states 'Theorem 3 no longer applies directly' (Section 3.4). The experimental sections (linear-Gaussian tree, pendulum, pose) are empirical and not circular—they do not fit to the quantities they later report—but they validate correlation with true latents, not the sufficiency claim in Theorem 3. Self-citations (Walker et al. 2023; Hromadka et al. 2025) supply the RPM free-energy and variational-bound machinery used in the construction; they are not the source of the recovery conclusion and do not by themselves raise the circularity score beyond the tautological theorem. Overall, one central 'prediction' reduces by construction; score 6.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

RAMP does not introduce new physical or ontological entities; the amortised message-passing network is a functional component, and latent variables are standard model variables. The load-bearing assumptions are the identifiability condition (Assumption 1), the existence/attainability condition in Theorem 3, and the exponential-family projection that places the implementation outside the theorem's scope.

free parameters (3)
  • Latent dimensionality per node = 2 (pendulum), 6 (pose)
    Chosen by hand; central to representational capacity, with no model-selection criterion reported.
  • Message-passing network widths = 32–128 hidden units depending on experiment
    Selected by grid search in the pendulum experiment (Appendix B.4) or fixed for other experiments; affects all reported results.
  • Learning rate, batch size, training iterations = 0.001/0.0005/0.0001; N=500 or 200; 500–3000 iterations
    Chosen by hand per experiment; no comprehensive sensitivity analysis is given, only a free-energy ranking for pendulum sweeps.
assumptions (6)
  • domain assumption Observations can be taken as leaves of the tree without loss of generality
    Footnote 1 in Section 2.2: an internal observed node would split the tree into independent subtrees within which inference and learning proceed independently. This restriction is used throughout and shapes the nodewise partition argument.
  • domain assumption Assumption 1: every pair of latent variables induces a distinct partition of the observed leaves
    Section 2.2.1. Needed for Lemma 2 and to identify latents from conditional-independence partitions. If two latents induce the same partition, they are not separately identifiable and the tree structure is not recoverable.
  • domain assumption The empirical measure p_emp(x_p) is atomic: (1/N)Σ δ(x_p − x_p^(n))
    Inherited from RPM (Walker et al., 2023, Eq. 1). Makes F_k^j computable as an average over data (Eq. 8) and is used in the interior variational bound in Appendix A.4.
  • ad hoc to paper There exists a single shared amortised message-passing network that exactly computes the true beliefs and is attained by optimizing Eq. (10)
    Theorem 3 proof, Section 3.3: “there exist functions {f_0^k}, {f_p^k} and {G_{j→k}} defining accurate beliefs” and the optimiser reaches that point. This existence/attainability assumption is not proved.
  • domain assumption Exponential-family message parametrisation with the interior variational bound approximating F_j^k by f_0^k
    Section 3.4 and Appendix A.4. Restricts all messages at a node to one exponential family and uses f_0 as the variational approximating function. The paper states Theorem 3 no longer applies directly under this approximation.
  • domain assumption The data distribution is compatible with a tree-structured latent model satisfying Assumption 1
    Theorem 3 premise. For real data (pendulum, pose), this is a structural assumption about the data-generating process and is not verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAMP: Recognition parametrisation by Amortised Message Passing." pith.science (2026). https://pith.science/paper/2SLVT3ZH

@misc{pith2026260718883,
  author       = {Pith},
  title        = {Pith review of: RAMP: Recognition parametrisation by Amortised Message Passing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SLVT3ZH}},
  note         = {Machine review of arXiv:2607.18883}
}
read the original abstract

A central aim of unsupervised learning is to uncover latent factors that explain dependencies among observations. Probabilistic models typically achieve this by introducing multiple latent variables linked through a graph of conditional relationships, with distributional parameters and their dependence learnt from data. Learning relies either on distributional choices that allow tractable belief propagation, or on approximations that scale poorly with model size and complexity. We build on the recently developed recognition-parametrised modelling paradigm to propose an alternative approach: RAMP, a method that implicitly defines latent structure by learning a flexible, nonlinear, amortised message-passing framework. We show that RAMP enables efficient likelihood-based recovery of latent-variable distributions within expressive nonlinear models acting on complex high-dimensional data.

Figures

Figures reproduced from arXiv: 2607.18883 by the authors.

Figure 1
Figure 1. RAMP works by learning a system of amortised message-passing transformations (directed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) Hierarchical structure with 3 levels of latent variables (color coded) arranged on a binary tree and observations on the leaves. (b) Under a linear-Gaussian data generating process, Pearson correlations between the posterior means inferred by RAMP and the ground truth latents (solid lines) reach the optimal bound achieved by exact inference with the true generative model (dashed lines). (c) Data as in (b), but R… view at source ↗
Figure 3
Figure 3. Inferred posterior means of learnt 2D latents for the pendulum dataset. Left: points [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Pose reconstruction. (a) Images from LSP (left) were used to create simplified stick-figure [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 3 canonical work pages

  1. [1]

    JMLR , volume =

    An Introduction to Variable and Feature Selection , author =. JMLR , volume =

  2. [2]

    Guyon and C

    I. Guyon and C. Aliferis and A. Elisseeff , title =

  3. [3]

    Bidirectional recurrent neural network language models for automatic speech recognition

    Ebru Arisoy, Abhinav Sethy, Bhuvana Ramabhadran, and Stanley Chen. Bidirectional recurrent neural network language models for automatic speech recognition. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5421--5425. IEEE, 2015

    2015

  4. [5]

    Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL : Automatic estimation of 3 D human pose and shape from a single image. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision -- ECCV 2016, pages 561--578, Cham, 2016. Springer International Publishing. ISBN 978-3...

    2016

  5. [6]

    Approximate learning of dynamic models

    Xavier Boyen and Daphne Koller. Approximate learning of dynamic models. In M. S. Kearns, S. A. Solla, and D. A. Cohn, editors, Advances in Neural Information Processing Systems, volume 11. MIT Press, 1999

    1999

  6. [8]

    A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum Likelihood from Incomplete Data via the EM Algorithm . Journal of the Royal Statistical Society. Series B (Methodological), 39 0 (1): 0 1--38, 1977. ISSN 0035-9246

    1977

  7. [9]

    Contrastive learning as goal-conditioned reinforcement learning

    Benjamin Eysenbach, Tianjun Zhang, Sergey Levine, and Russ R Salakhutdinov. Contrastive learning as goal-conditioned reinforcement learning. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 35603--35620. Curran Associates, Inc., 2022. URL https://proceedings....

    2022

  8. [10]

    A synopsis of linguistic theory, 1930-1955

    John Firth. A synopsis of linguistic theory, 1930-1955. Studies in linguistic analysis, pages 10--32, 1957

    1930

Show all 79 references
  1. [12]

    Maximum likelihood learning of latent dynamics without reconstruction, 2025

    Samo Hromadka, Kai Biegun, Lior Fox, James Heald, and Maneesh Sahani. Maximum likelihood learning of latent dynamics without reconstruction, 2025. URL https://arxiv.org/abs/2505.23569

    2025 arXiv

  2. [13]

    Particle belief propagation

    Alexander Ihler and David McAllester. Particle belief propagation. In Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS), volume 5, pages 256--263. PMLR, 2009

    2009

  3. [14]

    Composing graphical models with neural networks for structured representations and fast inference

    Matthew J Johnson, David K Duvenaud, Alex Wiltschko, Ryan P Adams, and Sandeep R Datta. Composing graphical models with neural networks for structured representations and fast inference. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural ...

    2016

  4. [15]

    Learning effective human pose estimation from inaccurate annotation

    Sam Johnson and Mark Everingham. Learning effective human pose estimation from inaccurate annotation. In Proceedings of Computer Vision and Pattern Recognition (CVPR) 2011, 2011

    2011

  5. [16]

    Probabilistic Graphical Models: Principles and Techniques

    Daphne Koller and Nir Friedman. Probabilistic Graphical Models: Principles and Techniques. MIT Press, Cambridge, MA, 2009. ISBN 9780262013192

    2009

  6. [17]

    Expectation particle belief propagation

    Thibaut Lienart, Yee Whye Teh, and Arnaud Doucet. Expectation particle belief propagation. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. URL https://proceedi...

    2015

  7. [18]

    Deep learning for universal linear embeddings of nonlinear dynamics

    Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Deep learning for universal linear embeddings of nonlinear dynamics. Nature communications, 9 0 (1): 0 4950, 2018

    2018

  8. [19]

    Thomas P. Minka. Expectation propagation for approximate bayesian inference. In UAI, pages 362--369, 2001

    2001

  9. [20]

    3d human pose estimation from a single image via distance matrix regression

    Francesc Moreno-Noguer. 3d human pose estimation from a single image via distance matrix regression. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1561--1570, 2017. doi:10.1109/CVPR.2017.170

    2017 doi

  10. [21]

    Murphy, Yair Weiss, and Michael I

    Kevin P. Murphy, Yair Weiss, and Michael I. Jordan. Loopy belief propagation for approximate inference: an empirical study. In Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence, UAI'99, page 467–475, San Francisco, CA, USA, 1999. Morgan Kaufmann...

    1999

  11. [22]

    Temporally-consistent koopman autoencoders for forecasting dynamical systems

    Indranil Nayak, Ananda Chakrabarti, Mrinal Kumar, Fernando L Teixeira, and Debdipta Goswami. Temporally-consistent koopman autoencoders for forecasting dynamical systems. Scientific Reports, 15 0 (1): 0 22127, 2025

    2025

  12. [23]

    Neal and Geoffrey E

    Radford M. Neal and Geoffrey E. Hinton. A view of the EM algorithm that justifies incremental, sparse, and other variants. In Michael I. Jordan, editor, Learning in Graphical Models, pages 355--370. Kluwer Academic Press, 1998

    1998

  13. [24]

    Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference

    Judea Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1988. ISBN 1558604790

    1988

  14. [25]

    Black box variational inference

    Rajesh Ranganath, Sean Gerrish, and David Blei. Black box variational inference. In Artificial intelligence and statistics, pages 814--822. PMLR, 2014

    2014

  15. [26]

    A gentle introduction to graph neural networks

    Benjamin Sanchez-Lengeling, Emily Reif, Adam Pearce, and Alexander B Wiltschko. A gentle introduction to graph neural networks. Distill, 6 0 (9): 0 e33, 2021

    2021

  16. [27]

    Schuster and K.K

    M. Schuster and K.K. Paliwal. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45 0 (11): 0 2673--2681, 1997. doi:10.1109/78.650093

    1997 doi

  17. [29]

    Sudderth, A.T

    E.B. Sudderth, A.T. Ihler, W.T. Freeman, and A.S. Willsky. Nonparametric belief propagation. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., volume 1, pages I--I, 2003. doi:10.1109/CVPR.2003.1211409

    2003 arXiv

  18. [30]

    Representation learning with contrastive predictive coding

    A \"a ron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. ArXiv, abs/1807.03748, 2018

    2018 arXiv

  19. [31]

    Wainwright and Michael I

    Martin J. Wainwright and Michael I. Jordan. Graphical Models , Exponential Families , and Variational Inference . Foundations and Trends in Machine Learning, 1 0 (1--2): 0 1--305, 2008. ISSN 1935-8237, 1935-8245

    2008

  20. [32]

    Tree-based reparameterization for approximate inference on loopy graphs

    Martin J Wainwright, Tommi Jaakkola, and Alan Willsky. Tree-based reparameterization for approximate inference on loopy graphs. In T. Dietterich, S. Becker, and Z. Ghahramani, editors, Advances in Neural Information Processing Systems, volume 14. MIT Press, 2001. URL https://p...

    2001

  21. [33]

    Walker, Hugo Soulat, Changmin Yu, and Maneesh Sahani

    William I. Walker, Hugo Soulat, Changmin Yu, and Maneesh Sahani. Unsupervised representation learning with recognition-parametrised probabilistic models. In Francisco Ruiz, Jennifer Dy, and Jan-Willem van de Meent, editors, Proceedings of The 26th International Conference on A...

    2023

  22. [35]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning

    Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8 0 (3): 0 229--256, 1992

    1992

  23. [36]

    John Winn and Christopher M. Bishop. Variational message passing. Journal of Machine Learning Research, 6 0 (23): 0 661--694, 2005. URL http://jmlr.org/papers/v6/winn05a.html

    2005

  24. [37]

    Generalized belief propagation

    Jonathan S Yedidia, William Freeman, and Yair Weiss. Generalized belief propagation. In T. Leen, T. Dietterich, and V. Tresp, editors, Advances in Neural Information Processing Systems, volume 13. MIT Press, 2000. URL https://proceedings.neurips.cc/paper_files/paper/2000/file/...

    2000

  25. [38]

    Yu, Krishna V

    Byron M. Yu, Krishna V. Shenoy, and Maneesh Sahani. Expectation propagation for inference in non-linear dynamical models with P oisson observations. In Proceedings of the Nonlinear Statistical Signal Processing Workshop. IEEE, 2006

    2006

  26. [39]

    Revisiting structured variational autoencoders

    Yixiu Zhao and Scott Linderman. Revisiting structured variational autoencoders. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning, volume 202 ...

    2023

  27. [40]

    2009 , publisher =

    Probabilistic Graphical Models: Principles and Techniques , author =. 2009 , publisher =

    2009

  28. [41]

    1988 , isbn =

    Pearl, Judea , title =. 1988 , isbn =

    1988

  29. [42]

    Proceedings of the Second AAAI Conference on Artificial Intelligence , pages =

    Pearl, Judea , title =. Proceedings of the Second AAAI Conference on Artificial Intelligence , pages =. 1982 , publisher =

    1982

  30. [43]

    Tree-based reparameterization for approximate inference on loopy graphs , url =

    Wainwright, Martin J and Jaakkola, Tommi and Willsky, Alan , booktitle =. Tree-based reparameterization for approximate inference on loopy graphs , url =

  31. [44]

    Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS) , volume =

    Particle Belief Propagation , author =. Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS) , volume =. 2009 , publisher =

    2009

  32. [45]

    Foundations and Trends in Machine Learning , author =

    Graphical. Foundations and Trends in Machine Learning , author =. 2008 , pages =

    2008

  33. [46]

    , author=

    Variational message passing. , author=

  34. [47]

    Journal of the Royal Statistical Society

    Maximum. Journal of the Royal Statistical Society. Series B (Methodological) , author =

  35. [48]

    Neal and Geoffrey E

    Radford M. Neal and Geoffrey E. Hinton , title =

  36. [49]

    Proceedings of The 26th International Conference on Artificial Intelligence and Statistics , pages =

    Unsupervised representation learning with recognition-parametrised probabilistic models , author =. Proceedings of The 26th International Conference on Artificial Intelligence and Statistics , pages =. 2023 , editor =

    2023

  37. [50]

    2025 , eprint=

    Maximum Likelihood Learning of Latent Dynamics Without Reconstruction , author=. 2025 , eprint=

    2025

  38. [51]

    Contrastive Learning as Goal-Conditioned Reinforcement Learning , url =

    Eysenbach, Benjamin and Zhang, Tianjun and Levine, Sergey and Salakhutdinov, Russ R , booktitle =. Contrastive Learning as Goal-Conditioned Reinforcement Learning , url =

  39. [52]

    ArXiv , year=

    Representation Learning with Contrastive Predictive Coding , author=. ArXiv , year=

  40. [53]

    2011 , booktitle =

    Learning Effective Human Pose Estimation from Inaccurate Annotation , author =. 2011 , booktitle =

    2011

  41. [54]

    Composing graphical models with neural networks for structured representations and fast inference , url =

    Johnson, Matthew J and Duvenaud, David K and Wiltschko, Alex and Adams, Ryan P and Datta, Sandeep R , booktitle =. Composing graphical models with neural networks for structured representations and fast inference , url =

  42. [55]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Revisiting Structured Variational Autoencoders , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , editor =

    2023

  43. [56]

    ICLR , author =

    Auto-Encoding Variational. ICLR , author =

  44. [57]

    Stochastic

    Rezende, Danilo Jimenez and Mohamed, Shakir and Wierstra, Daan , month = jan, year =. Stochastic

  45. [58]

    and Neal, Radford M

    Dayan, Peter and Hinton, Geoffrey E. and Neal, Radford M. and Zemel, Richard S. , year = 1995, pages =. The

    1995

  46. [59]

    2015 , eprint=

    Deep Kalman Filters , author=. 2015 , eprint=

    2015

  47. [60]

    Two problems with variational expectation maximisation for time series models , booktitle=

    Turner, Richard Eric and Sahani, Maneesh , year=. Two problems with variational expectation maximisation for time series models , booktitle=

  48. [61]

    Artificial intelligence and statistics , pages=

    Black box variational inference , author=. Artificial intelligence and statistics , pages=. 2014 , organization=

    2014

  49. [62]

    UAI , pages=

    Expectation Propagation for approximate Bayesian inference , author=. UAI , pages=

  50. [63]

    Advances in Neural Information Processing Systems , year = 1999, editor =

    Xavier Boyen and Daphne Koller , title =. Advances in Neural Information Processing Systems , year = 1999, editor =

    1999

  51. [64]

    Yu and Krishna V

    Byron M. Yu and Krishna V. Shenoy and Maneesh Sahani , title =. Proceedings of the Nonlinear Statistical Signal Processing Workshop , publisher =

  52. [65]

    and Weiss, Yair and Jordan, Michael I

    Murphy, Kevin P. and Weiss, Yair and Jordan, Michael I. , title =. Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence , pages =. 1999 , isbn =

    1999

  53. [66]

    Generalized Belief Propagation , url =

    Yedidia, Jonathan S and Freeman, William and Weiss, Yair , booktitle =. Generalized Belief Propagation , url =

  54. [67]

    Scientific Reports , volume=

    Temporally-consistent koopman autoencoders for forecasting dynamical systems , author=. Scientific Reports , volume=. 2025 , publisher=

    2025

  55. [68]

    Nature communications , volume=

    Deep learning for universal linear embeddings of nonlinear dynamics , author=. Nature communications , volume=. 2018 , publisher=

    2018

  56. [69]

    arXiv preprint arXiv:2102.12086 , year=

    Modern Koopman theory for dynamical systems , author=. arXiv preprint arXiv:2102.12086 , year=

    arXiv

  57. [70]

    and Paliwal, K.K

    Schuster, M. and Paliwal, K.K. , journal=. Bidirectional recurrent neural networks , year=

  58. [71]

    2005 , note =

    Framewise phoneme classification with bidirectional LSTM and other neural network architectures , journal =. 2005 , note =. doi:https://doi.org/10.1016/j.neunet.2005.06.042 , url =

    2005 doi

  59. [72]

    Bioinformatics , volume =

    Baldi, Pierre and Brunak, Søren and Frasconi, Paolo and Soda, Giovanni and Pollastri, Gianluca , title =. Bioinformatics , volume =. 1999 , month =. doi:10.1093/bioinformatics/15.11.937 , url =

    1999 doi

  60. [73]

    2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Bidirectional recurrent neural network language models for automatic speech recognition , author=. 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2015 , organization=

    2015

  61. [74]

    Distill , volume=

    A gentle introduction to graph neural networks , author=. Distill , volume=

  62. [75]

    arXiv preprint arXiv:2407.00494 , year=

    Graph neural networks gone hogwild , author=. arXiv preprint arXiv:2407.00494 , year=

    arXiv

  63. [76]

    Studies in linguistic analysis , pages=

    A synopsis of linguistic theory, 1930-1955 , author=. Studies in linguistic analysis , pages=

    1930

  64. [77]

    Machine learning , volume=

    Simple statistical gradient-following algorithms for connectionist reinforcement learning , author=. Machine learning , volume=. 1992 , publisher=

    1992

  65. [78]

    Learning in Graphical Models , booktitle =

  66. [79]

    Yuille and Wen Gao , title =

    Chunyu Wang and Yizhou Wang and Zhouchen Lin and Alan L. Yuille and Wen Gao , title =. 2014. 2014 , url =. doi:10.1109/CVPR.2014.303 , timestamp =

    2014 doi

  67. [80]

    Keep It SMPL : Automatic Estimation of 3 D Human Pose and Shape from a Single Image

    Bogo, Federica and Kanazawa, Angjoo and Lassner, Christoph and Gehler, Peter and Romero, Javier and Black, Michael J. Keep It SMPL : Automatic Estimation of 3 D Human Pose and Shape from a Single Image. Computer Vision -- ECCV 2016. 2016

    2016

  68. [81]

    3D Human Pose Estimation from a Single Image via Distance Matrix Regression , year=

    Moreno-Noguer, Francesc , booktitle=. 3D Human Pose Estimation from a Single Image via Distance Matrix Regression , year=

  69. [82]

    Expectation Particle Belief Propagation , url =

    Lienart, Thibaut and Teh, Yee Whye and Doucet, Arnaud , booktitle =. Expectation Particle Belief Propagation , url =

  70. [83]

    and Ihler, A.T

    Sudderth, E.B. and Ihler, A.T. and Freeman, W.T. and Willsky, A.S. , booktitle=. Nonparametric belief propagation , year=

  71. [84]

    Bishop , title =

    John Winn and Christopher M. Bishop , title =. Journal of Machine Learning Research , year =

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.