CoChat

Neural Dynamics under Active Inference: Plausibility and Efficiency of Information Processing

Open access
Da Costa, Lancelot, Parr, Thomas, Sengupta, Biswa, Friston, Karl · 2021 · Entropy

The DOI matches its Crossref record; 129 of the 130 references it cites were found in a registry or an index.

Abstract

Active inference is a normative framework for explaining behaviour under the free energy principle-a theory of self-organisation originating in neuroscience. It specifies neuronal dynamics for state-estimation in terms of a descent on (variational) free energy-a measure of the fit between an internal (generative) model and sensory observations. The free energy gradient is a prediction error-plausibly encoded in the average membrane potentials of neuronal populations. Conversely, the expected probability of a state can be expressed in terms of neuronal firing rates. We show that this is consistent with current models of neuronal dynamics and establish face validity by synthesising plausible electrophysiological responses. We then show that these neuronal dynamics approximate natural gradient descent, a well-known optimisation algorithm from information geometry that follows the steepest descent of the objective in information space. We compare the information length of belief updating in both schemes, a measure of the distance travelled in information space that has a direct interpretation in terms of metabolic cost. We show that neural dynamics under active inference are metabolically efficient and suggest that neural representations in biological agents may evolve by approximating steepest descent in information space towards the point of optimal inference.

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 · 10.3390/e23040454

Cites

129 of 130 references verified0 citation errors · 1 not checkable

Statistics

Figures

Figure 1 Active inference for state estimation (discrete state-space). Panel (a) summarises the problem of finding the minimum of a function (e.g., the free energy). One possibility would be taking the shortest path, which involves climbing up a hill; however, in the nescience of the minimum, a viable strategy consists of myopically taking the direction of steepest descent. In panel (b), we depict an example of a trajectory of an agent’s beliefs during the process of perception, which consists of updating beliefs about the states of the external world to reach the point of optimal inference (i.e., free energy minimum). In this example, the state-space comprises only three states (e.g., three different locations in a room). As they are probabilities over states, the components of s are non-negative and sum to one; hence, the agent’s beliefs naturally live on a triangle in three-dimensional space. Mathematically, this object is called a (two-dimensional) simplex. This constitutes the belief space, or information space, on which the free energy is defined. Technically, this object is a smooth statistical manifold, which corresponds to the set of parameters of a categorical distribution. To optimise metabolic and computational efficiency, agents must update their beliefs to reach the free energy minimum via the shortest possible path on this manifold. In panel (c), we exhibit simulated local field potentials that arise by interpreting the rate of change of v in terms of depolarisations, over a sequence of eight observations (e.g., saccadic eye-movements). As the rate of change is given by the free energy gradients, the decay of these local field potentials to zero coincides with the reaching of the free energy minimum (at which the gradient is zero by definition). Each colour expresses the mean voltage potential of a neural population, which encodes beliefs about one hidden state. These were obtained during the first eight trials of the T-Maze numerical simulation described in Section 6.1.1. Local field potentials accompanying the rule learning simulation—and for subsequent trials of the T-Maze task—can be found in [39,57], respectively. For more details on the generation of simulated electrophysiological responses, see [57].
Figure 1 Active inference for state estimation (discrete state-space). Panel (a) summarises the problem of finding the minimum of a function (e.g., the free energy). One possibility would be taking the shortest path, which involves climbing up a hill; however, in the nescience of the minimum, a viable strategy consists of myopically taking the direction of steepest descent. In panel (b), we depict an example of a trajectory of an agent’s beliefs during the process of perception, which consists of updating beliefs about the states of the external world to reach the point of optimal inference (i.e., free energy minimum). In this example, the state-space comprises only three states (e.g., three different locations in a room). As they are probabilities over states, the components of s are non-negative and sum to one; hence, the agent’s beliefs naturally live on a triangle in three-dimensional space. Mathematically, this object is called a (two-dimensional) simplex. This constitutes the belief space, or information space, on which the free energy is defined. Technically, this object is a smooth statistical manifold, which corresponds to the set of parameters of a categorical distribution. To optimise metabolic and computational efficiency, agents must update their beliefs to reach the free energy minimum via the shortest possible path on this manifold. In panel (c), we exhibit simulated local field potentials that arise by interpreting the rate of change of v in terms of depolarisations, over a sequence of eight observations (e.g., saccadic eye-movements). As the rate of change is given by the free energy gradients, the decay of these local field potentials to zero coincides with the reaching of the free energy minimum (at which the gradient is zero by definition). Each colour expresses the mean voltage potential of a neural population, which encodes beliefs about one hidden state. These were obtained during the first eight trials of the T-Maze numerical simulation described in Section 6.1.1. Local field potentials accompanying the rule learning simulation—and for subsequent trials of the T-Maze task—can be found in [39,57], respectively. For more details on the generation of simulated electrophysiological responses, see [57].CC BY · PubMed Central
Figure 2 Statistical manifolds and information length. Panels (a,b) illustrate the statistical manifolds associated with two well-known probability distributions: the normal distribution and the categorical distribution, respectively. The statistical manifold associated with a probability distribution is the set of all possible parameters that it can take. For the univariate normal distribution, parameterised with mean μ and positive standard deviation ς, the associated statistical manifold is the upper half plane (panel a). For the categorical distribution, in the case of three possible states, the statistical manifold is the 2-dimensional simplex (panel b). More generally, in the case of n possible states, the statistical manifold of the categorical distribution is the set of all vectors with positive components that sum to one, i.e., the (n − 1)-dimensional simplex. This is a higher-dimensional version of the triangle or the tetrahedron. In panels (c,d) we illustrate why the Euclidean distance is ill-suited to measure the information distance between probability distributions. To show this, we selected four distributions that correspond to points on the statistical manifold of the normal distribution. One can see that the Euclidean distance between the modes of the red and the blue distributions is the same as that from the orange and the green; however, the difference in information of each respective pair is quite different. In panel (c), the two distributions correspond to two drastically different beliefs, since there is such little overlap; on the contrary, the beliefs in panel (d) are much more similar. This calls for a different notion of distance that measures the difference in (Fisher) information between distributions, namely, the information length.
Figure 2 Statistical manifolds and information length. Panels (a,b) illustrate the statistical manifolds associated with two well-known probability distributions: the normal distribution and the categorical distribution, respectively. The statistical manifold associated with a probability distribution is the set of all possible parameters that it can take. For the univariate normal distribution, parameterised with mean μ and positive standard deviation ς, the associated statistical manifold is the upper half plane (panel a). For the categorical distribution, in the case of three possible states, the statistical manifold is the 2-dimensional simplex (panel b). More generally, in the case of n possible states, the statistical manifold of the categorical distribution is the set of all vectors with positive components that sum to one, i.e., the (n − 1)-dimensional simplex. This is a higher-dimensional version of the triangle or the tetrahedron. In panels (c,d) we illustrate why the Euclidean distance is ill-suited to measure the information distance between probability distributions. To show this, we selected four distributions that correspond to points on the statistical manifold of the normal distribution. One can see that the Euclidean distance between the modes of the red and the blue distributions is the same as that from the orange and the green; however, the difference in information of each respective pair is quite different. In panel (c), the two distributions correspond to two drastically different beliefs, since there is such little overlap; on the contrary, the beliefs in panel (d) are much more similar. This calls for a different notion of distance that measures the difference in (Fisher) information between distributions, namely, the information length.CC BY · PubMed Central
Figure 3 Generative model and T-Maze task. This figure describes the generative model used in the simulations and the specificities of the T-Maze task. The generative model is a partially observed Markov decision process (POMDP), a canonical model for sequential decision-making tasks under uncertainty. On the left-hand-side, the generative model is expressed as a factor graph [122]: the circles represent random variables while the squares represent factors. The shaded variables are observations, which the agent samples in discrete time. The unshaded variables are random variables that are inferred by the agent. These are either hidden states (s) or policies (π, i.e., action sequences), which need to be inferred from observations. The factors encode the relationships between the variables that they link. For example, D is a vector of non-negative components that sum to one that expresses the prior belief of being in this or another hidden state; B is a stochastic matrix that encodes the transition probabilities between hidden states given actions; similarly, A is a matrix that gives the probability of observations given states. See [7] for more details. Importantly, the factors can themselves be learned over time. On the right-hand side, we further unpack the generative model of the T-Maze task. The task is as pictured in the upper-right corner. The rat’s initial location is the centre of the T-Maze. At each time-step, the rat can choose between four control states (i.e., actions) that lead the rat to move to one of the four maze locations (middle, top-left, top-right or bottom). There are four hidden states that correspond to the location of the rat in the T-Maze and another two that correspond to the location of the reward (right arm or left arm), which leads to eight hidden states in total, meaning that the associated simplex on which the free energy is defined is seven-dimensional. There are seven possible observations that correspond to being in the middle of the T-Maze, in each one of the upper arms with (in red) or without the reward (in grey), and to collecting the cue, which reveals the location of the reward. The agent has a preference for each of these observations: the agent wants to claim the reward and avoid the unbaited arm. The agent is neutral w.r.t. all other observations. See [57] for more details of this paradigm.
Figure 3 Generative model and T-Maze task. This figure describes the generative model used in the simulations and the specificities of the T-Maze task. The generative model is a partially observed Markov decision process (POMDP), a canonical model for sequential decision-making tasks under uncertainty. On the left-hand-side, the generative model is expressed as a factor graph [122]: the circles represent random variables while the squares represent factors. The shaded variables are observations, which the agent samples in discrete time. The unshaded variables are random variables that are inferred by the agent. These are either hidden states (s) or policies (π, i.e., action sequences), which need to be inferred from observations. The factors encode the relationships between the variables that they link. For example, D is a vector of non-negative components that sum to one that expresses the prior belief of being in this or another hidden state; B is a stochastic matrix that encodes the transition probabilities between hidden states given actions; similarly, A is a matrix that gives the probability of observations given states. See [7] for more details. Importantly, the factors can themselves be learned over time. On the right-hand side, we further unpack the generative model of the T-Maze task. The task is as pictured in the upper-right corner. The rat’s initial location is the centre of the T-Maze. At each time-step, the rat can choose between four control states (i.e., actions) that lead the rat to move to one of the four maze locations (middle, top-left, top-right or bottom). There are four hidden states that correspond to the location of the rat in the T-Maze and another two that correspond to the location of the reward (right arm or left arm), which leads to eight hidden states in total, meaning that the associated simplex on which the free energy is defined is seven-dimensional. There are seven possible observations that correspond to being in the middle of the T-Maze, in each one of the upper arms with (in red) or without the reward (in grey), and to collecting the cue, which reveals the location of the reward. The agent has a preference for each of these observations: the agent wants to claim the reward and avoid the unbaited arm. The agent is neutral w.r.t. all other observations. See [57] for more details of this paradigm.CC BY · PubMed Central
Figure 4 Information length of perceptual inference in active inference and natural gradient. The upper-half of the figure plots the results of the numerical experiments with T-Maze task, while the lower half the abstract rule learning task. In Panel (a), the histograms show the information length of perceptual inference in active inference (in red) and natural gradient (in blue). Specifically, this is (mean = 7075, median = 2.9776) for active inference, and (mean = 2.7793, median = 3.1163) for the natural gradient; in the rule learning task, the information lengths are (mean = 14.3573, median = 16.2099) for active inference and (mean = 13.8588, median = 16.1058) for the natural gradient. Information trajectories are longer in the abstract rule learning task, due to its increased complexity. In other words, the greater number of hidden states—in the generative model—mandate more belief updating. This means the model used for this high-dimensional task entails greater complexity in the technical sense: as belief updating moves posterior beliefs further from their priors. (Parenthetically, this means that the agents took much longer to update their beliefs). The reasons for systematic variation in information length across trials within tasks is that (1) the task configuration varied from trial to trial and (2) the generative models were themselves learned over trials. The sudden increase in information length in the T-Maze task at the ninth trial occurs when agents learn that the reward’s location is constant. As the reward’s location is unambiguous, belief trajectories are longer as beliefs about context move from the middle of the simplex to the vertex representing the reward’s location at each step of the paradigm. This means that we can improve the biological plausibility of the implementation by starting each belief update at the previous free energy minimum instead of an agnostic location in the middle of the simplex. This would avoid beliefs travelling to the same vertex multiple times in the same trial and hence would make belief trajectories shorter once the reward location is learned. This consideration does not affect our characterisation, which simply compares the information length given consistent starting and end points. As the context could be deduced only at the last step of the task, a similar sudden change in information length did not occur in the abstract rule learning task. However, the slight increase in information length across trials occurs for the same reason: learning the generative model made inferences increasingly precise. In Panel (b), the violin plots illustrate the difference in information length by subtracting the information length of the natural gradient from the information length of active inference. In the T-Maze paradigm, active inference mostly takes shorter paths, which is why the violin plot’s values are mostly negative (mean= −0.0718, median= −0.1146). However, we obtain the opposite pattern in the abstract rule learning task (mean = 0.4985, median= 0.5401). Panel (c) shows an example of the belief trajectory taken during state estimation in the abstract rule learning task. This was possible to include as the generative model includes a hidden state dimension with three possible alternatives [39], enabling a two-dimensional representation of the associated simplex. The red trajectory corresponds to active inference, the blue is natural gradient descent, and the orange is the shortest path to the free energy minimum (i.e., the geodesic, see Appendix D). This example is not representative of the average and was chosen for purely illustrative purposes as the trajectories are very distinct, lengthy and do not coincide with the geodesic. The fact that both schemes were significantly suboptimal in this example was unsurprising as beliefs evolve to the free energy minimum using only local information about the free energy landscape. Note that this suboptimality occurred only in a small minority of the trials considered here.
Figure 4 Information length of perceptual inference in active inference and natural gradient. The upper-half of the figure plots the results of the numerical experiments with T-Maze task, while the lower half the abstract rule learning task. In Panel (a), the histograms show the information length of perceptual inference in active inference (in red) and natural gradient (in blue). Specifically, this is (mean = 7075, median = 2.9776) for active inference, and (mean = 2.7793, median = 3.1163) for the natural gradient; in the rule learning task, the information lengths are (mean = 14.3573, median = 16.2099) for active inference and (mean = 13.8588, median = 16.1058) for the natural gradient. Information trajectories are longer in the abstract rule learning task, due to its increased complexity. In other words, the greater number of hidden states—in the generative model—mandate more belief updating. This means the model used for this high-dimensional task entails greater complexity in the technical sense: as belief updating moves posterior beliefs further from their priors. (Parenthetically, this means that the agents took much longer to update their beliefs). The reasons for systematic variation in information length across trials within tasks is that (1) the task configuration varied from trial to trial and (2) the generative models were themselves learned over trials. The sudden increase in information length in the T-Maze task at the ninth trial occurs when agents learn that the reward’s location is constant. As the reward’s location is unambiguous, belief trajectories are longer as beliefs about context move from the middle of the simplex to the vertex representing the reward’s location at each step of the paradigm. This means that we can improve the biological plausibility of the implementation by starting each belief update at the previous free energy minimum instead of an agnostic location in the middle of the simplex. This would avoid beliefs travelling to the same vertex multiple times in the same trial and hence would make belief trajectories shorter once the reward location is learned. This consideration does not affect our characterisation, which simply compares the information length given consistent starting and end points. As the context could be deduced only at the last step of the task, a similar sudden change in information length did not occur in the abstract rule learning task. However, the slight increase in information length across trials occurs for the same reason: learning the generative model made inferences increasingly precise. In Panel (b), the violin plots illustrate the difference in information length by subtracting the information length of the natural gradient from the information length of active inference. In the T-Maze paradigm, active inference mostly takes shorter paths, which is why the violin plot’s values are mostly negative (mean= −0.0718, median= −0.1146). However, we obtain the opposite pattern in the abstract rule learning task (mean = 0.4985, median= 0.5401). Panel (c) shows an example of the belief trajectory taken during state estimation in the abstract rule learning task. This was possible to include as the generative model includes a hidden state dimension with three possible alternatives [39], enabling a two-dimensional representation of the associated simplex. The red trajectory corresponds to active inference, the blue is natural gradient descent, and the orange is the shortest path to the free energy minimum (i.e., the geodesic, see Appendix D). This example is not representative of the average and was chosen for purely illustrative purposes as the trajectories are very distinct, lengthy and do not coincide with the geodesic. The fact that both schemes were significantly suboptimal in this example was unsurprising as beliefs evolve to the free energy minimum using only local information about the free energy landscape. Note that this suboptimality occurred only in a small minority of the trials considered here.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 130 references

Research works printed but not linked: 4

  1. Da Costa, L., Sajid, N., Parr, T., Friston, K., and Smith, R. (2021, January 31). The Relationship between Dynamic Programming and Active Inference: The Discrete, Finite-Horizon Case. arXiv:200908111 [cs, math, q-bio]. Available online: http://arxiv.org/abs/2009.08111. · arXiv
  2. Barlow, H.B. (2021, April 09). Possible Principles Underlying the Transformations of Sensory Messages; The MIT Press: 1961. Available online: https://www.universitypressscholarship.com/view/10.7551/mitpress/9780262518420.001.0001/upso-9780262518420-chapter-13. · doi:10.7551/mitpress/9780262518420.001.0001/upso-9780262518420-chapter-13
  3. Bishop, C.M. (2006). Pattern Recognition and Machine Learning, Springer.
  4. Stone, J.V. (2019). Artificial Intelligence Engines: A Tutorial Introduction to the Mathematics of Deep Learning, Sebtel Press.

Cited by

See all 47

Metadata

Authors
Da Costa, L., Parr, T., Sengupta, B., & Friston, K.
Affiliations
  • Department of Mathematics, Imperial College London, London SW7 2AZ, UK
  • Wellcome Centre for Human Neuroimaging (GB)
Venue
Entropy
Publisher
MDPI AG
Published
2021-04-12
Volume
23
Issue
4
Pages
454
ISSN
1099-4300 (ISSN-L)
OpenAlex
W3126619093
Cited by
38
References
110
Citations by year
2026: 2 · 2025: 7 · 2024: 5 · 2023: 9 · 2022: 7 · 2021: 8
FWCI
3.81
Type
journal article
Language
en
Open access
Locations
  • Entropy (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Funders
  • Fonds National de la Recherche Luxembourg (13568875)
  • Wellcome Trust (088130/Z/09/Z)
  • Rosetrees Trust (173346)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Embodied and Extended Cognition
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Cognitive Science and Education Research
Keywords
active inference, free energy principle, neural dynamics, prediction error, natural gradient descent, information geometry, information length, metabolic cost, belief updating, generative models, electrophysiological responses, neural representations, state estimation

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: Neural Dynamics under Active Inference: Plausibility and Efficiency of Information Processing. https://app.cochat.ai/papers/10.3390/e23040454
The paper
Da Costa, L., Parr, T., Sengupta, B., & Friston, K. (2021). Neural Dynamics under Active Inference: Plausibility and Efficiency of Information Processing. Entropy. https://doi.org/10.3390/e23040454

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.