Karl Friston – 1959 Present

Biography

The history of modern neuroscience is frequently chronicled through the development of its visual and conceptual instruments. Just as the compound microscope enabled Santiago Ramón y Cajal to discern the individual architecture of the neuron, the late twentieth-century confluence of medical physics, computer science, and statistical theory transformed human neuroimaging from a descriptive cartographic exercise into a rigorous, mechanistic science. At the very center of this conceptual and methodological revolution stands Karl John Friston (born July 12, 1959), a British neuroscientist, biophysicist, and psychiatrist whose intellectual output has systematically reshaped contemporary understandings of the human brain. Working out of the Wellcome Centre for Human Neuroimaging at University College London, Friston has achieved a level of academic influence and citation impact that places him among the most prolific and consequential scientists in recorded history.

Friston’s scientific trajectory is marked by a rare and audacious trajectory of scale: beginning with the acute, qualitative phenomenological dilemmas of severe clinical psychiatry, he pioneered the foundational mathematics of functional neuroimaging, progressed into biophysical models of directed neuronal communication, and ultimately synthesized these diverse domains into an overarching, first-principles framework known as the Free Energy Principle (FEP) and its operational corollary, active inference. What began as a practical effort to distinguish genuine hemodynamic brain activation from background stochastic noise has culminated in a universal mathematical physics of self-organizing, adaptive, and sentient systems, bridging non-equilibrium thermodynamics, information geometry, Bayesian statistics, and philosophy of mind.

This comprehensive intellectual biography charts the evolution of Friston’s thought across five decades. From his early training in Cambridge natural philosophy and bedside psychiatric rotations at the Maudsley Hospital, to the development of Statistical Parametric Mapping (SPM), Dynamic Causal Modelling (DCM), hierarchical predictive coding, and the formal topology of Markov blankets, this treatise examines the mathematical structures, biophysical mechanisms, and philosophical implications that define his oeuvre. In tracing this intellectual odyssey, we illuminate not merely the biography of a singular thinker, but the emergence of a unified, computational paradigm for understanding life, cognition, and the physical architecture of consciousness itself.

1. Early Life, Academic Formations, and Clinical Foundations (1959–1980s)

1.1 Formative Years and Natural Philosophy Background

Karl John Friston was born on July 12, 1959, in York, England, entering a post-war British society undergoing rapid technological, industrial, and social transformation. Raised in an intellectually encouraging domestic environment, Friston exhibited an early and acute sensitivity to the lawful regularities of the physical world. His childhood curiosity was fundamentally systemic; rather than engaging with natural phenomena as isolated biological specimens, he demonstrated a precocious fascination with how mechanical, mathematical, and dynamical principles governed living forms. Family lore and personal recollection recount an adolescent immersion in classical physics, geometry, and the differential equations of dynamic physical systems, prefiguring an intellectual orientation that would later refuse to acknowledge any rigid ontological boundary between the inanimate physics of matter and the animate biology of mind.

Friston matriculated at Gonville and Caius College, Cambridge, where he read Natural Sciences. The Cambridge natural science tripos at that time remained deeply steeped in the legacy of British natural philosophy, demanding a rigorous cross-disciplinary literacy spanning mathematics, theoretical physics, physiology, and evolutionary biology. This pedagogical environment proved profoundly formative. Rather than adopting an exclusively reductionist view of biological phenomena, Friston assimilated classical Newtonian mechanics, thermodynamics, and dynamical systems theory alongside the complex wetware of neurophysiology. During these undergraduate years, he absorbed the fundamental principles of homeostatic regulation, cybernetics, and physical equilibria, contemplating how open thermodynamic systems manage to preserve their structural and functional integrity against the universal drift toward thermodynamic dissipation.

The intellectual milieu of Cambridge introduced Friston to the mathematical formalisms that would later underpin his revolutionary work in quantitative neurobiology. He was particularly drawn to the capacity of differential calculus and linear algebra to capture the temporal trajectories of dynamic, interacting networks. By synthesizing the rigorous mechanical determinism of physics with the non-linear, stochastic realities of living organisms, Friston laid the conceptual foundations for what would later become a lifelong campaign: the reduction of subjective cognitive experience and physiological organization to formal, principled physical equations.

1.2 Medical Qualification and Psychiatric Specialization

Following his undergraduate scientific training at Cambridge, Friston pursued clinical medicine, completing his primary medical qualification at King’s College Hospital Medical School in London in 1983. Clinical training introduced a profound pragmatic shift. The abstract, elegant regularities of mathematical physics and neurophysiology were suddenly juxtaposed with the visceral, chaotic suffering of the hospital ward. Confronted with the somatic vulnerabilities of the human organism, Friston developed an abiding clinical discipline, grounding his mathematical proclivities in the diagnostic imperative to understand pathological failure modes of human physiology.

Drawn to the profound explanatory mysteries of human consciousness and behavioral disintegration, Friston undertook specialist residency training in psychiatry at the renowned Maudsley Hospital in South London. The Maudsley was historically celebrated as the intellectual epicenter of academic British psychiatry, characterized by an empirical, phenomenologically rigorous approach to mental illness. During his rotations, Friston was immersed in the direct management of severe, treatment-refractory psychopathology, encountering patients afflicted by chronic schizophrenia, profound bipolar affective disorders, and severe neurological catatonia. The bedside confrontation with clinical psychosis—marked by systemic delusions, auditory-verbal hallucinations, and the radical dissolution of the coherent self—left an indelible mark on his intellectual development.

It was within the wards of the Maudsley that Friston experienced a pivotal epistemological crisis regarding clinical psychiatry. He grew acutely frustrated by the field’s profound lack of an objective, quantitative, and mechanistic framework. While clinical psychiatry excelled at qualitative taxonomy—categorizing subjective distress into descriptive diagnostic manuals such as the DSM and ICD—it lacked an underlying computational or physical theory capable of explaining why a patient should experience a hallucination, or how neural circuitry could systematically miscalculate reality. Diagnostic boundaries appeared fundamentally arbitrary, reliant on statistical consensus rather than formal biophysical principles. This dissatisfaction ignited in Friston a passionate determination to uncover the quantitative, computational pathophysiology of disordered thought, transforming him from a bedside clinician into an analytical researcher.

1.3 Transition from Bedside Psychiatry to Quantitative Neuroimaging

Friston’s clinical frustration coincided with a technological watershed: the dawn of non-invasive human functional neuroimaging. In the late 1980s, the Medical Research Council (MRC) Cyclotron Unit at Hammersmith Hospital in London was emerging as a premier center for Positron Emission Tomography (PET). Recognizing that the living human brain could finally be interrogated in real time, Friston transitioned from full-time clinical psychiatry to a research fellowship at Hammersmith under the mentorship of Richard Frackowiak. The Hammersmith Cyclotron Unit was a hive of physicists, radiochemists, computer scientists, and neurologists, providing the ideal sandbox for Friston’s multidisciplinary genius.

At the MRC Cyclotron Unit, PET was being deployed to measure regional cerebral blood flow ($rCBF$) as an indirect surrogate for localized neuronal activity. However, the raw data produced by these early scanners was profoundly noisy, spatially distorted, and conceptually opaque. Investigators were essentially confronted with three-dimensional arrays of radioactive counts subject to immense inter-individual anatomical variation, head movement artifacts, and Poisson measurement noise. The dominant methods of analysis were rudimentary: researchers manually traced regions of interest (ROIs) on coarse low-resolution anatomical scans, a process that was notoriously subjective, non-reproducible, and blinded to distributed network dynamics.

Friston immediately identified an opportunity to apply his Cambridge-derived mathematical training. He recognized that the primary challenge of neuroimaging was not merely radiological, but fundamentally statistical and computational. Working alongside colleagues at Hammersmith, he formulated early, revolutionary hypotheses regarding the nature of schizophrenia, positing that the disorder was not an isolated structural lesion, but rather a functional “disconnection syndrome”—an aberrant, desynchronized pattern of dynamic functional connectivity among anatomically distributed brain networks. To demonstrate this empirical claim, however, the field desperately required an objective, voxel-by-voxel mathematical architecture that could map functional physiological changes across the entire living brain without subjective anatomical bias. This imperative inaugurated the first phase of Friston’s scientific revolution: the birth of modern neuroimaging methodology.

2. The Methodological Revolution: Statistical Parametric Mapping (SPM)

2.1 Conceptual Architecture of Statistical Parametric Mapping

Prior to the late 1980s and early 1990s, analyzing human brain activation was a fractured, idiosyncratic craft. Every neuroimaging laboratory maintained its own ad-hoc heuristics for identifying activations, rendering cross-study meta-analyses and universal reproducibility practically impossible. Karl Friston confronted this crisis by systematically designing a comprehensive, voxel-by-voxel computational framework: Statistical Parametric Mapping (SPM). Rather than restricting analysis to predetermined anatomical regions, Friston envisioned an objective, automated pipeline capable of interrogating every elementary three-dimensional volume element (voxel) within the brain uniformly, objectively, and simultaneously.

The conceptual architecture of SPM rested upon three methodological pillars: spatial normalisation, the mass-univariate application of the General Linear Model (GLM), and spatial smoothing. Spatial normalisation addressed the immense morphological variability of human craniums and cortical folding. Friston developed non-linear spatial transformation algorithms to register an individual subject’s brain scan into a standardized stereotaxic coordinate space—initially defined by the atlas of Jean Talairach and Gabor Tournoux, and subsequently refined into the standardized templates of the Montreal Neurological Institute (MNI). By warping distinct brains into a shared coordinate space, researchers could perform cross-subject averaging and group-level statistical comparisons.

Once spatially normalized and filtered with an isotropic Gaussian kernel (which smoothed the data to improve the signal-to-noise ratio and conform to statistical distribution assumptions), the data were subjected to a mass-univariate statistical analysis. Friston recognized that the classical General Linear Model—which subsumed multiple linear regression, analysis of variance (ANOVA), and t-tests—could be deployed at every single voxel independently:
$$Y = X\beta + \epsilon$$
Here, $Y$ represented the time series of observed hemodynamic signals at a given voxel, $X$ denoted the design matrix encoding experimental conditions, stimulus onsets, and physiological nuisance regressors, $\beta$ represented the unknown parameter estimates of regional response magnitude, and $epsilon$ represented identically and independently distributed Gaussian error terms. By calculating statistical contrasts across the estimated parameter weights, researchers produced a continuous three-dimensional lattice of test statistics: a Statistical Parametric Map (such as an SPM{$t$} or SPM{$F$}). In a move of visionary academic generosity, Friston and his team released SPM as open-source code via MATLAB, democratizing computational neuroimaging globally and transforming it into the lingua franca of cognitive neuroscience.

2.2 Mathematical Precision via Gaussian Random Field Theory

While the mass-univariate GLM provided an elegant means of estimating experimental effects across the brain, it instantly introduced an astronomical statistical dilemma: the massive multiple comparisons problem. A typical volumetric fMRI scan contains anywhere from 100,000 to over 1,000,000 voxels. If an investigator adopted a standard uncorrected alpha threshold of $p < 0.05$, a completely null scan containing pure white noise would yield tens of thousands of false-positive activations purely by chance. Traditional Bonferroni correction, which divides the significance threshold by the total number of tests, was devastatingly conservative; because adjacent voxels in biological tissue are strongly correlated due to spatial smoothing and intrinsic hemodynamic autocorrelation, treating each voxel as an independent statistical hypothesis ignored the spatial continuity of brain signals and eliminated true biological signal.

Friston resolved this monumental challenge through a groundbreaking collaboration with the Canadian mathematical statistician Keith Worsley, introducing Gaussian Random Field Theory (GRFT) to the neuroimaging domain. GRFT shifted the unit of statistical inference from discrete, isolated voxels to continuous, smooth topological random surfaces. Instead of asking how many individual voxels surpassed a statistical threshold, GRFT treated the thresholded statistical map as a continuous topological excursion set above a critical value, asking what the probability was of observing an excursion set of a given spatial extent or peak height under the null hypothesis of a smooth stationary Gaussian field.

The mathematical breakthrough relied on calculating the Euler characteristic ($\chi$) of the thresholded random field. At high thresholds, the expected value of the Euler characteristic corresponds precisely to the number of disconnected clusters of activation surviving the threshold:
$$\mathbb{E}[\chi] = \sum_{d=0}^{D} R_d(V) \rho_d(u)$$
where $R_d(V)$ represents the resel (resolution element) counts reflecting the volume and spatial smoothness of the search space $V$, and $\rho_d(u)$ denotes the Euler characteristic density for dimension $d$ at statistical threshold $u$. This mathematical formalization permitted rigorous family-wise error (FWE) rate correction across three distinct levels of topological inference:

• Voxel-level inference: Assessing the probability of finding a peak of an observed height anywhere within the search volume.
• Cluster-level inference: Evaluating whether the contiguous spatial extent of voxels exceeding a primary threshold is larger than expected by chance.
• Set-level inference: Testing the global number of distinct clusters across the entire brain space.

This elegant solution validated the transition from PET to Functional Magnetic Resonance Imaging (fMRI), giving the global neuroscientific community a rigorous, mathematically unimpeachable basis for mapping human cognitive architecture.

2.3 Voxel-Based Morphometry (VBM) and Structural Analysis

Having established the dominance of SPM in analyzing dynamic functional neuroimaging data, Friston, working alongside his colleague John Ashburner, realized that the identical computational logic could be extended to the structural analysis of brain morphology. Historically, macroscopic brain morphometry relied on manual planimetry and subjective volumetric measurement of selected brain regions on high-resolution T1-weighted structural MRI scans. This approach was labor-intensive, suffered from low inter-rater reliability, and remained limited to coarse structures with visually distinct anatomical boundaries, such as the hippocampus or lateral ventricles.

In response, Ashburner and Friston developed Voxel-Based Morphometry (VBM). VBM completely automated the structural analysis of tissue across the whole brain. The algorithm operated by segmenting high-resolution structural MRI scans into probabilistic tissue classes: grey matter, white matter, and cerebrospinal fluid (CSF). These probabilistic tissue segmentations were then non-linearly registered into stereotaxic space. Crucially, a subsequent innovation—the “modulation” step—multiplied the spatially normalized tissue concentration by the Jacobian determinants of the deformation fields, preserving the absolute regional volume of tissue that had been stretched or compressed during spatial registration.

Once smoothed, these voxel-wise structural maps were subjected to the standard mass-univariate GLM framework, enabling investigators to identify microscopic and macroscopic structural differences between healthy cohorts and clinical populations without any subjective tracing. VBM sparked a revolution across cognitive neurology, psychiatric genetics, and developmental cognitive neuroscience. The technique became famous through iconic investigations, such as Eleanor Maguire’s study of London taxi drivers, which documented localized neuroplastic changes in the posterior hippocampus as a function of environmental navigational experience. VBM transformed the global understanding of neuroplasticity, healthy human aging, and neurodegenerative disease progression, cementing Friston’s status as the most cited and methodologically influential neuroscientist on the planet.

3. From Mapping to Mechanisms: Dynamic Causal Modelling (DCM)

3.1 Functional Connectivity Versus Effective Connectivity

By the turn of the twenty-first century, Statistical Parametric Mapping had established human functional brain mapping as an internationally recognized empirical discipline. However, Karl Friston became deeply cognizant of a profound epistemological ceiling inherent in functional segregation and functional localization paradigms. Identifying which isolated cortical territory exhibited a hemodynamic fluctuation during a cognitive task—what critics occasionally disparaged as “neophrenology”—provided virtually zero insight into the underlying biophysical mechanisms through which distributed populations of neurons communicate, integrate information, and coordinate collective computational states.

Friston addressed this limitation by formally defining the critical taxonomic distinction between functional connectivity and effective connectivity. Functional connectivity, he argued, was fundamentally descriptive and statistical: it quantified the simple temporal correlation or covariance between spatially remote neurophysiological events. While correlations provided valuable macroscopic summaries of synchronized networks (such as the default mode network), they were essentially agnostic to the underlying mechanism. Correlational metrics could not determine the direction of information flow, could be driven entirely by unmeasured third-party confounding sources, and could not differentiate direct synaptic pathways from indirect, polysynaptic cascades.

In contrast, Friston defined effective connectivity as the directed, causal influence that one neuronal system exerts over another. Moving to effective connectivity required a decisive shift from passive observational phenomenology to formal biophysical hypothesis testing. The objective was no longer to ask “which areas covary?”, but rather to construct explicit, mathematically parameterized models of hidden neuronal populations and test how specific psychological perturbations, pharmacological challenges, or task demands dynamically modulate the directed synaptic gains between specific nodes of a cortical circuit.

3.2 Biophysical and State-Space Mathematical Modeling

To mathematically operationalize effective connectivity, Friston, alongside colleagues such as Lee Harrison and Will Penny, introduced Dynamic Causal Modelling (DCM) in 2003. DCM departed radically from traditional regression-based neuroimaging methodologies by establishing a continuous-time, non-linear state-space architecture. The core conceptual insight of DCM was the explicit mathematical partition of the neuroimaging problem into two distinct levels:

1. The Hidden Neuronal State Model: A set of differential equations describing the dynamic interactions among latent, unobserved neurophysiological states.
2. The Biophysical Forward Observation Model: A mathematically validated biophysical model mapping hidden neuronal activity to measurable macroscopic signals (e.g., the BOLD contrast in fMRI or electromagnetic fields in M/EEG).

In its classical bilinear fMRI formulation, the neuronal state equation governs the rate of change of hidden neuronal states $z$:
$$\frac{dz}{dt} = \left( A + \sum_{j} u_j B^{(j)} \right) z + C u$$
where $z$ represents the vector of neuronal activities across modeled brain regions, $u$ is the vector of external experimental inputs, the matrix $A$ encodes the baseline endogenous, directed connectivity between regions in the absence of experimental perturbation, the matrices $B^{(j)}$ represent the modulatory effects of experimental condition $u_j$ on directed synaptic connections (capturing contextual changes in synaptic gain), and the matrix $C$ models the direct, driving influence of sensory stimuli on specific neuronal nodes.

This latent neuronal trajectory is subsequently transformed into the observed hemodynamic signal via a forward biophysical model, integrating the Balloon-Windkessel model of Richard Buxton and colleagues. This forward model explicitly accounts for regional neurovascular coupling, cerebral blood flow dynamics, venous ballooning volume changes, and deoxyhemoglobin content. Because the parameters governing this non-linear biophysical cascade are latent and unobservable directly, Friston deployed a fully Bayesian inversion framework. Utilizing the Variational Laplace approximation, DCM computes the posterior probability distributions over the biophysical connectivity parameters alongside the variational free energy as an approximation of the marginal likelihood (model evidence). Through Bayesian Model Selection (BMS), researchers could rigorously compare competing mechanistic wiring diagrams of the human brain, fundamentally advancing cognitive neuroscience from correlational mapping to mechanistic causal inference.

3.3 Neural Mass and Neural Field Frameworks

Recognizing the temporal sluggishness of the hemodynamic BOLD response, Friston rapidly generalized Dynamic Causal Modelling from fMRI to the millisecond temporal resolution of electroencephalography (EEG), magnetoencephalography (MEG), and local field potentials (LFPs). Translating DCM to electrophysiology required abandoning phenomenological bilinear equations in favor of biologically grounded neural mass and neural field models.

Drawing on the classical traditions of Walter Freeman, Frank Lopes da Silva, and Hugh Wilson and Jack Cowan, Friston formulated DCM for electrophysiology based on canonical microcircuits. These biophysical models divided a cortical column into discrete laminar populations of neurons: deep and superficial excitatory pyramidal cells, spiny stellate cells in granular layer IV, and inhibitory GABAergic interneurons. The dynamics of each laminar subpopulation were mathematically modeled via coupled second-order differential equations transforming average presynaptic firing rates into postsynaptic membrane potentials via convolution operators, followed by static non-linear sigmoid functions converting membrane potentials into population firing rates.

This extension enabled DCM to model spectral neurodynamics, event-related potentials (ERPs), steady-state responses, and cross-frequency oscillatory coupling. Crucially, it empowered neuroscientists to test hypotheses regarding microscopic synaptic physiology non-invasively in the living human subject. By fitting neural mass models to electrophysiological recordings, investigators could infer pharmacological alterations in specific neurotransmitter receptor conductances, such as AMPA, NMDA, and GABA receptors. Friston’s neural mass modelling bridges the macroscopic scale of clinical non-invasive electrophysiology with the microscopic mechanisms of pharmacological synaptic modulation, providing an indispensable foundation for the emerging field of computational psychiatry.

4. The Bayesian Brain Hypothesis and Hierarchical Predictive Coding

4.1 Historical Ancestry: From Helmholtz to MacKay and Dayan

As Karl Friston formalized Dynamic Causal Modelling, he became increasingly preoccupied with the computational problem that the brain itself solves. Why are cortical circuits wired with such dense, recurrent, and reciprocal laminar connectivity? In addressing this question, Friston synthesized a profound intellectual lineage that had challenged the classical, passive stimulus-response view of perception.

The philosophical ancestor of this perspective was the nineteenth-century German polymath Hermann von Helmholtz. In his monumental Treatise on Physiological Optics (1867), Helmholtz posited that visual perception is fundamentally an ongoing process of unconscious inference (unbewusster Schluss). Helmholtz recognized that sensory receptors are bombarded by ambiguous, impoverished physical signals; an identical retinal pattern of photons can be generated by an infinite array of distinct three-dimensional real-world scenes (the inverse problem). Consequently, perception cannot simply be a passive, bottom-up readout of sensory arrays. Rather, the brain must actively infer the most probable hidden environmental causes capable of generating those raw sensory impressions, utilizing prior knowledge derived from past experience.

During the mid-twentieth century, this Helmholtzian insight was translated into the language of cybernetics and information theory by thinkers such as Donald MacKay, who proposed that the nervous system operates as an active hypothesis-testing organ, continuously matching internal predictions against incoming sensory traffic to reduce redundant computational throughput. In the 1990s, this lineage was operationalized in computational neuroscience through the groundbreaking work of Peter Dayan, Geoffrey Hinton, Radford Neal, and Richard Zemel in their formulation of the “Helmholtz Machine”, alongside the seminal predictive coding vision models of Rajesh Rao and Dana Ballard (1999). Friston embraced these computational concepts, but recognized that they had not yet been fully reconciled with the anatomical realities of cortical laminar architecture or unified into a single, comprehensive neurobiological theory. He resolved to formulate a mathematically rigorous, physiologically grounded architecture for the Bayesian brain.

4.2 Mechanics of Hierarchical Predictive Processing

Friston’s formalization of Hierarchical Predictive Coding recast the entire cerebral cortex as a multi-tiered, hierarchical inference engine. Rather than viewing the brain as an information-processing pipeline where sensory features are progressively extracted and passed up an ascending ladder toward higher cognitive centers, Friston inverted the functional direction of information flow.

In the predictive coding architecture, cortical regions are organized hierarchically, mirroring the nested temporal and spatial structure of the natural environment. The primary informational currents within this hierarchy are rigorously segregated across distinct anatomical pathways:

• Descending (Top-Down) Pathways: Originating from deep pyramidal cells in higher cortical tiers, these backward projections carry empirical prior predictions regarding the hidden causes of sensory inputs. These projections convey the brain’s internal best guess of what the lower tier is experiencing.
• Ascending (Bottom-Up) Pathways: Crucially, ascending forward pathways do not carry raw sensory information. Instead, superficial pyramidal cells transmit only the prediction error ($varepsilon$)—the residual mathematical difference between the actual incoming input and the top-down prediction:
  $$\varepsilon^{(i)} = y^{(i)} – g(x^{(i)})$$
  where $y^{(i)}$ is the sensory signal or representation at level $i$, and $g(x^{(i)})$ is the non-linear prediction generated by hidden states $x$ at the hierarchical level above.

The entire cortical hierarchy engages in a continuous, bidirectional recurrent message-passing cascade. Descending predictions strive to suppress and cancel out ascending prediction errors across every level of the hierarchy, from association cortices all the way down to primary sensory processing areas. At the same time, unpredicted prediction errors ascend to dynamically update and refine the higher-level representations, adjusting internal generative models until predictions match sensory reality. This continuous Bayesian updating eliminates the need to transmit redundant, predictable information, maximizing metabolic and computational efficiency while enabling real-time, robust perception under conditions of profound sensory uncertainty.

4.3 Precision Weighting and Attentional Mechanisms

A computational inference engine cannot treat all prediction errors equally. In a complex, fluctuating, and stochastic world, some sensory signals are crystal clear and reliable, while others are noisy, ambiguous, or irrelevant. Karl Friston realized that to make optimal Bayesian inferences, the brain must possess a mechanism to estimate the uncertainty or reliability of its own prediction errors. This computational imperative led to the formalization of precision weighting.

In mathematical statistics, precision ($Pi$) is the inverse of variance ($\Pi = \sigma^{-2}$); it measures the reliability, certainty, or signal-to-noise ratio of a probability distribution. In Friston’s predictive coding architecture, prediction errors are explicitly multiplied by their estimated precision before they ascend the cortical hierarchy:
$$\xi^{(i)} = \Pi^{(i)} \cdot \left( y^{(i)} – g(x^{(i)}) \right)$$
If an agent estimates that a sensory channel is shrouded in heavy stochastic noise (for example, attempting to visually discern an object through a dense fog or across a degraded visual monitor), the estimated precision of those sensory prediction errors is lowered. Consequently, the prediction error is computationally silenced; it lacks the synaptic potency to propagate upward and modify higher-level beliefs, forcing the agent to rely heavily on its prior expectations. Conversely, when precision is estimated to be high, prediction errors are amplified, exerting a dominant, corrective influence on high-level cognitive models.

Friston achieved a historic theoretical synthesis by demonstrating that precision weighting is the formal computational equivalent of attention. Attention, in this framework, is not an ephemeral cognitive spotlight, but rather the systematic optimization of synaptic gain on prediction errors. Neurally, this gain modulation is implemented through classical neuromodulatory systems—specifically dopamine, acetylcholine, and norepinephrine—acting upon the apical dendrites of superficial pyramidal cells, alongside synchronous gamma-band oscillatory dynamics. By altering the balance of synaptic precision, the brain dynamically regulates how open or closed it is to new sensory evidence. As we will see, this single mathematical mechanism—precision weighting—would unlock Friston’s revolutionary computational models of neuropsychiatric disease.

5. The Free Energy Principle: Mathematical Foundations

5.1 Thermodynamics, Information Theory, and Statistical Mechanics

While hierarchical predictive coding offered an elegant model of perceptual neuroanatomy, Karl Friston harbored an infinitely more ambitious aspiration. He sought a single, foundational physical principle that could account not merely for sensory perception in the mammalian brain, but for the fundamental existence, self-organization, and persistence of any biological entity whatsoever. This monumental quest culminated in the formulation of the Free Energy Principle (FEP).

The philosophical starting point of the FEP addresses what is often termed the “Schrödinger problem,” famously articulated by Erwin Schrödinger in his 1944 treatise What is Life?: How do living organisms resist the inexorable, decay-inducing drive of the Second Law of Thermodynamics? In an open, thermodynamic universe characterized by entropic dissipation, inanimate matter inevitably disperses into states of maximum disorder and thermal equilibrium. Yet living organisms—from the humblest prokaryote to the human being—manage to maintain their structural integrity, physiological boundaries, and internal chemical organization over prolonged evolutionary and somatic timescales.

Friston approached this challenge using the mathematics of non-equilibrium steady-state (NESS) statistical mechanics. An adaptive system that successfully persists must restrict its physiological states to a bounded, non-dispersive set of phenotypic configurations (its phase space attractor). In the language of probability theory, this means an organism must minimize the “surprise” (or surprisal) of its sensory interactions with the environment. Surprisal is formally defined as the negative log-evidence of sensory states:
$$Im(y) = -\ln p(y mid m)$$
where $y$ represents sensory sensations and $m$ denotes the agent’s evolutionary and structural generative model. An organism that routinely encounters states of high surprisal—such as a fish out of water, or a mammal experiencing catastrophic hypothermia—inevitably disintegrates and dies. Therefore, to live is to persistently minimize sensory surprisal.

However, an insurmountable computational and thermodynamic obstacle immediately presents itself: evaluating the true probability $p(y mid m)$ directly is impossible. Calculating marginal sensory likelihood requires integrating over all possible, unobservable configurations of the external world that could have caused those sensations—an intractable mathematical calculation. Friston resolved this paradox by importing the mathematical framework of variational inference, originally developed in quantum physics by Richard Feynman and adapted to machine learning by Geoffrey Hinton. Instead of minimizing surprisal directly, Friston showed that an organism can minimize an mathematically tractable upper bound on surprisal: Variational Free Energy ($F$).

5.2 Variational Inference and Generative Models

To understand the variational free energy formulation, one must formalize the fundamental ontological divide between the organism and the environment. There exists a true, physical external world whose states are hidden from direct view, termed the generative process. The organism maintains an internal, computational representation of this world, termed the generative model. The agent constructs an internal recognition density, $q(\vartheta)$, which represents its subjective, probabilistic beliefs regarding the unobserved environmental causes ($\vartheta$) of its sensations ($y$).

Variational free energy ($F$) is formally defined using the mathematics of the Kullback-Leibler (KL) divergence, a non-symmetric statistical measure of the informational difference between two probability distributions:
$$F = D_{\mathrm{KL}}[q(\vartheta) midmid p(\vartheta mid y)] – \ln p(y)$$
Because the KL divergence is mathematically proven to be strictly non-negative ($D_{\mathrm{KL}} ge 0$), the variational free energy is always greater than or equal to sensory surprisal ($F ge -ln p(y)$). Consequently, by adjusting its internal recognition density $q(\vartheta)$ to minimize $F$, an agent indirectly, inexorably minimizes its sensory surprise, rendering its internal beliefs maximally faithful to the true posterior probability distribution of the world: $p(\vartheta mid y)$.

Crucially, variational free energy can be algebraically rewritten into an illuminating alternative decomposition:
$$F = \text{Complexity} – \text{Accuracy} = D_{\mathrm{KL}}[q(\vartheta) midmid p(\vartheta)] – \mathbb{E}_{q(\vartheta)}[\ln p(y mid \vartheta)]$$
This formulation exposes the profound physical elegance of Friston’s framework:

• Accuracy: The expected log-likelihood of sensations given current internal beliefs. Maximizing accuracy means ensuring internal states effectively account for incoming sensory observations.
• Complexity: The KL divergence between the current recognition density $q(\vartheta)$ and prior beliefs $p(\vartheta)$. Minimizing complexity acts as a formal Occam’s razor, penalizing any internal representations that become unnecessarily convoluted or deviate gratuitously from basic prior constraints.

Thus, minimizing variational free energy forces the organism to construct models of the world that are as accurate as possible, while remaining as simple and generalizable as possible.

In continuous-time physical systems, Friston formalized the trajectory of internal states using stochastic differential equations, Fokker-Planck probability density flows, and Helmholtz decompositions. The motion of an organism’s states down the free energy landscape is characterized by two orthogonal dynamics: a dissipative, gradient descent flow directly toward the minimum of the free energy potential, and an anti-dissipative, solenoidal (rotational) flow along the isoprobability contours of the steady-state distribution. Through this formulation, life is shown to be a stable vortex of low entropy, persistently performing gradient descent on variational free energy to preserve its physical existence.

5.3 The Unification Agenda: Free Energy as a Grand Unified Theory

Karl Friston’s ambition with the Free Energy Principle extends far beyond computational neuroscience; it is openly presented as a Grand Unified Theory for the life sciences. Friston posits that the FEP provides a principled, single axiomatic language that can subsume, integrate, and contextualize nearly every major paradigm within twentieth-century biology and cognitive science.

Under the expansive umbrella of the FEP:

• Cybernetics: The cybernetic imperative of maintaining homeostatic viability (as articulated by W. Ross Ashby’s Law of Requisite Variety) is formalized as keeping internal physiological variables within a bounded, low-entropy phase space via free energy minimization.
• Reinforcement Learning and Optimal Control: The classical reward-seeking behavior of organisms is rewritten without the need for arbitrary, external reward functions. What reinforcement learning calls “rewards” are simply biological prior preferences built into the organism’s generative model. An organism does not seek rewards because they possess an arbitrary numerical value; rather, an organism naturally seeks out the conditions necessary to sustain its existence because it fundamentally predicts its own survival.
• Information Theory: Claude Shannon’s redundancy reduction principles and Horace Barlow’s efficient coding hypotheses are recovered as special, steady-state expressions of minimizing information-theoretic complexity.
• Evolutionary Biology: Natural selection itself is conceptualized by Friston as Bayesian Model Selection operating across evolutionary timescales. Genomes act as generative models, physical phenotypes constitute hypotheses tested against environmental niches, and survival and reproductive success represent the accumulation of model evidence—a phenomenon Friston famously terms self-evidencing.

The scale-free nature of the Free Energy Principle constitutes its most radical philosophical attribute. Because the mathematics of variational free energy relies strictly on statistical dependencies, it possesses no intrinsic spatial or temporal scale. It applies with equal mathematical validity to the metabolic homeostatic loops of an individual prokaryotic bacterium, the functional laminar dynamics of a neocortical column, the motor interactions of an embodied human organism, or the collective linguistic coordination of an entire human society. In Friston’s vision, the FEP represents a profound physical continuity connecting raw thermodynamics, biological morphogenesis, cognitive perception, and sociocultural evolution.

6. Active Inference: Unifying Perception, Action, and Epistemics

6.1 Beyond Passive Perception: Action as Prediction Fulfillment

The early formulations of predictive coding and Bayesian inference suffered from an acute epistemological limitation: they characterized the organism as a fundamentally passive observer. The classical Bayesian brain sat enclosed within the dark vault of the cranium, patiently receiving sensory inputs, calculating prediction errors, and updating internal representations to match the external world. Karl Friston realized that this passive view was biologically untenable. Organisms do not merely perceive the world; they act within it. They actively alter their sensory streams through movement, respiration, locomotion, and environmental modification.

This realization gave birth to Active Inference. Friston demonstrated that there exist two mathematically complementary and symmetric avenues through which an organism can minimize variational free energy and eliminate sensory prediction error:

1. Perception (Changing Beliefs): The organism can optimize its internal beliefs, $q(\vartheta)$, updating its generative model to accommodate incoming sensory inputs. This corresponds to classical, passive perceptual inference.
2. Action (Changing the World): The organism can directly change its sensory inputs, $y$, by executing physical motor actions that alter its physical relationship with the environment, forcing external sensory reality to conform to its internal, top-down predictions.

Friston illustrated this profound conceptual unity through the biophysics of the motor system. Classical neuroscience had long posited that the brain executes movement by computing complex, detailed motor commands that are sent down the spinal cord to instruct muscles to contract. Friston completely dismantled this paradigm. Under active inference, the motor system does not issue motor commands at all. Instead, the brain simply projects a descending, top-down proprioceptive prediction—a prediction that the limb is already in a new spatial position. When this descending prediction reaches the spinal cord, it collides with ascending sensory signals reporting that the limb is, in fact, still in the old position, creating a local proprioceptive prediction error. The classical spinal stretch reflex arc (the monosynaptic alpha-motor neuron loop) then operates to eliminate this prediction error by contracting the muscle until the limb attains the predicted position. Thus, motor action is nothing other than the mechanical fulfillment of descending predictions. Perception and action are revealed to be two functional sides of a single, indivisible physical imperative: the minimization of free energy.

6.2 Expected Free Energy and Policy Selection

While the minimization of instantaneous free energy accounts for real-time perception and reflex actions, adaptive organisms must also make prospective, forward-looking decisions across future temporal horizons. They must choose sequences of actions—termed policies ($\pi$)—that will protect them from encountering dangerous, high-surprisal states in the future. To formalize prospective planning and decision-making, Friston introduced the concept of Expected Free Energy (EFE), denoted mathematically as $G(\pi, \tau)$.

Expected Free Energy evaluates the variational free energy expected to occur at a future time $tau$ if a specific policy $\pi$ is pursued. Crucially, the mathematical decomposition of Expected Free Energy separates cleanly into two complementary, foundational terms:
$$G(\pi) \approx -\underbrace{\mathbb{E}_{q}[\ln p(o_{\tau} mid C)]}_{\text{Pragmatic Value (Exploitation)}} – \underbrace{\mathbb{E}_{q}[D_{\mathrm{KL}}[q(s_{\tau} mid o_{\tau}, \pi) midmid q(s_{\tau} mid \pi)]]}_{\text{Epistemic Value (Exploration)}}$$
where $o$ represents future observations, $s$ represents future hidden states, and $C$ denotes the agent’s prior evolutionary preferences over phenotypic outcomes.

This mathematical derivation elegantly resolves the notorious “exploration-versus-exploitation” dilemma that has haunted computer science, reinforcement learning, and behavioral ecology for decades:

• Pragmatic (Extrinsic) Value: Represents goal-directed, exploitative behavior. It drives the agent to select policies that bring about outcomes conforming to its prior preferences (e.g., obtaining food, warmth, and shelter).
• Epistemic (Intrinsic) Value: Represents curiosity-driven, exploratory behavior. It drives the agent to select policies that will maximize information gain—resolving uncertainty about the hidden states of the world. It compels an agent to explore novel, unfamiliar environments precisely because doing so eliminates future surprise.

In active inference, curiosity is not an arbitrary evolutionary afterthought, nor is it a heuristic bonus appended to a reward function. Epistemic exploration emerges naturally from the underlying mathematical physics of expected free energy. An active inference agent explores whenever it is uncertain, and exploits whenever it is confident, automatically orchestrating a seamless, optimal transition between information-seeking and goal-seeking behavior.

6.3 Temporal Depth, Planning, and Counterfactual Reasoning

The full formalization of active inference incorporates deep, hierarchical temporal structures. In sophisticated cognitive organisms such as primates and humans, generative models are not confined to reactive, instantaneous sensory-motor loops. Instead, they possess deep temporal depth, allowing agents to execute complex mental simulations across vast spatiotemporal horizons.

Friston mapped these computational dynamics onto discrete-time Partially Observable Markov Decision Processes (POMDPs). In these discrete hierarchical architectures, high-level cortical areas simulate trajectories of future hidden states that unfold over slow temporal scales, while lower tiers resolve immediate, rapid sensory-motor dynamics. Friston and his collaborators identified the prefrontal cortex and the hippocampus as critical biophysical substrates for this counterfactual prospective planning. Within the hippocampal-prefrontal axis, the brain continuously executes “trajectory rollouts”—simulating the counterfactual outcomes of prospective policies without physically enacting them in the external world. This enables high-level planning, abstract problem solving, and deliberative decision-making.

Furthermore, active inference extends this prospective logic into the internal somatic domain of interoception. Just as the brain maintains generative models of the external visual and auditory environment (exteroception), it actively models the physiological state of the viscera, heart, lungs, and gut. In this interoceptive formulation, emotions and affective states are mathematically conceptualized as inferences regarding the rate of free energy minimization over time. The experience of subjective valence—feeling positive or negative affect—is tied directly to whether an agent is minimizing Expected Free Energy faster or slower than anticipated. Affective cognition is thus integrated into the solitary imperative of active inference: an embodied, interoceptive appraisal of the organism’s capacity to navigate its environment successfully and preserve its homeostatic viability.

7. Markov Blankets and the Physics of Sentience

7.1 The Formal Topology of Markov Blankets

To assert that a system exists as a distinct biological entity, one must resolve a profound ontological question: Where does the physical organism end, and where does the external environment begin? How does nature carve out a localized, self-sustaining system from the ambient thermodynamic background? Karl Friston solved this foundational boundary problem by appropriating and physicalizing a concept originally formulated by the computer scientist Judea Pearl within the domain of probabilistic graphical models: the Markov blanket.

In formal probability theory, the Markov blanket of a designated variable in a directed graph is the minimal set of nodes containing its parents, its children, and any other parents of its children. Possessing knowledge of this blanket renders the designated variable statistically conditionally independent of all remaining variables in the network. Friston recognized that this statistical topology could be translated into physical dynamical systems to define the structural boundary of an autonomous agent.

Friston partitions the states of the universe into four mutually exclusive categories:

• External States ($eta$): Environmental configurations outside the organism, which cannot be perceived or affected directly.
• Internal States ($\mu$): The physiological, computational, and structural configurations of the organism itself.
• Sensory States ($s$): States influenced by external configurations, which in turn directly affect internal states.
• Active States ($a$): States influenced by internal configurations, which in turn directly affect external states.

The union of Sensory States and Active States constitutes the Markov Blanket ($b = {s, a}$). Under this formal partition, the internal states and the external states are conditionally independent given the Markov blanket:
$$p(\mu, \eta mid b) = p(\mu mid b) , p(\eta mid b)$$
The Markov blanket acts as an informational and statistical veil. The internal states never have direct access to the external world; they “see” the external world exclusively through the veil of sensory states, and they “act” upon the external world exclusively through active states. Friston demonstrated that whenever a system possesses a non-equilibrium steady state, a Markov blanket emerges naturally as an objective, statistical boundary separating the system from its embedding environment.

7.2 Autopoiesis, Boundary Formation, and Living Organization

The mathematical formalization of the Markov blanket allowed Friston to build a formal bridge to one of the most celebrated philosophical concepts in theoretical biology: the theory of autopoiesis (self-creation and self-maintenance), pioneered in the 1970s by Chilean biologists Humberto Maturana and Francisco Varela. Maturana and Varela had argued that the defining attribute of a living system is its circular organization; an autopoietic system is an autonomous network of biochemical processes that continuously produces, regenerates, and maintains the very boundaries that define it.

Friston provided the rigorous mathematical physics that autopoietic theory historically lacked. Through the Free Energy Principle, the maintenance of an autopoietic boundary is shown to be a direct mathematical consequence of variational free energy minimization. Consider the archetypal biological boundary: the phospholipid bilayer of a single biological cell. The membrane is a literal, biophysical Markov blanket. The physical receptors embedded within the lipid membrane act as sensory states, registering biochemical concentrations in the extracellular medium. The ion channels, enzymatic pumps, and active transporters act as active states, expelling metabolites and maintaining internal osmotic equilibrium.

If the cell fails to minimize its variational free energy, the precision of its Markov blanket degrades. The cell can no longer distinguish its internal states from the external bath; ions rush across the electrochemical gradient, entropic dissipation takes over, the membrane dissolves, and the cell undergoes lytic necrosis. Thus, the physical maintenance of a biological boundary is structurally identical to the continuous informational optimization of a Markov blanket. Under Friston’s physics of sentience, to exist as an organism is to maintain a Markov blanket.

7.3 Nested Blankets and Scale-Free Cognitive Architectures

One of the most profound and far-reaching dimensions of Karl Friston’s topological framework is the concept of nested Markov blankets. Because the Free Energy Principle is inherently scale-free, Markov blankets do not exist in isolation at a single biological tier; rather, they are recursively and hierarchically nested within one another across nature.

Consider the scale-free architecture of human biology:

• At the subcellular scale, an organelle like the mitochondrion possesses a double membrane that serves as a Markov blanket, separating internal biochemical reactions from the cellular cytoplasm.
• The containing cell itself possesses a plasma membrane—a Markov blanket that wraps around its constituent organelles.
• Aggregations of cells form tissues and organs (such as the brain or the liver), which are insulated by epithelial membranes and vascular blood-brain barriers.
• These organs form the embodied organism, whose skin and sensory receptors constitute the macroscopic Markov blanket mediating interaction with the broader physical ecology.
• Extending outward, individual human organisms, coupled via linguistic communication, acoustic signals, and cultural artifacts, constitute nested social groups, institutions, and ecosystems, each possessing its own collective statistical blankets.

This nested topology dissolves the classical Cartesian dichotomy between the inner mental realm and the external physical world. Mind and cognition are revealed to be distributed phenomena. Through coupled active inference, two organisms can synchronize their internal states by coordinating their mutual sensory and active states—a process Friston models as collective or shared active inference. The theory of nested Markov blankets provides a rigorous, mathematical taxonomy for biological individuality, explaining how agency, cognition, and subjectivity can emerge at multiple, hierarchical layers of living organization.

8. Applications in Computational Psychiatry and Clinical Neurobiology

8.1 Schizophrenia, Hallucinations, and Precision Dysregulation

The ultimate test of any foundational physical theory of mind is its capacity to resolve the very clinical mysteries that originally motivated it. Returning full circle to his early psychiatric rotations at the Maudsley Hospital, Karl Friston systematically applied the principles of predictive coding and active inference to revolutionize computational psychiatry, beginning with the mechanistic deconstruction of schizophrenia and clinical psychosis.

Under Friston’s framework, psychosis is not a generalized cognitive collapse, but rather a specific failure of sensory attenuation driven by the aberrant dysregulation of precision weighting. In healthy individuals, when a self-generated motor or cognitive action is initiated (such as speaking or moving an arm), the brain predicts the sensory consequences of that action and transiently down-weights (attenuates) the sensory precision of the corresponding ascending prediction errors. This sensory attenuation explains why an individual cannot tickle themselves: the brain accurately predicts the tactile sensations and computationally silences the resulting prediction errors.

In schizophrenia, Friston posits a primary biophysical failure in the regulation of synaptic precision, heavily implicating NMDA receptor hypofunction on cortical interneurons and aberrant dopaminergic transmission:

• Failure of Sensory Attenuation: The brain fails to down-weight sensory precision during self-generated actions. Consequently, when a schizophrenic patient speaks or thinks, the sensory feedback arrives with anomalous, hyper-intense precision. The patient cannot recognize the sensation as self-generated, interpreting their own internal monologue as an alien, external auditory-verbal hallucination or a delusion of control (passivity phenomenon).
• Hallucinations: As high-level prior predictions become decoupled from reality, compensatory hyper-precision is assigned to top-down priors. A hallucination occurs when descending empirical predictions are imbued with such overwhelming precision that they literally override ascending sensory inputs, generating vivid, conscious perceptual experiences in the total absence of external stimuli.
• Delusions: Delusional ideation (such as paranoia or grandiose conspiracy theories) represents a secondary, optimal cognitive inference. The patient’s higher-level cortical tiers are continuously bombarded by non-attenuated, aberrant prediction errors carrying immense subjective significance (“aberrant salience”). Delusions are constructed by the generative model as the only possible, logically coherent hypothesis capable of explaining away these relentlessly anomalous signals.

Through this model, Friston successfully bridged the explanatory gap between microscale molecular receptor dysfunctions and the macroscale, bizarre subjective phenomenology of psychotic illness.

8.2 Autism Spectrum Conditions: Sensory Hyper-Precision and World Invariance

Friston, in profound collaboration with cognitive scientists such as Sander Van de Cruys and Rebecca Lawson, extended active inference to provide an authoritative computational model of autism spectrum conditions (ASC). The predominant framework, known as the High Inflexible Precision of Prediction Errors in Autism (HIPPEA) model, explains the complex triad of autistic phenomenology through a solitary computational alteration: an abnormally high, inflexible baseline setting of low-level sensory precision.

In the neurotypical brain, precision weighting is highly contextual and dynamic. In complex, noisy, or fluctuating environments, low-level sensory precision is automatically dialed down, allowing the agent to ignore minor visual or auditory fluctuations and grasp global, gestalt meanings. In the autistic brain, however, low-level sensory prediction errors are perpetually assigned rigid, hyper-intense precision. This leads to a profound cascade of cognitive consequences:

• Sensory Overload: Because sensory precision cannot be attenuated, every minute sensory variation—fluorescent lighting hums, textural shifts in fabric, subtle background noises—is registered by the brain as an urgent, highly salient prediction error demanding cognitive resolution. The physical world becomes overwhelmingly intense and unpredictable.
• Insistence on Sameness: Stereotyped movements, repetitive behaviors (stimming), and an insistence on rigid daily routines are recast as rational active inference strategies. When an agent experiences the sensory world as chronically hyper-volatile and unpredictable, the most effective way to minimize overwhelming sensory free energy is to actively impose total invariance upon the environment, engaging in repetitive, predictable motor loops.
• Atypical Social Interaction: Human social communication is arguably the most complex, unpredictable, and context-dependent signal domain in existence. Sarcasm, irony, metaphorical gaze shifts, and facial expressions require flexible, dynamic attenuation of low-level details in favor of high-level context. The hyper-precision of low-level sensory signals prevents the autistic brain from flexibly abstracting away the micro-fluctuations of human social behavior, leading to challenges in reciprocal communication.

8.3 Affective Disorders, Interoception, and Parkinsonian Pathology

The clinical reach of Karl Friston’s computational framework extends across the spectrum of affective and motor pathology. In major depressive disorder and learned helplessness, Friston, working alongside colleagues like Ray Dolan and Jonathan Roiser, conceptualized depression as an aberrant meta-prediction regarding the agent’s own agency. The depressed individual’s generative model has learned that no policy it selects will succeed in minimizing expected free energy. Because the agent infers that its actions are computationally impotent to resolve environmental and visceral stress, it down-regulates the precision of all active policies, manifesting as profound psychomotor retardation, anhedonia, and vegetative withdrawal.

In the realm of interoceptive active inference, Friston and neuroscientists like Anil Seth and Hugo Critchley demonstrated how dysfunctions of the insular cortex generate psychosomatic disorders, panic conditions, and chronic fatigue syndrome. In these conditions, the brain forms hyper-precise, catastrophic predictions regarding visceral autonomic signals—heart rate, gastrointestinal motility, metabolic exhaustion. The resulting interoceptive prediction errors dominate the cognitive hierarchy, causing systemic bodily sensations to be experienced as acute medical crises.

In motor neurology, Friston applied active inference to demystify Parkinson’s disease. Classical neurology struggled to explain why Parkinsonian patients suffer from akinesia (an inability to initiate voluntary movement) yet can catch a ball thrown at them unexpectedly (paradoxical kinesia). Friston showed that initiating voluntary movement requires an agent to temporarily down-weight the precision of its current proprioceptive states, allowing the descending motor prediction to dominate. Because dopamine in the basal ganglia is the biophysical substrate that encodes the precision of action policies, the dopaminergic degeneration of Parkinson’s renders the patient unable to lower proprioceptive precision, locking them in their current posture. However, when an external object (such as a ball) suddenly looms into view, powerful exteroceptive prediction errors bypass the damaged basal ganglia via subcortical visual pathways, instantly triggering reflexive motor action. Through these formulations, Friston’s framework provides a unified, mathematically tractable nosology for clinical medicine.

9. Morphogenesis, Cellular Cognition, and Multiscale Biology

9.1 Collaborations with Michael Levin: Morphogenetic Active Inference

In recent years, Karl Friston’s scientific vision has expanded beyond the brain into the broader landscape of theoretical and developmental biology. Through a groundbreaking, highly productive collaboration with the developmental biologist Michael Levin at Tufts University, Friston has integrated the Free Energy Principle with developmental biology and bioelectricity, formulating the revolutionary paradigm of Morphogenetic Active Inference.

Historically, developmental biology viewed embryogenesis and tissue patterning as a rigid, feed-forward biochemical cascade governed strictly by genetic code execution. Friston and Levin challenged this paradigm, proposing that non-neural cells, tissues, and developing organs operate as active inference agents navigating a high-dimensional “morphological space”. Just as neurons utilize action potentials to communicate, all somatic cells utilize endogenous bioelectric gradients—mediated by resting membrane potentials ($V_{mem}$) coordinated through gap junctions—to share informational models of anatomical target structures.

In this framework, cells maintain a collective generative model of the organism’s anatomical target configuration. Bioelectric patterns encode the “prior predictions” of what the tissue should look like (e.g., the correct spatial positioning of an eye or a limb in a developing amphibian). When tissue damage occurs—such as amputation in a planarian flatworm—a morphogenetic prediction error emerges between the current aberrant shape and the bioelectric target morphology. The cells actively minimize this free energy by migrating, proliferating, and differentiating until the anatomical prediction error is canceled out, successfully restoring anatomical homeostasis. This framework has immense implications for regenerative medicine, suggesting that clinicians can repair congenital birth defects or induce organ regeneration not through painstaking gene editing, but by providing bioelectric informational prompts that harness the innate active inference capacity of somatic tissue.

9.2 Basal Cognition and Pre-Neural Agency

The collaboration between Friston and Levin naturally demolished the neurocentric dogma that cognition requires a centralized nervous system. If the Free Energy Principle is a universal mathematical imperative of life, then basal cognition must be ubiquitous across all biological scales, from simple prokaryotes to botanical networks.

Friston illustrated this by demonstrating that simple organisms such as Escherichia coli, the slime mold Physarum polycephalum, and vascular plants execute active inference without possessing a single neuron:

• Bacterial Chemotaxis: An E. coli swimming through a chemical concentration gradient operates an active inference loop. Its biochemical signaling cascades (such as the CheA-CheY pathway) calculate prediction errors between current and desired nutrient concentrations, switching motor flagellar rotation between directed swimming (exploitation) and random tumbling (epistemic exploration).
• Slime Molds: The acellular slime mold Physarum resolves complex mathematical challenges, such as navigating complex labyrinths and designing optimal transport networks, by using rhythmic, bioelectric actomyosin contractions to perform physical gradient descent on free energy landscapes.
• Subcellular Metabolism: At the most fundamental level, intracellular metabolic networks operate as predictive computational engines, maintaining homeostatic equilibrium by continuously minimizing free energy against fluctuating biochemical stress.

Cognitive awareness, under Friston’s multiscale lens, is not an all-or-nothing phenomenon that suddenly appeared with the evolution of the mammalian cerebral cortex. Rather, human consciousness is merely a technologically and hierarchically sophisticated elaboration of a primordial, scale-free physical sentience that has defined biological matter since the origin of life on Earth.

9.3 Cancer Biology as a Breakdown of Multi-Scale Blankets

The application of nested Markov blankets and active inference to multiscale biology has yielded a transformative etiology of oncogenesis and cancer biology. Traditional oncology has long viewed cancer through the Somatic Mutation Theory: an irreversible accumulation of random genetic mutations within a rogue cell that drives unregulated proliferation.

Friston and Levin re-conceptualized cancer as a radical communicative and statistical failure: the breakdown of nested Markov blankets. In healthy tissue, a cell is integrated into a multi-tiered communal blanket. Through continuous bioelectric and chemical coupling via gap junctions, the individual cell’s generative model is enslaved by the macroscopic generative model of the higher-order tissue. The cell minimizes its free energy by acting in harmony with the organ’s homeostatic needs, executing apoptosis when instructed, and regulating its own division.

Carcinogenesis occurs when a cell becomes functionally and informationally disconnected from its communal tissue blanket—often initiated by the closure or down-regulation of gap junctions. Severed from the collective predictive signaling of the organ, the cell’s Markov blanket shrinks to its own individual boundary. Cut off from higher-order constraints, the cell reverts to a primordial, single-celled evolutionary imperative: it treats the rest of the body as an alien, hostile environment. The tumor cell minimizes its free energy by maximizing its own individual survival, scavenging resources, and proliferating uncontrollably. Metastatic invasion is thus mathematically modeled as aberrant active inference: the rogue cell, seeking predictable, viable niches, migrates across the body. This paradigm suggests revolutionary oncological interventions designed not to carpet-bomb tumors with cytotoxic chemotherapy, but rather to pharmacologically reopen gap junctions, re-integrating the malignant cells into the collective Markov blanket of the healthy organism.

10. Artificial Intelligence, Neuromorphic Architectures, and Robotics

10.1 Active Inference Versus Deep Reinforcement Learning

As the contemporary artificial intelligence landscape reached a historic inflection point in the 2010s and 2020s, Karl Friston’s theories emerged as a powerful, radically divergent paradigm to the mainstream dogmas of Deep Learning and Deep Reinforcement Learning (DRL). While acknowledging the impressive empirical successes of massive Transformer models and deep neural networks trained via backpropagation, Friston and his artificial intelligence collaborators have pointed out severe, structural limitations inherent in classical deep learning:

Traditional deep reinforcement learning is notoriously sample-inefficient, often requiring millions of training trials to master basic tasks. It suffers from “catastrophic forgetting,” behaves as a brittle, inscrutable black box lacking explainability, and relies on human engineers to hand-craft arbitrary scalar “reward functions.” When deployed in real-world environments outside their narrow training distribution, deep neural networks display catastrophic failures and remain profoundly vulnerable to tiny, adversarial perturbations.

Active Inference offers a biologically grounded, first-principles alternative:

• No External Reward Engineering: Active inference agents operate without scalar reward functions. Instead, an agent’s goals are encoded as intrinsic prior beliefs within a generative model. Motivation is intrinsic, derived directly from the physical mandate to minimize free energy.
• Sample Efficiency and Epistemic Drive: Because Expected Free Energy formally incorporates epistemic value (information gain), active inference agents do not require brute-force random exploration. They systematically seek out data that resolves uncertainty, learning causal models of the environment in a fraction of the trials required by DRL.
• Adversarial Robustness: In deep learning, an adversarial image tricks a feed-forward network by introducing noise that flips a static classification layer. An active inference agent, by contrast, is an ongoing, dynamical predictive system. It continuously matches sensory evidence against top-down predictions; aberrant adversarial noise is instantly flagged as an anomalous prediction error and resolved through real-time belief updating.
• Explainable AI (XAI): Because active inference models are built on transparent, probabilistic causal architectures, every decision, action, and inference made by the agent can be directly traced and understood through Bayesian mathematics.

10.2 Neuromorphic Hardware and Energy-Efficient Computing

A staggering environmental and computational crisis facing contemporary artificial intelligence is its astronomical energy consumption. Training and running modern trillion-parameter deep learning models on silicon GPUs requires gigawatts of electrical power and massive cooling infrastructure. In sharp contrast, the human brain—the most sophisticated computing machine in the known universe—operates on a metabolic budget of approximately 20 watts, roughly equivalent to a dim incandescent light bulb.

Karl Friston recognized that this monumental disparity in energy efficiency stems directly from computational architecture. Contemporary AI relies on the von Neumann architecture, shuttling continuous, floating-point weights back and forth between separate memory and processing units while performing global backpropagation across dense, synchronized layers. The biological brain, however, is a localized, asynchronous, event-driven predictive engine. Active inference and hierarchical predictive coding operate strictly via local message passing: a cortical column updates its internal weights using only the prediction errors arriving from its immediate neighbors, eliminating the need for biologically impossible and energetically ruinous global backpropagation.

To realize this vision in silicon, Friston has actively collaborated with neuromorphic engineers and semiconductor physicists. Predictive coding architectures are uniquely suited for implementation on neuromorphic hardware and Spiking Neural Networks (SNNs)—such as Intel’s Loihi or custom physical memristor arrays. In these ultra-low-power edge computing systems, artificial neurons communicate via discrete, sparse electrical spikes only when an unexpected event (a prediction error) occurs. When the environment is predictable, the system falls computationally silent, consuming near-zero power. Furthermore, Friston is pioneering thermodynamic computing paradigms that leverage the natural non-equilibrium physics of physical substrates to perform variational inference directly, promising a revolution in sustainable, energy-efficient artificial intelligence.

10.3 Spatial Computing, VERSES AI, and Autonomous Systems

In a major transition from purely academic theory to industrial application, Karl Friston stepped into a prominent role in the technology sector, being appointed as Chief Scientist at VERSES AI, a cognitive computing company dedicated to next-generation artificial intelligence. This appointment signaled a decisive effort to deploy active inference as the foundational operating system for autonomous robotics, smart cities, and distributed spatial computing.

Working alongside AI systems architects, Friston has contributed to the development of standardized protocols for autonomous intelligence, collaborating on initiatives such as the Spatial Web and the IEEE P2874 working group. The objective is to construct an interoperable digital ecosystem where heterogeneous autonomous agents—such as autonomous drones, self-driving vehicles, robotic warehouse fleets, and industrial Internet-of-Things (IoT) sensors—can interact within a shared spatial context.

Rather than relying on isolated, monolithic large language models, this vision implements multi-agent active inference networks governed by shared Markov blankets. Each autonomous agent maintains a localized generative model of its specific task and spatial coordinates. By communicating via standardized, sparse probabilistic messages, these multi-agent networks can dynamically coordinate collective behavior, self-organize logistical workflows, and navigate complex, unexpected industrial environments without human intervention. This industrial trajectory positions active inference not merely as an abstract biological theory, but as an engineering framework for environmentally situated, biologically plausible Artificial General Intelligence (AGI).

11. Philosophical, Epistemological, and Foundational Critiques

11.1 The Falsifiability and Tautology Debate

Despite its vast influence, Karl Friston’s Free Energy Principle has ignited intense philosophical debate and methodological controversy. A prominent coalition of philosophers of science and theoretical neurobiologists—including Colin Klein, J. Colin, and Matteo Colombo—has leveled a severe epistemic critique against the FEP: Is the Free Energy Principle actually an empirically falsifiable scientific hypothesis, or is it merely an untestable, mathematical tautology?

The critique argues that the FEP has been formulated with such expansive, unconstrained generality that no possible empirical observation could ever contradict it. If an animal behaves in a manner that appears to maximize reward, the FEP claims it is minimizing free energy. If an animal behaves self-destructively or irrationally, the FEP claims the animal is still minimizing free energy, but under aberrant priors or pathological precision weighting. If an organism sits motionless, it is minimizing free energy; if an organism frantically explores, it is minimizing expected free energy. Critics contend that a principle capable of explaining every conceivable biological outcome explains nothing at all, failing the basic Popperian demarcation criterion of empirical science.

Karl Friston has mounted an unapologetic and rigorous defense against this charge, clarifying the precise epistemological status of the FEP. Friston emphasizes that the Free Energy Principle is not an empirical, biological hypothesis; it is a foundational mathematical mechanic. In this regard, Friston explicitly compares the FEP to William Rowan Hamilton’s Principle of Least Action in classical mechanics. The Principle of Least Action is not an empirical claim that can be disproven by an experiment; rather, it is a normative mathematical language, a foundational variational calculus through which specific physical systems are modeled and understood.

Friston argues that what must be empirically tested and falsified are not the foundational mathematics of the FEP itself, but rather the specific, instantiated process theories derived from it. For example, hierarchical predictive coding, specific laminar wiring diagrams, Dynamic Causal Models, and dopamine-as-precision hypotheses are fully falsifiable empirical claims. If an experiment demonstrates that superficial pyramidal cells do not transmit prediction errors, or that acetylcholine fails to track sensory precision, the process theory is falsified, leaving the foundational mathematical language intact. This distinction between foundational mechanics and empirical process theories remains a pivotal subject in contemporary philosophy of cognitive science.

11.2 Instrumentalism Versus Scientific Realism of Markov Blankets

A second major philosophical battleground concerns the ontological status of Markov blankets. Within cognitive science and philosophy of mind, scholars are sharply divided between two opposing camps: scientific realism and scientific instrumentalism.

The instrumentalist position—championed by philosophers who view cognitive modeling through an operational lens—posits that Markov blankets do not exist as literal, physical structures in the real world. Instead, they are merely pragmatic, heuristic computational tools invented by the human scientist. When an investigator models a cell or a brain, drawing a Markov blanket around it is simply a convenient statistical technique to define the boundaries of the model. The blanket is in the eye of the beholder, entirely dependent on how the researcher chooses to partition the state space. To reify the Markov blanket into an actual physical boundary is, for the instrumentalist, a dangerous category mistake.

Conversely, Friston and realist philosophers such as Michael Kirchhoff and Julian Kiverstein advocate for a nuanced form of structural realism. They argue that Markov blankets are not arbitrary human impositions; they are objectively emergent, physical features of non-equilibrium dynamical systems. When a cell forms a membrane, or when an organism maintains homeostasis, the statistical conditional independencies are physically real and causally potent. Without this objective statistical boundary, the system would undergo thermodynamic decay.

This debate intersects directly with contemporary theories of embodied cognition, radical enactivism, and the Extended Mind Thesis pioneered by Andy Clark and David Chalmers. Clark, an enthusiastic champion of predictive processing, has engaged in rich debates with Friston regarding where the cognitive blanket falls. If an individual interacts deeply with an external notebook or a smartphone, does the Markov blanket expand to incorporate the tool into the internal state of the agent? Friston’s mathematical framework provides the formal tools to debate these questions, redefining traditional Cartesian discussions of the mind-body problem into the rigorous mathematics of statistical boundary formation.

11.3 Computational Intractability and Dark Room Paradoxes

Beyond philosophical debates, the Free Energy Principle has faced substantial mathematical and behavioral challenges, none more famous than the notorious “Dark Room Paradox.”

The paradox, initially posed as an intuitive critique, runs as follows: If the fundamental physical imperative of any living organism is to minimize surprise and variational free energy, the simplest, most optimal behavioral strategy for any human being would be to immediately enter a pitch-black, silent, completely empty room and remain there indefinitely. In a completely dark room, sensory inputs are utterly invariant, sensory noise is zero, and prediction errors are completely extinguished. The dark room should be the absolute paradise of free energy minimization. Yet real-world organisms actively avoid dark rooms; if placed in one, they quickly grow restless, turn on lights, explore, and eventually leave to seek food, stimulation, and social interaction. Why does the FEP not lead to universal catatonia?

Friston’s resolution of the Dark Room Paradox is mathematically absolute and illuminates the heart of active inference: organisms are not empty inference engines; they are endowed with evolutionary prior expectations. A human being’s generative model does not predict that it will exist in a dark, empty void. An organism’s evolutionary priors—forged through millions of years of natural selection and embedded in its physiological genome—predict that it will maintain a body temperature of 37 degrees Celsius, maintain a specific blood glucose level, seek social interaction, and explore novel environments. Therefore, for a human being, remaining trapped in a dark room is, in fact, a state of immense, catastrophic free energy. As blood sugar drops, dehydration sets in, and sensory deprivation conflicts with expected environmental engagement, massive prediction errors cascade up the cortical hierarchy. To minimize free energy, the organism must take action: leave the dark room, hunt for calories, and find a mate. The dark room paradox dissolves entirely once it is understood that agents actively minimize free energy relative to their phenotypic, evolutionary priors.

A second, formidable critique concerns computational tractability. Classical variational inference relies on the Laplace approximation (assuming Gaussian probability distributions) or mean-field approximations (assuming statistically independent factors). Critics from theoretical computer science point out that the real world is radically non-Gaussian, highly multimodal, and fraught with combinatorial explosions. If the approximations required to compute variational free energy are too coarse, how can biological wetware successfully navigate real-world high-dimensional physics? Friston’s ongoing work in generalized coordinates of motion and non-linear, non-Gaussian path-integral formulations strives to address these challenges, pushing the absolute frontiers of mathematical physics to demonstrate that biological networks possess precisely the right mathematical heuristics to navigate intractable computational spaces.

12. Contemporary Horizons, Intellectual Legacy, and Future Trajectories

12.1 Scientific Accolades, Quantitative Impact, and Cultural Standing

Karl Friston’s sustained intellectual output has produced a quantitative and qualitative legacy that is virtually unprecedented in contemporary biological science. His academic impact defies conventional metrics. With an h-index exceeding 250 and total citations surpassing half a million, Friston has been consistently ranked by analytical ranking agencies (such as Clarivate and Semantic Scholar) as the most cited and influential neuroscientist on Earth. In 2006, his revolutionary methodological and theoretical contributions were formally recognized with his election as a Fellow of the Royal Society (FRS), Britain’s most prestigious scientific academy, followed by his election to the Academy of Medical Sciences (FMedSci).

Beyond raw bibliometrics, Friston’s true impact lies in the extraordinary international academic diaspora he has cultivated. The Wellcome Centre for Human Neuroimaging at Queen Square has operated for three decades as the world’s premier intellectual incubator for computational neuroscience. Friston has mentored, collaborated with, and inspired an entire generation of globally recognized scientists—scholars who have gone on to lead major neuroscience and philosophy departments across Europe, North America, Asia, and Australasia. His laboratory is legendary for its open, rigorous, and collaborative atmosphere, where psychiatrists, quantum physicists, software engineers, and continental philosophers sit side by side at the Queen Square blackboard.

Friston’s cultural standing has also permeated the broader intellectual zeitgeist. While his academic papers are celebrated (and occasionally lamented) for their extraordinary mathematical density, formidable vocabulary, and uncompromising analytical complexity, he has emerged as a captivating public intellectual. Profiles in publications such as Wired and major scientific podcasts have introduced his radical vision of the human mind to millions outside the academy. Friston presents a uniquely charismatic presence: soft-spoken, unfailingly modest, endlessly patient in unpacking Byzantine mathematical formulas, and relentlessly smoking his trademark cigarettes outside Queen Square while casually synthesizing quantum thermodynamics with Hegelian philosophy.

12.2 Planetary Active Inference and Ecosystemic Dynamics

In the contemporary era, Karl Friston has increasingly turned his analytical gaze toward the macroscopic challenges confronting human civilization. Recognizing that the Free Energy Principle is profoundly scale-free, Friston and a growing international cohort of complex systems theorists have begun formulating the mechanics of Planetary Active Inference.

This macroscale framework extends the mathematics of nested Markov blankets to the global biosphere:

• Ecosystemic Resilience and Climate Change: Planetary stability is conceptualized as an overarching, planetary-scale Markov blanket. The Earth’s climate system, carbon cycles, ocean currents, and forests form coupled homeostatic feedback loops that continuously dissipate solar energy while minimizing environmental entropy. Anthropogenic climate change represents a catastrophic rupture of these planetary homeostatic blankets—an ecological state of massive, systemic free energy where the biosphere’s predictive regulatory cycles are destabilized by industrial activity.
• Sociological Dynamics and Political Polarization: Friston has applied active inference to understand the breakdown of civic discourse in the digital age. Social communities operate as collective generative models. In the contemporary media landscape, algorithmic echo chambers act as artificial, hyper-insulated Markov blankets. To minimize free energy, insular online subcultures systematically reject contradictory evidence, treating outgroups with hyper-skepticism and accelerating political polarization.
• Pandemic Modelling: During the global COVID-19 crisis, Friston deployed dynamic causal modelling to epidemiologic data, creating real-time dynamic transmission models for the UK government that accounted for behavioral changes and viral transmission rates, showcasing how active inference can be leveraged to preserve human life during international health crises.

12.3 The Enduring Scientific Stature of Karl Friston

To assess the ultimate historical standing of Karl Friston is to locate his work within the grand, heroic lineage of British and European intellectual history. His career exhibits the visionary mathematical scope of Alan Turing, the systemic cybernetic brilliance of Norbert Wiener and W. Ross Ashby, and the foundational physical insight of Ludwig Boltzmann and Hermann von Helmholtz. Friston belongs to that vanishingly rare breed of intellectual polymaths who refuse to be imprisoned within the narrow confines of academic specialization.

Whether the Free Energy Principle will ultimately be canonized as the singular, undisputed “Grand Unified Theory of Biology” remains an active, vibrant debate within the scientific community. Empirical demonstrations are ongoing, mathematical formalisms continue to be refined, and philosophical arguments remain intensely contested. Yet what is indisputable is that Friston has irreversibly altered the trajectory of human knowledge. He has dismantled the naive stimulus-response models of twentieth-century behaviorism, liberated neuroimaging from unprincipled correlational mapping, constructed an unbreakable mathematical bridge between the mind and physical matter, and given humanity a unified vocabulary to comprehend how thought, life, and the physical universe cohere.

Friston’s work ultimately delivers a breathtaking, poetically profound vision of our place in the cosmos. We are not alienated spectators observing a detached, external physical reality. We are localized, self-evidencing vortices of low entropy, sculpted by billions of years of evolutionary physics. Through our sensory and active blankets, our brains continuously weave probabilistic hypotheses, striving to understand the hidden mysteries of the world while simultaneously sculpting that world into a livable home. In the mathematics of Karl Friston, the physics of the universe has become conscious of itself, perpetually minimizing its uncertainty, endlessly exploring the dark, and transforming sensory chaos into the structured, magnificent reality of human thought.

Conclusion

From the psychiatric wards of the Maudsley Hospital to the vanguard of theoretical physics and artificial intelligence, the intellectual journey of Karl Friston embodies one of the most astonishing scientific narratives of the modern era. By inventing Statistical Parametric Mapping, he democratized the human neuroimaging revolution; by creating Dynamic Causal Modelling, he taught neuroscience to infer the directed, causal mechanisms of the living mind; and by formulating the Free Energy Principle and active inference, he delivered a breathtaking, mathematically rigorous architecture capable of explaining the emergence of life, action, and consciousness from basic physical law.

Friston has revealed that living systems are fundamentally self-evidencing entities: open thermodynamic engines that preserve their physical existence against entropic dissolution by persistently inferring their world and acting to fulfill their own homeostatic expectations. In an academic world frequently fractured by hyper-specialization, Friston stands as a monumental synthesizer—a natural philosopher whose work unites the microscopic conductance of a synaptic channel with the macroscopic self-organization of human societies. His enduring legacy ensures that as science continues its quest to decipher the ultimate nature of mind and reality, the path forward will remain illuminated by the foundational equations, topological boundaries, and visionary insights of Karl John Friston.

References

• Ashburner, J., & Friston, K. J. (2000). Voxel-based morphometry—the methods. NeuroImage, 11(6), 805–821. https://doi.org/10.1006/nimg.2000.0582
• Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204. https://doi.org/10.1017/S0140525X12000477
• Friston, K. J. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society B: Biological Sciences, 360(1456), 815–836. https://doi.org/10.1098/rstb.2005.1622
• Friston, K. J. (2009). The free-energy principle: a rough guide to the brain? Trends in Cognitive Sciences, 13(7), 293–301. https://doi.org/10.1016/j.tics.2009.04.005
• Friston, K. J. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787
• Friston, K. J., Harrison, L., & Penny, W. (2003). Dynamic causal modelling. NeuroImage, 19(4), 1273–1302. https://doi.org/10.1016/S1053-8119(03)00202-7
• Friston, K. J., Holmes, A. P., Worsley, K. J., Poline, J. P., Frith, C. D., & Frackowiak, R. S. (1994). Statistical parametric maps in functional imaging: a general linear approach. Human Brain Mapping, 2(4), 189–210. https://doi.org/10.1002/hbm.460020402
• Friston, K. J., Stephan, K. E., Montague, R., & Dolan, R. J. (2014). Computational psychiatry: the brain as a phantastic organ. The Lancet Psychiatry, 1(2), 148–158. https://doi.org/10.1016/S2215-0366(14)70275-5
• Friston, K. J., Levin, M., Sengupta, B., & Pezzulo, G. (2015). Knowing one’s place: a free-energy approach to pattern regulation. Journal of the Royal Society Interface, 12(105), 20141383. https://doi.org/10.1098/rsif.2014.1383
• Hohwy, J. (2013). The Predictive Mind. Oxford University Press.
• Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. https://doi.org/10.1038/4580
• Worsley, K. J., Marrett, S., Neelin, P., Vandal, A. C., Friston, K. J., & Evans, A. C. (1996). A unified statistical approach for determining significant signals in images of cerebral activation. Human Brain Mapping, 4(1), 58–73. https://doi.org/10.1002/(SICI)1097-0193(1996)4:1<58::AID-HBM4>3.0.CO;2-O