Active inference casts the brain as a system that maintains an internal generative model of the world and minimizes variational free energy through perception, action, and learning. Neuroscience has been the primary proving ground for this framework since Karl Friston's 2010 "unified brain theory" paper, with a concentrated body of theoretical, simulation, and early clinical work testing how far the idea can be pushed. This page surveys what that literature actually shows, the modeling patterns and tools researchers use, and the open questions that still separate active inference from a validated neuroscientific theory.
Why the domain fits
Active inference treats perception as inference about hidden causes of sensory data, action as the selective sampling of observations that fulfill prior expectations, and learning as the slow adaptation of a generative model's parameters — a triad that maps directly onto core neuroscientific concerns. The brain's hierarchical cortical organization, with higher areas encoding slower, more abstract causes and lower areas encoding fast sensory detail, is a natural substrate for the hierarchical generative models the framework requires. Active inference also extends predictive-coding accounts of perception by explicitly modeling actions and policy selection as part of the same inferential loop, coupling perception and control through a single objective — expected free energy minimization.
State of the literature
The literature is anchored by a relatively concentrated set of foundational and review papers — Friston's 2010 Nature Reviews Neuroscience article, work formalizing variational free energy for perception and action, and later expository reviews aimed at making the principle's assumptions and scope explicit to neuroscientists. Simulation-based applications extend from there into deep temporal models of epistemic behavior (e.g., reading), oculomotion, working memory and attention, dopaminergic neuromodulation, and the emergence of habits from repeated policy optimization. Discrete-state formulations, cast as partially observable Markov decision processes, have been particularly influential for modeling cognitive tasks such as visual search and goal-directed planning. Most of this work originates from a core group centered on Karl Friston at University College London and the Wellcome Centre for Human Neuroimaging, with Thomas Parr a key collaborator on discrete-state formalization, and is increasingly extending into computational psychiatry and morphogenesis.
Application patterns and tools
The clearest neurophysiologically grounded demonstration is oculomotion: simulations using Bayesian filtering to implement planning as inference generate saccadic and smooth pursuit eye movements whose message-passing structure maps onto known brainstem and cerebellar connectivity and resembles single-unit recordings from relevant nuclei. Dopamine has been modeled as encoding precision (confidence) over policies, with tonic dopamine levels in simulation shaping exploration versus exploitation and movement vigor, and disruptions offered as a lens on Parkinson's disease and schizophrenia. Implementation has largely relied on the MATLAB-based SPM ecosystem and its dynamic expectation maximization (DEM) module for continuous-state models, and Python libraries for discrete-state POMDP-style models used in cognitive-task simulations — tools that remain research code rather than standardized, widely distributed software.
Open problems
The report identifies biological plausibility as unresolved: network models that approximate gradient descent on free energy reproduce oscillations and attractor dynamics, but the exact biophysical mechanisms in real neural tissue, and mappings like superficial pyramidal cells encoding prediction errors, remain hypothetical and largely untested. Scaling generative models beyond small state spaces to naturalistic, high-dimensional environments is a second open computational challenge. A third is empirical falsifiability — critics argue the free energy principle is broad enough to resist disconfirmation, and the report calls for experiments that discriminate active inference's predictions (epistemic exploration, precision-weighted neuromodulation) from reinforcement learning and predictive coding alternatives. Clinical translation in computational psychiatry is likewise described as promising but early-stage, resting mainly on theoretical and simulation work rather than large validated patient datasets.
Threads into philosophy: neurophenomenology
Active inference's account of neural function also connects directly to philosophy of mind through neurophenomenology, a field working to build a dialogue between measurable neuronal processes and philosophical descriptions of subjective experience. Lars Sandved-Smith, Casper Hesp, Jérémie Mattout, Karl Friston, Antoine Lutz, and Maxwell J. D. Ramstead (2021) took up this connection directly, modeling meta-awareness and attentional control with deep parametric active inference in Neuroscience of Consciousness (DOI: 10.1093/nc/niab018) — treating the same hierarchical generative models built to explain cortical dynamics as candidate models of what it is like to notice, and to be aware of noticing.
Threads into physics: the frame problem
The neuroscientific modeling surveyed above already assumes a frame in which concepts such as neuron, circuit, and prediction error are meaningful units to model over — a physics-level question about how any bounded system settles on the states and boundaries it treats as real before inference over them can even begin. Chris Fields, an Institute Scientific Advisor who has led its physics course and livestream series, has taken up this frame problem directly, including in an Identity Operator presentation (https://www.youtube.com/watch?v=RwDMP0qeyHo; also linked as https://www.youtube.com/live/RwDMP0qeyHo).
Reference Backbone
Karl J. Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. DOI: 10.1038/nrn2787. Christopher L. Buckley, Chang Sub Kim, Simon McGregor, Anil K. Seth (2017). The free energy principle for action and perception: A mathematical review. Journal of Mathematical Psychology. DOI: 10.1016/j.jmp.2017.09.004. Thomas Parr, Giovanni Pezzulo, Karl J. Friston (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press. Lancelot Da Costa, Thomas Parr, Noor Sajid, Sebastijan Veselic, Victorita Neacsu, Karl J. Friston (2020). Active inference on discrete state-spaces: A synthesis. Journal of Mathematical Psychology. DOI: 10.1016/j.jmp.2020.102447. André Bastos, W. Martin Usrey, Rick A. Adams, George R. Mangun, Pascal Fries, Karl J. Friston (2012). Canonical Microcircuits for Predictive Coding. Neuron. https://www.cell.com/neuron/fulltext/S0896-6273(12)00959-2. Takuya Isomura, Hideaki Shimazaki, Karl J. Friston (2022). Canonical neural networks perform active inference. Communications Biology. DOI: 10.1038/s42003-021-02994-2. Takuya Isomura, Kiyoshi Kotani, Yasuhiko Jimbo, Karl J. Friston (2023). Experimental validation of the free-energy principle with in vitro neural networks. Nature Communications. DOI: 10.1038/s41467-023-40141-z. Michael Levin (2023). Bioelectric networks: the cognitive glue enabling evolutionary scaling from physiology to mind. Animal Cognition. DOI: 10.1007/s10071-023-01780-3. Christopher J. Whyte, Andrew W. Corcoran, Jonathan Robinson, Ryan Smith, Rosalyn J. Moran, Thomas Parr, Karl J. Friston, Anil K. Seth, Jakob Hohwy (2024). On the Minimal Theory of Consciousness Implicit in Active Inference. arXiv.