paper
active
2020
ⓒ 225
paper:ramstead-2020-two-densities

A tale of two densities: active inference is enactive inference

ByMaxwell J. D. Ramsteadⓘ·Michael D. Kirchhoffⓘ·Karl J. Friston
DOI 10.1177/1059712319862774OpenAlex W2953421850
Active inference & free energy principleActive inference & free energy principle"A tale of two densities, as it were."

TL;DR

Ramstead, Kirchhoff, and Friston argue that generative models in active inference under the free energy principle (FEP) are control systems—not structural representations—and that this distinction has been systematically obscured by conflating active inference with brain-centered Bayesian frameworks such as predictive coding (Rao & Ballard, 1999), the Helmholtz machine (Dayan et al., 1995), and Bishop's (2006) variational machine learning. The load-bearing move is a technical one: under the FEP, generative models are entailed by the adaptive dynamics of an organism rather than encoded in physical neural states, while it is the recognition density that is embodied—parameterized by the sufficient statistics of internal Markov blanket states. The paper introduces the construct of enactive inference, a reinterpretation grounding the generative model as a normative control system in the tradition of Conant and Ross Ashby's (1970) good regulator theorem, distinguishing it sharply from the structural representationalist accounts advanced by Kiefer and Hohwy (2018, 2019) and Gładziejewski and Miłkowski (2017). Representationalists are correct that internal states encode exploitable structural similarities, but they misidentify the vehicle: those states parameterize the recognition density, not the generative model. The paper argues this implies that perception and action are inseparable moments of a single policy-selection process, that cognitive science should shift from asking how brains represent the world to how organisms enact attunement to their ecological niche, and that enactivism and the mathematical apparatus of active inference are mutually reinforcing rather than in tension.

What to take away

  1. 1. Under the free energy principle, generative models are entailed by an organism's adaptive behavior and are not encoded as physical neural states, making them categorically distinct from the recognition density, which is parameterized by internal Markov blanket states.
  2. 2. Structural representationalist accounts of generative models—most prominently defended by Kiefer and Hohwy (2018, 2019) and Gładziejewski and Miłkowski (2017)—are based on an older Bayesian framework (the Helmholtz machine, Dayan et al., 1995) rather than on active inference under the FEP, constituting a category error.
  3. 3. The paper introduces 'enactive inference' as a formal reinterpretation of active inference wherein the generative model functions as a normative control system (in the sense of Conant and Ross Ashby's 1970 good regulator theorem) and the recognition density is what the organism literally embodies.
  4. 4. Variational free energy F(Q) ≥ −ln P(s) is an upper bound on negative log evidence (surprise), and the paper argues that any system maintaining a Markov blanket necessarily implements a gradient descent on this quantity, licensing an inferential interpretation of all adaptive dynamics.
  5. 5. In active inference, 'perceptual inference' is not a parallel process to action but merely the state-estimation moment within the broader policy-selection process, meaning perception is constitutively action-dependent rather than action being an output of perception.
  6. 6. The recognition density—not the generative model—is the locus of exploitable structural similarity with the environment, because it is the recognition density whose sufficient statistics (expectations and precisions) are physically realized in internal states and updated through active inference.
  7. 7. A replication-ready methodology in this paper is the use of Forney-style factor graphs and Bayesian network diagrams (Figures 2 and 3) to formally represent the generative model for policy selection, explicitly separating the quantities that are updated (recognition density parameters) from those that define the optimization target (the generative model).
  8. 8. The paper raises the open question of whether the enactive inference framework can be extended to social and cultural cognition—citing Veissière, Constant, Ramstead, Friston, and Kirmayer (2020) on variational approaches to cognition and culture—as a domain where the control-system reading of generative models may need further elaboration.
  9. 9. Under active inference, the generative model is typically more deeply structured than the generative process describing the environment because agents author their own sensory samples through action—a point illustrated by the case of voluntary movement, where no external structure corresponds to articulated hand trajectories until the organism enacts them.
  10. 10. The paper predicts that properly distinguishing entailed generative models from embodied recognition densities will reconcile representationalist and enactivist research programs, because representationalists are right that internal states encode action-guiding structural information, while enactivists are right that this encoding is established and maintained through adaptive action rather than passive mirroring.

Peer brief — for seminar discussion

Ramstead, Kirchhoff, and Friston (2020, Adaptive Behavior) take aim at a specific interpretive claim that has become widespread in Bayesian cognitive science: that the generative models posited by the free energy principle (FEP) are structural representations—internal neural structures that carry representational content in virtue of an exploitable structural similarity to the environment, as argued most systematically by Kiefer and Hohwy (2018, 2019) and Gładziejewski and Miłkowski (2017). The paper's intervention is to show that this claim rests on importing assumptions from brain-centered Bayesian frameworks—predictive coding (Rao & Ballard, 1999), the Helmholtz machine (Dayan et al., 1995), and Bishop's (2006) variational machine learning—into active inference under the FEP, where those assumptions do not apply. The load-bearing finding is a technical distinction between two probability densities: the generative model, which is entailed by the organism's adaptive dynamics but not encoded in any physical neural state, and the recognition density, which is embodied in the sufficient statistics (expectations and precisions) of internal Markov blanket states. Because the generative model has no physically realized sufficient statistics of its own, it cannot function as a structural representation; it functions instead as a normative control system, in the precise sense of Conant and Ross Ashby's (1970) good regulator theorem. The paper introduces the construct of enactive inference to name this reinterpretation, arguing that active inference is enactive because generative models guide action without themselves being representational vehicles—they are enacted through the organism's behavior rather than encoded in it. The method used is close formal analysis of the active inference mathematical apparatus, including Markov blanket partitions, variational free energy F(Q) ≥ −ln P(s), and Forney-style factor graphs; an alternative approach the authors could have taken would be computational simulation of active inference agents to empirically track the formal separation between generative model entailment and recognition density updating. The core implication is that perception and action cannot be dissociated: active inference is not Bayesian perceptual inference with an action output appended, but a unified policy-selection process in which perception is one moment. The paper predicts this reframing will reconcile enactivism and representationalism: internal states do encode exploitable structural similarities (conceding the representationalist point), but these are properties of the recognition density, established and maintained through active inference, not of the generative model. A critical reader should push back on the scope of the enactive inference claim: the formal argument establishes that the generative model has no physically realized sufficient statistics within the standard FEP formalism, but the paper does not demonstrate that no consistent extension of the formalism could assign the generative model a representational role—it is possible that richer implementations, such as hierarchical deep generative models with explicit architectural priors, would blur the boundary the authors draw, and the paper does not engage seriously with this possibility. The paper also relies heavily on the Conant-Ashby good regulator theorem as philosophical support, but that theorem concerns deterministic control systems, and its application to stochastic, nonequilibrium biological agents is asserted rather than formally derived.

Claims (4)

Questions (1)

Original abstract (expand)

The aim of this article is to clarify how best to interpret some of the central constructs that underwrite the free-energy principle (FEP) – and its corollary, active inference – in theoretical neuroscience and biology: namely, the role that generative models and variational densities play in this theory. We argue that these constructs have been systematically misrepresented in the literature, because of the conflation between the FEP and active inference, on the one hand, and distinct (albeit closely related) Bayesian formulations, centred on the brain – variously known as predictive processing, predictive coding or the prediction error minimisation framework. More specifically, we examine two contrasting interpretations of these models: a structural representationalist interpretation and an enactive interpretation. We argue that the structural representationalist interpretation of generative and recognition models does not do justice to the role that these constructs play in active inference under the FEP. We propose an enactive interpretation of active inference – what might be called enactive inference. In active inference under the FEP, the generative and recognition models are best cast as realising inference and control – the self-organising, belief-guided selection of action policies – and do not have the properties ascribed by structural representationalists.

Related work— refs + corpus + external arXiv

Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.

+18 more

Similar preprints — Semantic Scholar

Cited by (1)

Cross-corpus bridges (1)

same_concept_as · Nomic cosine

External markdown files that talk about the same concept as this entity.

  • aboutblank_kb
    Active Inferenceframeworks/active-inference.md0.858