Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 94 of the 95 references it cites were found in a registry or an index.

Neuronal message passing using Mean-field, Bethe, and Marginal approximations

Open access
Parr, Thomas, Markovic, Dimitrije, Kiebel, Stefan J., Friston, Karl J. · 2019 · Scientific Reports

Abstract

Neuronal computations rely upon local interactions across synapses. For a neuronal network to perform inference, it must integrate information from locally computed messages that are propagated among elements of that network. We review the form of two popular (Bayesian) message passing schemes and consider their plausibility as descriptions of inference in biological networks. These are variational message passing and belief propagation – each of which is derived from a free energy functional that relies upon different approximations (mean-field and Bethe respectively). We begin with an overview of these schemes and illustrate the form of the messages required to perform inference using Hidden Markov Models as generative models. Throughout, we use factor graphs to show the form of the generative models and of the messages they entail. We consider how these messages might manifest neuronally and simulate the inferences they perform. While variational message passing offers a simple and neuronally plausible architecture, it falls short of the inferential performance of belief propagation. In contrast, belief propagation allows exact computation of marginal posteriors at the expense of the architectural simplicity of variational message passing. As a compromise between these two extremes, we offer a third approach – marginal message passing – that features a simple architecture, while approximating the performance of belief propagation. Finally, we link formal considerations to accounts of neurological and psychiatric syndromes in terms of aberrant message passing.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41598-018-38246-3

Cites

94 of 94 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 A graphical representation of a probabilistic model. For a generative model, expressed as a joint probability distribution, it is possible to write down the associated factor graph by following a few simple steps. First, the model may be expressed in terms of the factors (prior and conditional distributions) that make up the joint probability. Square nodes are then associated with each of these factors. These nodes are connected whenever they are functions of the same random variable. The result is known as a normal factor graph13,20. For comparison, we present the same generative model expressed according to two alternative graphical representations. The Bayes net shown on the left places random variables in circular nodes and connects these with arrows corresponding to the conditional distributions. An alternative factor graph representation is shown on the right. This combines the normal factor graph formalism with that of the Bayes net; incorporating both factor and variable nodes. For the rest of this paper, we adopt the normal factor graph formalism as this provides a natural way to think about the form of local message passing.
Figure 1 A graphical representation of a probabilistic model. For a generative model, expressed as a joint probability distribution, it is possible to write down the associated factor graph by following a few simple steps. First, the model may be expressed in terms of the factors (prior and conditional distributions) that make up the joint probability. Square nodes are then associated with each of these factors. These nodes are connected whenever they are functions of the same random variable. The result is known as a normal factor graph13,20. For comparison, we present the same generative model expressed according to two alternative graphical representations. The Bayes net shown on the left places random variables in circular nodes and connects these with arrows corresponding to the conditional distributions. An alternative factor graph representation is shown on the right. This combines the normal factor graph formalism with that of the Bayes net; incorporating both factor and variable nodes. For the rest of this paper, we adopt the normal factor graph formalism as this provides a natural way to think about the form of local message passing.CC BY · PubMed Central
Figure 2 Commonly used (static) generative models as factor graphs. This figure illustrates the generative models that underwrite many common inferential procedures. In each factor graph, small blue squares indicate observable data, while squares with an equality sign relate their adjoining edges via a delta function factor (shown explicitly in lower right inset in Fig. 3 - Hidden Markov model). Numbered squares relate factors to those in the probabilistic models in the blue panels. This Figure illustrates static models of the sort that underlie factor analysis (FA), probabilistic principle component analysis (PPCA), and principal component analysis (PCA)21. Each of these dimensionality reduction techniques relies upon the same generative model, but with different assumptions about the covariance structure of latent causes or sources. Adding in non-linear functions allows this generative model to be extended to incorporate independent component analysis (ICA)94, while using two different linear transforms leads to probabilistic canonical variates analysis (CVA)95. Incorporating discrete random variables gives mixture models including mixtures of Gaussians (MoG), which form the basis of many clustering algorithms96. The notation N indicates a normal (Gaussian) distribution, while Cat means a categorical distribution.
Figure 2 Commonly used (static) generative models as factor graphs. This figure illustrates the generative models that underwrite many common inferential procedures. In each factor graph, small blue squares indicate observable data, while squares with an equality sign relate their adjoining edges via a delta function factor (shown explicitly in lower right inset in Fig. 3 - Hidden Markov model). Numbered squares relate factors to those in the probabilistic models in the blue panels. This Figure illustrates static models of the sort that underlie factor analysis (FA), probabilistic principle component analysis (PPCA), and principal component analysis (PCA)21. Each of these dimensionality reduction techniques relies upon the same generative model, but with different assumptions about the covariance structure of latent causes or sources. Adding in non-linear functions allows this generative model to be extended to incorporate independent component analysis (ICA)94, while using two different linear transforms leads to probabilistic canonical variates analysis (CVA)95. Incorporating discrete random variables gives mixture models including mixtures of Gaussians (MoG), which form the basis of many clustering algorithms96. The notation N indicates a normal (Gaussian) distribution, while Cat means a categorical distribution.CC BY · PubMed Central
Figure 3 Commonly used (dynamic) generative models. This shows the dynamic generative models that support Kalman filtering97 and generalised filtering98 (the basis of predictive coding schemes99 and inversion of dynamic causal models100). A discrete state space model that exhibits temporal dynamics is a hidden Markov model. This is the generative model we will take as our example for the remainder of this paper and, for this reason, we have expressed this in full, with a sequence of transitions over time. The notation N indicates a normal (Gaussian) distribution, while Cat means a categorical distribution.
Figure 3 Commonly used (dynamic) generative models. This shows the dynamic generative models that support Kalman filtering97 and generalised filtering98 (the basis of predictive coding schemes99 and inversion of dynamic causal models100). A discrete state space model that exhibits temporal dynamics is a hidden Markov model. This is the generative model we will take as our example for the remainder of this paper and, for this reason, we have expressed this in full, with a sequence of transitions over time. The notation N indicates a normal (Gaussian) distribution, while Cat means a categorical distribution.CC BY · PubMed Central
Figure 4 Message passing in a hidden Markov model. This schematic illustrates the scheduling of belief propagation (left) and variational message passing (right) in a hidden Markov model. Each row shows a single step in a round of message passing, ordered from top to bottom. Under belief propagation, for a message to be sent across a factor to an edge, the factor requires all of the other adjoining edges to provide a message. Initially, in the first step, this is only true for likelihood factors (that compute their message from the data), and priors (that are associated with just one edge). This enforces a strict scheduling that starts with the computation of messages at the extremities of the graph. More proximal factors then use these messages from the extremes to compute their own. Eventually, when all factors have passed their message, the incoming messages to each edge can be combined to compute the marginal posterior belief (Q) about the associated random variable (last row). For directed acyclic graphs, one round of message passing is sufficient. For cyclic graphs, multiple rounds may be required. In contrast, variational message passing computes messages from the current beliefs associated with each edge, and not from other incoming messages. This means that variational message passing simply alternates between message passing (of all messages in parallel) and updating posterior expectations.
Figure 4 Message passing in a hidden Markov model. This schematic illustrates the scheduling of belief propagation (left) and variational message passing (right) in a hidden Markov model. Each row shows a single step in a round of message passing, ordered from top to bottom. Under belief propagation, for a message to be sent across a factor to an edge, the factor requires all of the other adjoining edges to provide a message. Initially, in the first step, this is only true for likelihood factors (that compute their message from the data), and priors (that are associated with just one edge). This enforces a strict scheduling that starts with the computation of messages at the extremities of the graph. More proximal factors then use these messages from the extremes to compute their own. Eventually, when all factors have passed their message, the incoming messages to each edge can be combined to compute the marginal posterior belief (Q) about the associated random variable (last row). For directed acyclic graphs, one round of message passing is sufficient. For cyclic graphs, multiple rounds may be required. In contrast, variational message passing computes messages from the current beliefs associated with each edge, and not from other incoming messages. This means that variational message passing simply alternates between message passing (of all messages in parallel) and updating posterior expectations.CC BY · PubMed Central
Figure 5 Belief propagation as neuronal message passing. The equations on the left show the form of belief propagation (Equations 2 and 3) when expressed in a neuronally plausible form. These equations are written in terms of the sufficient statistics of the probability distributions and auxiliary variables representing prediction errors (ετ) and membrane potentials (vτ). The softmax (σ) functions act as neuronal transfer functions, converting presynaptic potentials to firing rates (sτ), which represent the sufficient statistics of the posterior beliefs. Forwards and backwards messages across the transition factors (B) are written as (μτ→)^B,(μτ←)^B respectively. Red indicates an excitatory connection, while blue is inhibitory. The starred connection represents the subtraction of the ascending message from that passed on to other neuronal populations. This plays the role of an ascending ‘loop’, as in circular inference accounts of neuronal computation4,37. The analogous descending loops are the inhibitory connections between the neurons representing messages in opposite directions. This formulation assumes that the neurons representing the messages have much shorter time constants than those representing marginal beliefs, allowing the former to be ‘enslaved’ by the latter15,101. Although omitted here (and in later figures) for simplicity, the normalisation induced by the softmax functions could be mediated via recurrent inhibitory connections within a layer of neurons.
Figure 5 Belief propagation as neuronal message passing. The equations on the left show the form of belief propagation (Equations 2 and 3) when expressed in a neuronally plausible form. These equations are written in terms of the sufficient statistics of the probability distributions and auxiliary variables representing prediction errors (ετ) and membrane potentials (vτ). The softmax (σ) functions act as neuronal transfer functions, converting presynaptic potentials to firing rates (sτ), which represent the sufficient statistics of the posterior beliefs. Forwards and backwards messages across the transition factors (B) are written as (μτ→)^B,(μτ←)^B respectively. Red indicates an excitatory connection, while blue is inhibitory. The starred connection represents the subtraction of the ascending message from that passed on to other neuronal populations. This plays the role of an ascending ‘loop’, as in circular inference accounts of neuronal computation4,37. The analogous descending loops are the inhibitory connections between the neurons representing messages in opposite directions. This formulation assumes that the neurons representing the messages have much shorter time constants than those representing marginal beliefs, allowing the former to be ‘enslaved’ by the latter15,101. Although omitted here (and in later figures) for simplicity, the normalisation induced by the softmax functions could be mediated via recurrent inhibitory connections within a layer of neurons.CC BY · PubMed Central
Figure 6 Variational (and marginal) message passing. The equations on the left show variational message passing (upper panel) and marginal message passing (lower panel) expressed as gradient descents on the variational (or marginal) free energy. These equations are implemented by the neuronal network shown on the right. Notably, this is much simpler than the network of Fig. 5, requiring fewer neurons and a simpler connectivity structure. The primary reason for the simplicity of this structure is that these schemes take into account the current marginal beliefs of adjacent variables. The messages do not need to be recursively computed from other messages, Fig. 5. This limits the number of auxiliary variables required. We have introduced the notation B^(†) for the transpose of the transition matrix with normalised columns.
Figure 6 Variational (and marginal) message passing. The equations on the left show variational message passing (upper panel) and marginal message passing (lower panel) expressed as gradient descents on the variational (or marginal) free energy. These equations are implemented by the neuronal network shown on the right. Notably, this is much simpler than the network of Fig. 5, requiring fewer neurons and a simpler connectivity structure. The primary reason for the simplicity of this structure is that these schemes take into account the current marginal beliefs of adjacent variables. The messages do not need to be recursively computed from other messages, Fig. 5. This limits the number of auxiliary variables required. We have introduced the notation B^(†) for the transpose of the transition matrix with normalised columns.CC BY · PubMed Central
Figure 7 Marginal (forward-backward) models. This schematic illustrates the steps that motivate the marginal free energy. On the left, we show an HMM that is divided in two different ways. For the future part (lower row), we reverse the direction of the transition probabilities by normalising with respect to the earlier times. On the right, we take these partitioned generative models and sum over all variables within the dashed boxes. This leads to two, marginal, generative models – one that progresses from the past to the future, and one that reverses this. By approximating the free energies for each model, and mixing them in equal parts, we define a marginal free energy with empirical priors that are independently constrained by the future and the past.
Figure 7 Marginal (forward-backward) models. This schematic illustrates the steps that motivate the marginal free energy. On the left, we show an HMM that is divided in two different ways. For the future part (lower row), we reverse the direction of the transition probabilities by normalising with respect to the earlier times. On the right, we take these partitioned generative models and sum over all variables within the dashed boxes. This leads to two, marginal, generative models – one that progresses from the past to the future, and one that reverses this. By approximating the free energies for each model, and mixing them in equal parts, we define a marginal free energy with empirical priors that are independently constrained by the future and the past.CC BY · PubMed Central
Figure 8 Probability distributions for simulated HMM. The probability distributions here make up the generative model we used to simulate neuronal message passing under both schemes. These probabilities have been chosen to make several points. Firstly, we separated the hidden state into two distinct state factors (light blue shapes and light green shapes). This allowed us to generate data (darker blue shapes) that depends upon only one of these (light blue). The likelihood mapping illustrates how data are (probabilistically) generated from these states (P(oτ=i|sτ=j)=Aij). We used deterministic priors for the initial states (P(s1=i)=Di) and stochastic transitions (P(sτ+1=i|sτ=j)=Bij).
Figure 8 Probability distributions for simulated HMM. The probability distributions here make up the generative model we used to simulate neuronal message passing under both schemes. These probabilities have been chosen to make several points. Firstly, we separated the hidden state into two distinct state factors (light blue shapes and light green shapes). This allowed us to generate data (darker blue shapes) that depends upon only one of these (light blue). The likelihood mapping illustrates how data are (probabilistically) generated from these states (P(oτ=i|sτ=j)=Aij). We used deterministic priors for the initial states (P(s1=i)=Di) and stochastic transitions (P(sτ+1=i|sτ=j)=Bij).CC BY · PubMed Central
Figure 9 Simulated neuronal message passing for the three schemes: variational message passing, belief propagation and Marginal message passing. These plots show the consequences of generating data from the model described in Fig. 8 and solving the equations of Figs 5 and 6 for these data. The upper plots show the beliefs (s˜) throughout the trial in terms of the sufficient statistics of the categorical distributions (i.e. probability of each alternative state at each time). These depict belief updating in terms of expectations about the two hidden factors, each with three levels. Crucially, these beliefs are about each hidden state at (15) different points in time. Each line is then the posterior probability that a given hidden state at a given time takes on a specific value. The colour-coding of these lines is consistent between the plots along the upper row. These plots are important in that they give a sense of the time-course of belief updating. While variational message passing shows rapid changes at the very start of the simulation (nearly every line reaches an extreme value within the first time-step) and few thereafter, the updates of belief propagation and marginal message passing happen over a much longer time scale. The beliefs after each successive outcome are shown in the second and third rows (with black = 1 and white = 0). Each line in the plots in the first row therefore represents a cell in the second or third row – drawn at the time point encoded by each line. Red dots indicate the ‘true’ states used to generate the data. Note that there are no red dots associated with green circles, as these never occur when the initial green state is a square (see the transition matrix in Fig. 8). The final row shows the sequence of observations presented to each message passing scheme. The second and third rows show the posterior beliefs about the two factors comprising the hidden state space as shown in Fig. 8.
Figure 9 Simulated neuronal message passing for the three schemes: variational message passing, belief propagation and Marginal message passing. These plots show the consequences of generating data from the model described in Fig. 8 and solving the equations of Figs 5 and 6 for these data. The upper plots show the beliefs (s˜) throughout the trial in terms of the sufficient statistics of the categorical distributions (i.e. probability of each alternative state at each time). These depict belief updating in terms of expectations about the two hidden factors, each with three levels. Crucially, these beliefs are about each hidden state at (15) different points in time. Each line is then the posterior probability that a given hidden state at a given time takes on a specific value. The colour-coding of these lines is consistent between the plots along the upper row. These plots are important in that they give a sense of the time-course of belief updating. While variational message passing shows rapid changes at the very start of the simulation (nearly every line reaches an extreme value within the first time-step) and few thereafter, the updates of belief propagation and marginal message passing happen over a much longer time scale. The beliefs after each successive outcome are shown in the second and third rows (with black = 1 and white = 0). Each line in the plots in the first row therefore represents a cell in the second or third row – drawn at the time point encoded by each line. Red dots indicate the ‘true’ states used to generate the data. Note that there are no red dots associated with green circles, as these never occur when the initial green state is a square (see the transition matrix in Fig. 8). The final row shows the sequence of observations presented to each message passing scheme. The second and third rows show the posterior beliefs about the two factors comprising the hidden state space as shown in Fig. 8.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 95 references

Research works printed but not linked: 6

  1. Deneve, S. In Advances in neural information processing systems. 353–360.
  2. Dauwels, J. In Information Theory, 2007. ISIT 2007. IEEE International Symposium on. 2546–2550 (IEEE).
  3. Jaakkola, T., Singh, S. P. & Jordan, M. I. In Advances in neural information processing systems. 345–352.
  4. Heskes, T. In Advances in neural information processing systems. 359–366.
  5. Weller, A., Tang, K., Sontag, D. & Jebara, T. In 30th Conference on Uncertainty in Artificial Intelligence, UAI. (AUAI Press, 2014).
  6. Nowlan, S. J. In Advances in neural information processing systems. 574–582.

Cited by

See all 132

Metadata

Authors
Parr, T., Markovic, D., Kiebel, S. J., & Friston, K. J.
Affiliations
  • Wellcome Centre for Human Neuroimaging (GB)
  • Technische Universität Dresden (DE)
Corresponding author
Parr, Thomas
Venue
Scientific Reports
Published
2019-02-13
Volume
9
Issue
1
Pages
1889-1889
ISSN
2045-2322 (ISSN-L)
OpenAlex
W2918957485
Cited by
128
References
95
Citations by year
2026: 10 · 2025: 10 · 2024: 7 · 2023: 16 · 2022: 17 · 2021: 32 · 2020: 27 · 2019: 9
FWCI
9.69
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • UCL Discovery (University College London) (submittedVersion)
Funders
  • Rosetrees Trust (173346)
  • Deutsche Forschungsgemeinschaft (SFB 940/2, Project A9)
  • Wellcome Trust (088130/Z/09/Z)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Bayesian Modeling and Causal Inference
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
Keywords
variational message passing, belief propagation, mean-field approximation, Bethe approximation, free energy principle, hidden Markov models, factor graphs, marginal posterior, neuropsychiatric disorders

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Neuronal message passing using Mean-field, Bethe, and Marginal approximations. https://app.cochat.ai/papers/10.1038/s41598-018-38246-3
The paper
Parr, T., Markovic, D., Kiebel, S. J., & Friston, K. J. (2019). Neuronal message passing using Mean-field, Bethe, and Marginal approximations. Scientific Reports. https://doi.org/10.1038/s41598-018-38246-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.