Pith. sign in

REVIEW 1 major objections 5 minor 114 references

Deep Gaussian processes on DAGs keep input distinctions alive with depth and recover explaining-away under partial observations.

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

2026-07-13 01:29 UTC pith:44UZIWGX

load-bearing objection Solid first generalisation of DGPs to DAGs with real non-collapse theory and a usable structured VI scheme; known-graph assumption is the main scope limit, not a crack in the claims. the 1 major comments →

arxiv 2607.09645 v1 pith:44UZIWGX submitted 2026-07-10 stat.ML cs.LGmath.STstat.COstat.MEstat.TH

Deep Gaussian Processes on Directed Acyclic Graphs

Federico L. Perlino , Oliver Hamelijnck , Adam M. Johansen , Theodoros Damoulas This is my paper
classification stat.ML cs.LGmath.STstat.COstat.MEstat.TH MSC 62M3062F1568T05
keywords deep Gaussian processesdirected acyclic graphsprior collapsestructured variational inferenceexplaining-awaymulti-fidelity emulationcompositional uncertainty
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.

Many scientific systems are compositions of latent functions wired along a known directed acyclic graph: causal mechanisms, multi-fidelity simulators, gene-regulatory networks. Data arrive only at some nodes, often noisy and unevenly sampled. This paper places independent Gaussian-process priors on every non-root node and obtains Deep Gaussian Processes over DAGs. It proves almost-sure lower bounds on how often two distinct inputs remain distinguishable as one moves deeper in the graph, shows how topology (in- and out-degree) and intermediate observations act as refresh mechanisms, and proves the long-standing conjecture that input connections prevent prior collapse on chains. A structured variational family built on the moralised ancestral graph keeps compositional uncertainty and the classical explaining-away dependence of colliders. On a protein-signalling network and a multi-fidelity heavy-ion collision task the method matches or beats specialised baselines while recovering how lower-fidelity branches contribute to the high-fidelity prediction.

Core claim

A Deep Gaussian Process defined directly on a DAG, with node-wise fusion kernels and observations at arbitrary nodes, does not suffer prior collapse whenever sufficiently many v-star-separating nodes appear across progressive antichains; the same architecture admits a structured variational approximation that retains posterior dependence among co-parents and yields state-of-the-art multi-fidelity emulation while recovering low-fidelity contributions.

What carries the argument

v-star-separating nodes (fresh GP modules whose two-case contrast covariance is bounded below almost surely) together with progressive antichain decompositions; they convert local non-degeneracy into almost-sure lower bounds on the asymptotic frequency of non-collapsed depths (Theorem 1) and turn intermediate observations into stochastic skip connections (Theorem 2).

Load-bearing premise

The directed acyclic graph itself is known and correctly specified; every non-collapse guarantee and the moralised precision construction stand or fall with that fixed topology.

What would settle it

On a chain or layered DAG whose kernels satisfy the paper’s conditions, draw many independent prior realisations of two distinct inputs and check whether the empirical frequency of depths with maximum contrast above a fixed epsilon falls below the claimed almost-sure lower bound 1-(1-p_epsilon)^s.

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

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

1 major / 5 minor

Summary. The paper introduces Deep Gaussian Processes on Directed Acyclic Graphs (DAG-DGPs): independent GP priors on nodewise maps with fusion kernels for multi-parent nodes, and heterogeneous noisy observations at arbitrary nodes. It proves almost-sure lower bounds on the asymptotic frequency of depths at which two-case input distinctions are preserved under progressive antichain decompositions when antichains contain sufficiently many v*-separating nodes (Theorem 1), characterises indegree/outdegree effects under radial fusion, and shows that intermediate observations act as stochastic skip connections (Theorem 2). It also proves non-collapse for the input-connected chain of Dunlop et al. (2018). A structured variational family (DAG-SVI) is built from the moralised ancestral graph and chordal completion so that explaining-away and compositional uncertainty are retained; mean-field DAG-VI is recovered as a special case. Theory is checked by Monte Carlo, and the method is demonstrated on a latent collider, the Sachs protein network, and multi-fidelity heavy-ion emulation, where it improves on published baselines and recovers low-fidelity contributions via Shapley values.

Significance. If the results hold, the paper supplies a single probabilistic framework that unifies chain DGPs, multi-fidelity DGPs, and graphical multi-fidelity emulators, with the first non-collapse theory for DAG compositions and a proof of the Dunlop et al. input-connection observation. The structured variational construction is a genuine extension of existing chain DGP approximations and is shown to capture explaining-away, which mean-field cannot. Empirical gains on the heavy-ion task (Table 2) and recovery of low-fidelity contributions are practically useful for scientific multi-fidelity modelling. Strengths include detailed appendix proofs (progressive antichains, conditional non-degeneracy, Azuma–Hoeffding frequency bounds, route retention), explicit kernel classes that certify separation, and multi-seed/fold reporting with competitive baselines.

major comments (1)
  1. No load-bearing technical error was found in the central theorems or the variational construction under the paper’s stated premises (known DAG, stated kernel classes, progressive antichains). The known-DAG assumption is a genuine scope limitation for causal and structure-discovery settings, but it is already flagged in Sec. 8.2 and does not invalidate the non-collapse bounds or the reported empirical gains under that assumption.
minor comments (5)
  1. Sec. 8.2 and the Sachs experiment: the 2-D latent confounder layer is ad hoc; a short sensitivity check (e.g., dimension 1 vs 2, or no confounder) would strengthen the claim that DAG-SVI’s extrapolation gains are not driven by that modelling choice alone.
  2. Fig. 2 and App. E.5: the branching-tree scaling is clear, but a one-sentence statement of when the sparse backend’s RH term remains favourable for denser moral graphs would help practitioners choose between dense and sparse DAG-SVI.
  3. Notation: the two filtrations F_ℓ (module-generated) and H_ℓ (state-generated) are carefully defined in the appendix; a brief cross-reference in Sec. 4 would reduce the chance of confusion for readers who stay in the main text.
  4. Table 1 / App. G.4: reporting PICP and NLPD for the Sachs interpolation task (as done for extrapolation in Table 3) would make the interpolation–extrapolation comparison more complete.
  5. Typographical: arXiv line breaks occasionally split math (e.g., product fusion displays); ensure final camera-ready equations are unbroken.

Circularity Check

0 steps flagged

No significant circularity: non-collapse bounds and variational constructions are derived from the prior and graph structure without reducing to fitted targets or self-justifying definitions.

full rationale

The central theoretical claims (Theorem 1 on almost-sure lower bounds for contrast frequency under repeated v*-separating nodes; the recovery of Dunlop et al. input-connection non-collapse as Corollary 2; intermediate-observation refresh as Theorem 2) are obtained from the DAG-DGP prior recursion, progressive antichain decompositions, conditional Gaussian non-degeneracy of separating nodes, and standard martingale/Azuma–Hoeffding/Borel–Cantelli arguments. Separation is certified by explicit kernel classes (additive/ANOVA root retention, multi-fidelity discrepancy kernels) whose contrast covariances are computed directly from the kernel definition, not fitted to the target frequency. The structured variational family (DAG-SVI) is constructed from the moralised ancestral graph and chordal completion, recovering known chain and mean-field special cases by restriction; the ELBO and ancestral sampling follow from the usual variational lower bound without circular dependence on the predictive metrics. Empirical results (collider explaining-away visualisation, Sachs interpolation/extrapolation, heavy-ion RMSE/CRPS/Shapley) evaluate held-out predictive performance and post-hoc attributions under a known DAG; they do not feed back into the statements of the theorems. Self-citations are limited to standard DGP/multi-fidelity baselines and do not supply load-bearing uniqueness or ansatzes for the main derivations. The known-DAG modelling assumption is an explicit scope limitation, not an internal circular reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central claims rest on standard GP finite-dimensional distributions, a known DAG, nodewise fusion kernels that produce v*-separation, and a chordal completion of the moralised ancestral graph. Hyperparameters and inducing locations are free parameters optimised by ELBO maximisation; the DAG itself and the choice of fusion rule are domain assumptions. No new physical entities are postulated.

free parameters (4)
  • kernel length-scales and variances (per node)
    ARD RBF (and multi-fidelity) hyperparameters are optimised by maximising the ELBO; they control both the prior contrast variance v* and predictive performance.
  • number and locations of inducing points M_w
    Chosen by the user (e.g. 20 per node on trees, 200 on heavy-ion L1/L2); locations are either fixed after initialisation or jointly optimised.
  • observation-noise variances σ_u²
    Likelihood noise parameters appear in the Gaussian refresh formula (Eq. 9) and are learned; they directly scale the source strengths q_j.
  • variational mean m and precision Λ of q_H(U)
    Free parameters of the structured Gaussian posterior; optimised by stochastic gradients of the ELBO.
axioms (5)
  • domain assumption Nodewise GP modules are mutually independent a priori and the joint prior factorises according to the DAG recursion (Eq. 2).
    Standard hierarchical GP construction; used throughout the prior and the ELBO derivation.
  • domain assumption The DAG G=(V,E) is known, acyclic and correctly specified.
    All theory and the moralised precision construction condition on a fixed graph (Sections 2 and 8.2).
  • ad hoc to paper A node is v*-separating if its two-point contrast covariance is bounded below by v* I almost surely (Definition in App. A.3).
    Technical condition introduced to obtain uniform lower bounds; verified for root-retaining and multi-fidelity kernels.
  • standard math Chordal completion of the moralised ancestral graph admits a clique-separator Gaussian representation (Lauritzen 1996).
    Used to define the sparse precision family q_H (Section 3).
  • standard math Progressive antichain decompositions exist for every finite DAG (Prop. 1).
    Order-theoretic fact used to define depth for non-chain graphs.
invented entities (3)
  • DAG-DGP prior and fusion kernels Φ_w independent evidence
    purpose: Place independent GP priors on every non-root node and combine multi-parent inputs via additive/product/ANOVA or domain-specific rules.
    Core modelling object; recovers chain DGPs and multi-fidelity models as special cases.
  • v*-separating node no independent evidence
    purpose: Abstract the mechanisms (root retention, perfect intermediate observation) that inject uniform contrast variance and prevent prior collapse.
    Technical device for Theorem 1; verified for concrete kernel classes in App. B.5.
  • DAG-SVI structured variational family q_H independent evidence
    purpose: Gaussian inducing posterior whose precision is supported on a chordal completion of the moralised ancestral graph, retaining explaining-away.
    Inference contribution; specialises to Ustyuzhaninov et al. block-tridiagonal family on chains.

pith-pipeline@v1.1.0-grok45 · 69190 in / 3404 out tokens · 39626 ms · 2026-07-13T01:29:08.478351+00:00 · methodology

0 comments
read the original abstract

Many real-world processes can be represented as compositions of functions along a directed acyclic graph (DAG). In causal modelling, these correspond to the underlying mechanisms; in engineering, to multiple fidelity levels; and in gene-regulatory networks, to transcription factors. These functions are partially observed across the DAG, with noisy and heterogeneously sampled measurements, posing significant challenges for reconstruction, uncertainty propagation, and inference. To tackle these challenges, we place priors over functions and naturally arrive at Deep Gaussian Processes over DAGs. We theoretically study their prior-collapse behaviour, and the effect of graph topology and intermediate observations on the preservation of information. We obtain almost-sure lower bounds on the asymptotic frequency of depths at which the distinction between inputs is preserved, identify broad kernel classes for which these hold, and prove an observation by \cite{dunlop2018} on the role of input connections. We offer a structured variational approximation that retains graph dependencies, preserves compositional uncertainty, and captures the explaining-away behaviour of colliders. Finally, we empirically validate our theoretical results and our methodology, and model a latent-collider DAG, a protein signalling network, and a multi-fidelity heavy-ion collision emulation task, attaining state-of-the-art performance while recovering low-fidelity contributions and yielding interpretability of the simulator hierarchy.

Figures

Figures reproduced from arXiv: 2607.09645 by Adam M. Johansen, Federico L. Perlino, Oliver Hamelijnck, Theodoros Damoulas.

Figure 1
Figure 1. Figure 1: DAGs discussed in Secs. 2, 3, and 4 (top row) and the corresponding block sparsity patterns of the structured precision matrix Λ (bottom row): chain DGP, disjoint routes (Thm. 4), V-structure with branching, and a three-layered DAG. Circles denote latent nodes and squares nodes with partial observations. With observations placed at the terminal nodes, each DAG coincides with its moralised ancestral graph, … view at source ↗
Figure 2
Figure 2. Figure 2: Wall-clock time per ELBO evalua￾tion on increasingly deep branching trees. Scaling up to larger DAGs. The expectation in Eq. (7) is estimated by marginal ancestral sampling (see Prop. (9)). Since qH is in canonical form, the sampler requires marginal and con￾ditional moments, equivalently selected applications of Λ −1 . Let J = |U|, M = maxw dim(Uw), and K = SB for S Monte Carlo samples and minibatch size … view at source ↗
Figure 3
Figure 3. Figure 3: Empirical validation of the main theoretical results in Sec. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Latent-collider experiment. Left: ground-truth w3, input marked. Centre: parent posterior under DAG-VI and DAG-SVI. Right: w3 and per-method errors, with root mean square error (RMSE) and negative log likelihood (NLL) vs. observations. For Gaussian observations Y (i) u = F (i) u + ξ (i) u , ξ (i) u ∼ N (0, σ2 u ), set Γ = Γu(a, b) and DY = Y (a) u − Y (b) u . If Γ > 0, then conditionally on Hℓ0 , F (a) u −… view at source ↗
Figure 5
Figure 5. Figure 5: DAG for the Sachs protein sig￾nalling network. Diamond nodes denote ex￾ogenous/root inputs. Blue nodes correspond to observed variables on which deep GP priors are placed. The dotted node denotes a latent confounder, modelled using a 2D latent vari￾able layer (Salimbeni et al., 2019). The first latent dimension acts as input to PKA, and the second to Raf. We evaluate the proposed DAG-DGP framework on the r… view at source ↗
Figure 6
Figure 6. Figure 6: DAG-DGP for the heavy-ion emulation task. We evaluate DAG-DGPs on the heavy-ion collision real dataset of Ji et al. (2024), a graphical multi-fidelity emulation problem with a shared nine-dimensional input and a scalar pion-yield-ratio output. The elicited simulator graph has two lower-fidelity nodes, L1 and L2, feeding the high-fidelity node H. Instead of being sequential, these lower fidelities are compl… view at source ↗
Figure 7
Figure 7. Figure 7: Normalized Shapley shares for L1 and L2 on the published heavy-ion test set under DAG-SVI. Left: predictive mean of H. Right: posterior variance in H. 8 Discussion 8.1 DAG-DGPs as a General Framework DAG-DGPs recover and extend several Gaussian-process architectures by restricting three components: the DAG topology, the nodewise observation pattern, and the fusion rule. This yields two concrete benefits. F… view at source ↗
Figure 8
Figure 8. Figure 8: Progressive antichain sequence on an example DAG. Each highlighted row is an antichain: there are [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Route conventions below an antichain. In panel (a), the thick path is an admissible route [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Layered block as in Assumption 1. Each row is an antichain, and all displayed edges in the analysed block run from one layer to the next. The highlighted node w ∈ Aℓ illustrates the local indegree mechanism: all its parents lie in Aℓ−1, and its descendants lie in Aℓ+1. Dashed arrows indicate omitted portions of the surrounding DAG above and below the displayed block. By Schoenberg’s theorem (Schoenberg, 1… view at source ↗
Figure 11
Figure 11. Figure 11: shows a binary instance (b = 2) of this designated branching pattern inside the layered block. Aℓ Aℓ+1 Aℓ+2 w0 Bℓ Bℓ+1 Bℓ+2 [PITH_FULL_IMAGE:figures/full_fig_p039_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Three views built from one base graph. (a) The base graph G on K = 3 vertices, with a single self-loop on v1 (highlighted) and a feedback pair v2↔v3, hence cyclic. (b) DGPG: a standard chain DGP whose state at every layer is a signal on G—shown as an identical small copy of G inside each layer node—and whose layer map f ℓ is wired by G. (c) The DAG-DGP defined on the depth-unrolled graph Ge: each node v ℓ… view at source ↗
Figure 13
Figure 13. Figure 13: Illustration of the prior-theory experiments. [PITH_FULL_IMAGE:figures/full_fig_p069_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Explaining-away and compositional uncertainty in the synthetic V-structure (Section [PITH_FULL_IMAGE:figures/full_fig_p071_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Effect of high-fidelity pretraining in the pooled 10-fold heavy-ion protocol. Validation RMSE and [PITH_FULL_IMAGE:figures/full_fig_p075_15.png] 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

114 extracted references · 9 canonical work pages · 4 internal anchors

  1. [1]

    Bayesian Optimization using Partially Observable

    Kiroriwal, Saksham and Pfrommer, Julius and Beyerer, J. Bayesian Optimization using Partially Observable. NeurIPS 2025 Workshop MLxOR: Mathematical Foundations and Operational Integration of Machine Learning for Uncertainty-Aware Decision-Making , year=

    2025

  2. [2]

    Learning Bayesian Networks with the

    Marco Scutari , journal =. Learning Bayesian Networks with the. 2010 , volume =

    2010

  3. [3]

    doi:10.48550/arXiv.1805.03108 , url =

    Ramsey, Joseph and Andrews, Bryan , year =. doi:10.48550/arXiv.1805.03108 , url =. 1805.03108 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1805.03108

  4. [4]

    and Magliacane, Sara and Claassen, Tom , title =

    Mooij, Joris M. and Magliacane, Sara and Claassen, Tom , title =. Journal of Machine Learning Research , year =

  5. [5]

    Rasmussen, Carl Edward and Williams, Christopher K. I. , year =

  6. [6]

    Foundations and Trends in Machine Learning , year =

  7. [7]

    Advances in Neural Information Processing Systems , volume =

    Salimbeni, Hugh and Deisenroth, Marc , title =. Advances in Neural Information Processing Systems , volume =. 2017 , publisher =

    2017

  8. [8]

    Digraphs: Theory, Algorithms and Applications , edition =

    Bang-Jensen, J. Digraphs: Theory, Algorithms and Applications , edition =. 2009 , publisher =

    2009

  9. [9]

    1988 , publisher =

    Pearl, Judea , title =. 1988 , publisher =

    1988

  10. [10]

    Variational Learning of Inducing Variables in Sparse

    Titsias, Michalis , booktitle =. Variational Learning of Inducing Variables in Sparse. 2009 , editor =

    2009

  11. [11]

    Proceedings of the Twenty-Ninth Conference on Uncertainty in Artificial Intelligence , pages =

    Hensman, James and Fusi, Nicol. Proceedings of the Twenty-Ninth Conference on Uncertainty in Artificial Intelligence , pages =. 2013 , publisher =

    2013

  12. [12]

    Williams, Christopher K. I. and Rasmussen, Carl Edward , title =. Advances in Neural Information Processing Systems , volume =. 1995 , publisher =

    1995

  13. [13]

    Advances in Neural Information Processing Systems , volume =

    Snelson, Edward and Ghahramani, Zoubin , title =. Advances in Neural Information Processing Systems , volume =. 2005 , publisher =

    2005

  14. [14]

    2006 , doi =

    Uncertain Judgements: Eliciting Experts' Probabilities , author =. 2006 , doi =

    2006

  15. [15]

    Qualitative Combination of

    Del Sagrado, Jos. Qualitative Combination of. International Journal of Intelligent Systems , year =

  16. [16]

    A Discrete-Time

    Boudali, Hichem and Bechta Dugan, Joanne , journal =. A Discrete-Time. 2005 , volume =

    2005

  17. [17]

    , journal =

    Heckerman, David and Geiger, Dan and Chickering, David M. , journal =. Learning. 1995 , volume =

    1995

  18. [18]

    and Girolami, Mark A

    Dunlop, Matthew M. and Girolami, Mark A. and Stuart, Andrew M. and Teckentrup, Aretha L. , journal =. How Deep Are Deep. 2018 , url =

    2018

  19. [19]

    Proceedings of the Seventeenth International Conference on Artificial Intelligence and Statistics , pages =

    Avoiding Pathologies in Very Deep Networks , author =. Proceedings of the Seventeenth International Conference on Artificial Intelligence and Statistics , pages =. 2014 , editor =

    2014

  20. [20]

    , series =

    Neal, Radford M. , series =. 1996 , doi =

    1996

  21. [21]

    and Michiardi, Pietro and Filippone, Maurizio , booktitle =

    Cutajar, Kurt and Bonilla, Edwin V. and Michiardi, Pietro and Filippone, Maurizio , booktitle =. Random Feature Expansions for Deep. 2017 , editor =

    2017

  22. [22]

    and Gonz

    Cutajar, Kurt and Pullin, Mark and Damianou, Andreas and Lawrence, Neil D. and Gonz. Deep. 2019 , journal =. doi:10.48550/arXiv.1903.07320 , url =. 1903.07320 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1903.07320 2019

  23. [23]

    Variational Auto-encoded Deep Gaussian Processes

    Dai, Zhenwen and Damianou, Andreas and Gonz. Variational Auto-encoded Deep. 2015 , journal =. doi:10.48550/arXiv.1511.06455 , url =. 1511.06455 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1511.06455 2015

  24. [24]

    Journal of Machine Learning Research , year =

    Knoblauch, Jeremias and Jewson, Jack and Damoulas, Theodoros , title =. Journal of Machine Learning Research , year =

  25. [25]

    Robust Deep Gaussian Processes

    Knoblauch, Jeremias , title =. arXiv preprint arXiv:1904.02303 , year =. doi:10.48550/arXiv.1904.02303 , url =. 1904.02303 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1904.02303 1904

  26. [26]

    and Walker, Stephen G

    Bissiri, Pier Giovanni and Holmes, Chris C. and Walker, Stephen G. , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year =

  27. [27]

    2011 , publisher =

    Goldstein, Harvey , title =. 2011 , publisher =. doi:10.1002/9780470973394 , isbn =

    work page doi:10.1002/9780470973394 2011

  28. [28]

    , title =

    Bollen, Kenneth A. , title =. 1989 , publisher =. doi:10.1002/9781118619179 , isbn =

    doi:10.1002/9781118619179 1989

  29. [29]

    , title =

    Lauritzen, Steffen L. , title =. 1996 , publisher =

    1996

  30. [30]

    and Ruppert, David and Stefanski, Leonard A

    Carroll, Raymond J. and Ruppert, David and Stefanski, Leonard A. and Crainiceanu, Ciprian M. , title =. 2006 , edition =. doi:10.1201/9781420010138 , isbn =

    doi:10.1201/9781420010138 2006

  31. [31]

    Advances in Neural Information Processing Systems , volume =

    Lindinger, Jakob and Reeb, David and Lippert, Christoph and Rakitsch, Barbara , title =. Advances in Neural Information Processing Systems , volume =. 2020 , publisher =

    2020

  32. [32]

    and Dawid, A

    Cowell, Robert G. and Dawid, A. Philip and Lauritzen, Steffen L. and Spiegelhalter, David J. , title =. 1999 , publisher =

    1999

  33. [33]

    Introduction to Statistical Relational Learning , editor =

    Koller, Daphne and Friedman, Nir and Getoor, Lise and Taskar, Ben , title =. Introduction to Statistical Relational Learning , editor =. 2007 , chapter =. doi:10.7551/mitpress/7432.003.0004 , isbn =

    work page doi:10.7551/mitpress/7432.003.0004 2007

  34. [34]

    Blair, Jean R. S. and Peyton, Barry W. , title =. Graph Theory and Sparse Matrix Computation , editor =. 1993 , publisher =

    1993

  35. [35]

    2005 , publisher =

    Rue, H. 2005 , publisher =. doi:10.1201/9780203492024 , isbn =

    doi:10.1201/9780203492024 2005

  36. [36]

    and Tarjan, Robert E

    Rose, Donald J. and Tarjan, Robert E. and Lueker, George S. , title =. SIAM Journal on Computing , volume =. 1976 , doi =

    1976

  37. [37]

    Compositional Uncertainty in Deep

    Ustyuzhaninov, Ivan and Kazlauskaite, Ieva and Kaiser, Markus and Bodin, Erik and Campbell, Neill and Ek, Carl Henrik , booktitle =. Compositional Uncertainty in Deep. 2020 , editor =

    2020

  38. [38]

    , title =

    Bishop, Christopher M. , title =. 2006 , publisher =

    2006

  39. [39]

    , title =

    Mirsky, L. , title =. The American Mathematical Monthly , volume =. 1971 , doi =

    1971

  40. [40]

    A Graphical Multi-Fidelity

    Ji, Yi and Mak, Simon and Soeder, Derek and Paquet, Jean-Fran. A Graphical Multi-Fidelity. Technometrics , volume =. 2024 , doi =

    2024

  41. [41]

    and Nickisch, Hannes and Rasmussen, Carl Edward , title =

    Duvenaud, David K. and Nickisch, Hannes and Rasmussen, Carl Edward , title =. Advances in Neural Information Processing Systems , volume =. 2011 , publisher =

    2011

  42. [42]

    Journal of Multivariate Analysis , volume =

    Durrande, Nicolas and Ginsbourger, David and Roustant, Olivier and Carraro, Luca , title =. Journal of Multivariate Analysis , volume =. 2013 , doi =

    2013

  43. [43]

    Tohoku Mathematical Journal , series =

    Azuma, Kazuoki , title =. Tohoku Mathematical Journal , series =. 1967 , doi =

    1967

  44. [44]

    Journal of the American Statistical Association , volume =

    Hoeffding, Wassily , title =. Journal of the American Statistical Association , volume =. 1963 , doi =

    1963

  45. [45]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Liu, Yuhao and Ajirak, Marzieh and Djuri. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2024 , doi =

    2024

  46. [46]

    and Jacobs, Robert A

    Jordan, Michael I. and Jacobs, Robert A. , title =. Neural Computation , volume =. 1994 , doi =

    1994

  47. [47]

    and Aitchison, Laurence , booktitle =

    Ober, Sebastian W. and Aitchison, Laurence , booktitle =. Global Inducing Point Variational Posteriors for. 2021 , editor =

    2021

  48. [48]

    and Ney, Peter E

    Athreya, Krishna B. and Ney, Peter E. , title =. 1972 , publisher =. doi:10.1007/978-3-642-65371-1 , isbn =

    doi:10.1007/978-3-642-65371-1 1972

  49. [49]

    Schoenberg, I. J. , title =. Annals of Mathematics , series =. 1938 , doi =

    1938

  50. [50]

    Schoenberg, I. J. , title =. Transactions of the American Mathematical Society , volume =. 1938 , doi =

    1938

  51. [51]

    1941 , publisher =

    Widder, David Vernon , title =. 1941 , publisher =

    1941

  52. [52]

    and O'Hagan, Anthony , title =

    Kennedy, Marc C. and O'Hagan, Anthony , title =. Biometrika , volume =. 2000 , doi =

    2000

  53. [53]

    and Karniadakis, George Em , title =

    Perdikaris, Paris and Raissi, Maziar and Damianou, Andreas and Lawrence, Neil D. and Karniadakis, George Em , title =. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume =. 2017 , doi =

    2017

  54. [54]

    and Nolan, Garry P

    Sachs, Karen and Perez, Omar and Pe'er, Dana and Lauffenburger, Douglas A. and Nolan, Garry P. , title =. Science , year =. doi:10.1126/science.1105809 , pmid =

    doi:10.1126/science.1105809

  55. [55]

    2009 , publisher =

    Pearl, Judea , title =. 2009 , publisher =

    2009

  56. [56]

    Review of Multi-Fidelity Models , journal =

    Fern. Review of Multi-Fidelity Models , journal =. 2023 , volume =

    2023

  57. [57]

    Science , year =

    Friedman, Nir , title =. Science , year =. doi:10.1126/science.1094068 , pmid =

    work page doi:10.1126/science.1094068

  58. [58]

    , title =

    Greenland, Sander and Pearl, Judea and Robins, James M. , title =. Epidemiology , year =

  59. [59]

    Little, Roderick J. A. and Rubin, Donald B. , title =. 2019 , publisher =. doi:10.1002/9781119482260 , isbn =

    doi:10.1002/9781119482260 2019

  60. [60]

    and O'Hagan, Anthony , title =

    Kennedy, Marc C. and O'Hagan, Anthony , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year =

  61. [61]

    and Matias, Catherine and Rhodes, John A

    Allman, Elizabeth S. and Matias, Catherine and Rhodes, John A. , title =. The Annals of Statistics , year =

  62. [62]

    Structural and Practical Identifiability Analysis of Partially Observed Dynamical Models by Exploiting the Profile Likelihood , journal =

    Raue, Andreas and Kreutz, Clemens and Maiwald, Thomas and Bachmann, Julie and Schilling, Marcel and Klingm. Structural and Practical Identifiability Analysis of Partially Observed Dynamical Models by Exploiting the Profile Likelihood , journal =. 2009 , volume =

    2009

  63. [63]

    , title =

    Neal, Radford M. , title =. 1995 , address =

    1995

  64. [64]

    Philip , title =

    Dawid, A. Philip , title =. Journal of the Royal Statistical Society: Series B (Methodological) , year =

  65. [65]

    and Jordan, Michael I

    Wainwright, Martin J. and Jordan, Michael I. , title =. Foundations and Trends in Machine Learning , year =. doi:10.1561/2200000001 , publisher =

    doi:10.1561/2200000001

  66. [66]

    , booktitle =

    Damianou, Andreas and Lawrence, Neil D. , booktitle =. Deep. 2013 , editor =

    2013

  67. [67]

    and Moore, Andrew J

    Lawrence, Neil D. and Moore, Andrew J. , booktitle =. Hierarchical. 2007 , publisher =. doi:10.1145/1273496.1273557 , url =

    doi:10.1145/1273496.1273557 2007

  68. [68]

    , booktitle =

    Lawrence, Neil D. , booktitle =. 2003 , publisher =

    2003

  69. [69]

    , journal =

    Lawrence, Neil D. , journal =. Probabilistic Non-Linear Principal Component Analysis with. 2005 , url =

    2005

  70. [70]

    and Hern

    Bui, Thang D. and Hern. Deep. Proceedings of the 33rd International Conference on Machine Learning , pages =. 2016 , volume =

    2016

  71. [71]

    Inference in Deep

    Havasi, Marton and Hern. Inference in Deep. Advances in Neural Information Processing Systems , volume =. 2018 , publisher =

    2018

  72. [72]

    and Higdon, David , journal =

    Sauer, Annie and Gramacy, Robert B. and Higdon, David , journal =. Active Learning for Deep. 2023 , doi =

    2023

  73. [73]

    Hern. Deep. Machine Learning and Knowledge Discovery in Databases , editor =. 2021 , publisher =. doi:10.1007/978-3-030-67664-3_29 , note =

    work page doi:10.1007/978-3-030-67664-3_29 2021

  74. [74]

    Alpha-Divergence Minimization for Deep

    Villacampa-Calvo, Carlos and Hern. Alpha-Divergence Minimization for Deep. International Journal of Approximate Reasoning , volume =. 2022 , doi =

    2022

  75. [75]

    Salimbeni, Hugh and Dutordoir, Vincent and Hensman, James and Deisenroth, Marc , booktitle =. Deep. 2019 , editor =

    2019

  76. [76]

    Advances in Neural Information Processing Systems , volume =

    Yu, Haibin and Chen, Yizhou and Low, Bryan Kian Hsiang and Jaillet, Patrick and Dai, Zhongxiang , title =. Advances in Neural Information Processing Systems , volume =. 2019 , publisher =

    2019

  77. [77]

    and Gonz

    Dai, Zhenwen and Damianou, Andreas C. and Gonz. Variational Auto-encoded Deep. Proceedings of the 4th International Conference on Learning Representations , year =. 1511.06455 , archivePrefix =

    Pith/arXiv arXiv

  78. [78]

    Recurrent

    Mattos, C. Recurrent. Proceedings of the International Conference on Learning Representations , year =. 1511.06644 , archivePrefix =

    Pith/arXiv arXiv

  79. [79]

    Matthews, Alexander G. de G. and Hensman, James and Turner, Richard and Ghahramani, Zoubin , booktitle =. On Sparse Variational Methods and the. 2016 , editor =

    2016

  80. [80]

    Nested Variational Compression in Deep Gaussian Processes

    Hensman, James and Lawrence, Neil D. , year =. Nested Variational Compression in Deep. doi:10.48550/arXiv.1412.1370 , url =. 1412.1370 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1412.1370

Showing first 80 references.