pith. sign in

arxiv: 2003.06804 · v1 · pith:CNSOFMHVnew · submitted 2020-03-15 · 📊 stat.ME · math.ST· stat.ML· stat.TH

Semi-Modular Inference: enhanced learning in multi-modular models by tempering the influence of components

Chris U. Carmona , Geoff K. Nicholls This is my paper

Pith reviewed 2026-05-24 14:24 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.MLstat.TH
keywords semi-modular inferencecut-model inferenceBayesian inferencemodel misspecificationmulti-modular modelsinfluence parametermeta-learning
0
0 comments X

The pith

Semi-Modular Inference introduces an influence parameter for tunable directed information flow between modules, recovering Bayesian inference when modules are correctly specified.

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

The paper defines a family of Semi-Modular Inference schemes inside a loss-based generalization of Bayesian inference. These schemes are indexed by an influence parameter that controls how much each module contributes to the overall posterior, with full Bayesian inference and cut-model inference as boundary cases. A meta-learning criterion selects the parameter from data and returns the Bayesian solution under correct specification. This construction applies to multi-modular models and permits directed rather than undirected control of information flow from well-specified to misspecified components.

Core claim

Within an existing coherent loss-based generalisation of Bayesian inference, the authors show that modular cut-model inference is coherent and introduce Semi-Modular Inference schemes indexed by an influence parameter. Bayesian inference and cut-models appear as special cases. A meta-learning procedure estimates the parameter so that the scheme returns Bayesian inference when there is no misspecification. The resulting family supports tunable directed information flow between modules in multi-modular models.

What carries the argument

The influence parameter that scales the contribution of each module's posterior or likelihood term, thereby controlling the strength and direction of information flow between modules.

If this is right

  • Cut-model inference arises when the influence parameter from misspecified modules is set to zero.
  • Full Bayesian inference is recovered when the influence parameter equals one and all modules are correctly specified.
  • The meta-learning criterion supplies a data-driven rule for balancing information exchange between modules.
  • SMI extends naturally to any multi-modular model in which some components may be misspecified.

Where Pith is reading between the lines

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

  • The same influence-parameter construction could be used to select partial pooling strengths in hierarchical models.
  • The meta-learning step suggests a route for automatic module weighting in composite machine-learning pipelines.
  • Directed tempering may interact with existing power-posterior methods to produce hybrid schemes with both direction and continuous strength control.

Load-bearing premise

The meta-learning criterion for selecting the influence parameter is well-behaved and recovers the Bayesian solution under correct model specification.

What would settle it

A simulation in which the data-generating process exactly matches the model yet the estimated influence parameter is not one, or a misspecification experiment in which the selected SMI scheme fails to improve predictive performance over standard Bayesian inference.

Figures

Figures reproduced from arXiv: 2003.06804 by Chris U. Carmona, Geoff K. Nicholls.

Figure 1
Figure 1. Figure 1: Graphical representation of a simple multi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Model assessment for biased data example. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Top: Estimated elpd as predictive criteria [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results expand on but support those reported [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Joint SMI posterior for θ1 and θ2 for the HPV model using MCMC on the SMI posterior with η ∈ [0, 1] −elpd binomial −elpd poisson 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0 500 1000 36 40 44 48 eta (over poisson module) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 1
Figure 1. Figure 1: Graphical representation of a simple multi-modular model. [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Posterior distribution of ϕ, θ and ˜θ for a single dataset (Z,Y ). A black horizontal line shows the true generative value. The posterior mean is the solid red line and we show intervals with ± one posterior std. dev. using dotted blue lines. 5.1.3 Expected log pointwise predictive density (elpd) The elpd is elpd = Z Z p ∗ (z, y) log psmi,η(z, y | Z, Y )dzdy where p ∗ is the distribution representing the t… view at source ↗
Figure 3
Figure 3. Figure 3: Mean Squared Error of the two main parameters under SMI posterior. [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of the bias estimation under SMI (theta) vs powered likelihood (theta tilde) [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ELPD under SMI posterior. 5.2 Agricultural data The aim of the study, described in detail in Styring et al. (2017), is to provide statistical evidence about a specific agricultural practice of the first urban centres in northern Mesopotamia. The hypothesis is that increased agricultural production to support growing urban populations was achieved by cultivation of larger areas of land, entailing lower manu… view at source ↗
Figure 6
Figure 6. Figure 6: Graphical representation of the model for the agricultural data. Squares denote observable [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Simplified representation of the model for the agricultural data. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: SMI posterior distribution for the parameter of interest in the agricultural model. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Choosing the best SMI posterior candidate by choosing the [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Bayes Factor for the hypothesis [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
read the original abstract

Bayesian statistical inference loses predictive optimality when generative models are misspecified. Working within an existing coherent loss-based generalisation of Bayesian inference, we show existing Modular/Cut-model inference is coherent, and write down a new family of Semi-Modular Inference (SMI) schemes, indexed by an influence parameter, with Bayesian inference and Cut-models as special cases. We give a meta-learning criterion and estimation procedure to choose the inference scheme. This returns Bayesian inference when there is no misspecification. The framework applies naturally to Multi-modular models. Cut-model inference allows directed information flow from well-specified modules to misspecified modules, but not vice versa. An existing alternative power posterior method gives tunable but undirected control of information flow, improving prediction in some settings. In contrast, SMI allows tunable and directed information flow between modules. We illustrate our methods on two standard test cases from the literature and a motivating archaeological data set.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Semi-Modular Inference (SMI), a family of coherent inference schemes for multi-modular models indexed by an influence parameter that includes standard Bayesian inference and cut-model inference as special cases. It proposes a meta-learning criterion together with an estimation procedure for selecting the influence parameter (and thus the inference scheme), with the property that the procedure recovers Bayesian inference under correct model specification. SMI is positioned as allowing tunable, directed information flow between modules (in contrast to undirected power-posterior tempering), and the methods are illustrated on two standard test cases plus an archaeological data set.

Significance. If the meta-learning criterion is shown to be well-behaved and to recover the Bayesian solution under correct specification, the framework would supply a principled, loss-based route to directed tempering of misspecified modules while preserving coherence with existing modular inference. The explicit contrast with power posteriors and the recovery guarantee under correct specification would be genuine strengths.

major comments (2)
  1. [Abstract / §3 (meta-learning criterion)] The central claim that the meta-learning criterion and estimation procedure recover Bayesian inference exactly when there is no misspecification is load-bearing for the whole contribution, yet the manuscript provides no explicit statement of the loss function, validation scheme, or convergence argument establishing that the selected influence parameter converges to the value that yields full Bayesian inference under correct specification.
  2. [§4] §4 (or wherever the directed-flow property is formalized): the claim that SMI permits directed information flow (well-specified modules to misspecified but not vice versa) while remaining coherent needs an explicit comparison showing that the chosen influence parameter does not inadvertently allow reverse flow or reduce to an undirected tempering scheme.
minor comments (2)
  1. [Throughout] Notation for the influence parameter and the modular decomposition should be introduced once and used consistently; several passages reuse symbols without redefinition.
  2. [Empirical section] The archaeological data example would benefit from a clearer statement of which modules are treated as misspecified and how the influence parameter is estimated in practice.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed report. We address the two major comments point by point below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / §3 (meta-learning criterion)] The central claim that the meta-learning criterion and estimation procedure recover Bayesian inference exactly when there is no misspecification is load-bearing for the whole contribution, yet the manuscript provides no explicit statement of the loss function, validation scheme, or convergence argument establishing that the selected influence parameter converges to the value that yields full Bayesian inference under correct specification.

    Authors: We agree that the presentation of the meta-learning criterion in §3 would benefit from greater explicitness. The current manuscript defines the criterion and states the recovery property, but does not isolate the loss function or supply a formal convergence argument. In the revision we will insert a dedicated paragraph (or short subsection) that (i) states the precise loss function, (ii) describes the validation scheme used for estimation, and (iii) provides a brief consistency argument showing that the selected influence parameter converges to the value corresponding to full Bayesian inference when all modules are correctly specified. revision: yes

  2. Referee: [§4] §4 (or wherever the directed-flow property is formalized): the claim that SMI permits directed information flow (well-specified modules to misspecified but not vice versa) while remaining coherent needs an explicit comparison showing that the chosen influence parameter does not inadvertently allow reverse flow or reduce to an undirected tempering scheme.

    Authors: The directed-flow property follows from the asymmetric role of the influence parameter in the factorised posterior; however, we accept that an explicit side-by-side comparison with power posteriors is currently only sketched. In the revised §4 we will add a short analytical subsection that (a) writes the SMI posterior factorisation for a two-module model, (b) shows that the influence parameter modulates information from the well-specified module to the misspecified module without the reverse effect, and (c) contrasts this with the symmetric tempering of power posteriors, including a brief derivation confirming that the scheme does not collapse to undirected tempering for any interior value of the influence parameter. revision: yes

Circularity Check

0 steps flagged

No circularity: meta-learning criterion presented as independent selection procedure

full rationale

The paper defines SMI as a parameterized family with Bayesian and cut-model inference as boundary cases, then separately introduces a meta-learning criterion plus estimation procedure to select the influence parameter, asserting (without shown reduction) that the procedure recovers the Bayesian solution under correct specification. No equation or step in the provided text equates the selection criterion to the fitted parameter by construction, nor does any load-bearing claim reduce to a self-citation chain or renamed ansatz. The coherence argument for cut-models and the directed-flow contrast with power posteriors are stated as derived properties independent of the meta-learner. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the influence parameter itself is introduced as part of the new family but its status (fitted or derived) is not stated.

pith-pipeline@v0.9.0 · 5705 in / 1173 out tokens · 15563 ms · 2026-05-24T14:24:04.082373+00:00 · methodology

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

24 extracted references · 24 canonical work pages

  1. [1]

    Bernardo, J. M. and Smith, A. F. (2000). Bayesian Theory . Wiley Series in Probability and Statistics. John Wiley & Sons, Inc., Hoboken, NJ, USA, 3 edition

    work page 2000

  2. [2]

    G., Holmes, C

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

    work page 2016

  3. [3]

    Cox, D. R. (1975). Partial likelihood . Biometrika , 62(2):269--276

    work page 1975

  4. [4]

    Gr \" u nwald, P. (2012). The Safe Bayesian . In Bshouty, N. H., Stoltz, G., Vayatis, N., and Zeugmann, T., editors, Algorithmic Learning Theory: 23rd International Conference, ALT 2012, Lyon, France, October 29-31, 2012. Proceedings , volume 7568 LNAI, pages 169--183. Springer Berlin Heidelberg

    work page 2012

  5. [5]

    and van Ommen, T

    Gr \" u nwald, P. and van Ommen, T. (2017). Inconsistency of Bayesian Inference for Misspecified Linear Models, and a Proposal for Repairing It . Bayesian Analysis , 12(4):1069--1103

    work page 2017

  6. [6]

    Holmes, C. C. and Walker, S. G. (2017). Assigning a value to a power likelihood in a general Bayesian model . Biometrika , 104(2):497--503

    work page 2017

  7. [7]

    E., Murray, L

    Jacob, P. E., Murray, L. M., Holmes, C. C., and Robert, C. P. (2017a). Better together? Statistical learning in models made of modules

    work page

  8. [8]

    E., O'Leary, J., and Atchad \' e , Y

    Jacob, P. E., O'Leary, J., and Atchad \' e , Y. F. (2017b). Unbiased Markov chain Monte Carlo with couplings

    work page

  9. [9]

    W., Divitini, M

    Knuiman, M. W., Divitini, M. L., Buzas, J. S., and Fitzgerald, P. E. (1998). Adjustment for Regression Dilution in Epidemiological Regression Analyses . Annals of Epidemiology , 8(1):56--63

    work page 1998

  10. [10]

    Little, R. J. A. and Rubin, D. B. (2002). Statistical Analysis with Missing Data . John Wiley & Sons, Inc., Hoboken, NJ, USA, 2nd edition

    work page 2002

  11. [11]

    J., and Berger, J

    Liu, F., Bayarri, M. J., and Berger, J. O. (2009). Modularization in Bayesian analysis, with emphasis on analysis of computer models . Bayesian Analysis , 4(1):119--150

    work page 2009

  12. [12]

    Lunn, D., Best, N., Spiegelhalter, D., Graham, G., and Neuenschwander, B. (2009). Combining MCMC with ‘sequential’ PKPD modelling . Journal of Pharmacokinetics and Pharmacodynamics , 36(1):19--38

    work page 2009

  13. [13]

    Maucort-Boulch, D., Franceschi, S., and Plummer, M. (2008). International Correlation between Human Papillomavirus Prevalence and Cervical Cancer Incidence . Cancer Epidemiology Biomarkers & Prevention , 17(3):717--720

    work page 2008

  14. [14]

    Meng, X.-L. (1994). Multiple-Imputation Inferences with Uncongenial Sources of Input . Statistical Science , 9(4):538--558

    work page 1994

  15. [15]

    Miller, J. W. and Dunson, D. B. (2018). Robust Bayesian Inference via Coarsening . Journal of the American Statistical Association , pages 1--13

    work page 2018

  16. [16]

    Plummer, M. (2015). Cuts in Bayesian graphical models . Statistics and Computing , 25(1):37--43

    work page 2015

  17. [17]

    J., Thomas, A., Best, N., and Lunn, D

    Spiegelhalter, D. J., Thomas, A., Best, N., and Lunn, D. (2014). OpenBUGS User Manual

    work page 2014

  18. [18]

    K., Charles, M., Fantone, F., Hald, M

    Styring, A. K., Charles, M., Fantone, F., Hald, M. M., McMahon, A., Meadow, R. H., Nicholls, G. K., Patel, A. K., Pitre, M. C., Smith, A., So?tysiak, A., Stein, G., Weber, J. A., Weiss, H., and Bogaard, A. (2017). Isotope evidence for agricultural extensification reveals how the world's first cities were fed . Nature Plants , 3(6)

    work page 2017

  19. [19]

    Vehtari, A., Gelman, A., and Gabry, J. (2017). Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC . Statistics and Computing , 27(5):1413--1432

    work page 2017

  20. [20]

    and Hjort, N

    Walker, S. and Hjort, N. L. (2001). On Bayesian consistency . Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 63(4):811--821

    work page 2001

  21. [21]

    Watanabe, S. (2009). Algebraic Geometry and Statistical Learning Theory . Cambridge Monographs on Applied and Computational Mathematics. Cambridge University Press

    work page 2009

  22. [22]

    and Meng, X.-L

    Xie, X. and Meng, X.-L. (2016). Dissecting multiple imputation from a multi-phase inference perspective: what happens when God’s, imputer’s and analyst’s models are uncongenial? Statistica Sinica , 27:1485--1594

    work page 2016

  23. [23]

    Carpenter, B., Gelman, A., Hoffman, M., Lee, D., Goodrich, B., Betancourt, M., Brubaker, M., Guo, J., Li, P., and Riddell, A. (2017). Stan: A Probabilistic Programming Language . Journal of Statistical Software, Articles , 76(1):1--32

    work page 2017

  24. [24]

    E., O'Leary, J., and Atchad \' e , Y

    Jacob, P. E., O'Leary, J., and Atchad \' e , Y. F. (2017). Unbiased Markov chain Monte Carlo with couplings

    work page 2017