Biological Psychology 186 (2024) 108741

Contents lists available at ScienceDirect

Biological Psychology
journal homepage: www.elsevier.com/locate/biopsycho

Review

Active inference as a theory of sentient behavior

Giovanni Pezzulo a, *, Thomas Parr b, Karl Friston c, d
a
Institute of Cognitive Sciences and Technologies, National Research Council, Rome, Italy
b
Nuffield Department of Clinical Neurosciences, University of Oxford, UK
c
Wellcome Centre for Human Neuroimaging, Queen Square Institute of Neurology, University College London, London, UK
d
VERSES AI Research Lab, Los Angeles, CA 90016, USA

A R T I C L E I N F O A B S T R A C T

Keywords: This review paper offers an overview of the history and future of active inference—a unifying perspective on
Active inference action and perception. Active inference is based upon the idea that sentient behavior depends upon our brains’
Predictive coding implicit use of internal models to predict, infer, and direct action. Our focus is upon the conceptual roots and
Generative model
development of this theory of (basic) sentience and does not follow a rigid chronological narrative. We trace the
evolution from Helmholtzian ideas on unconscious inference, through to a contemporary understanding of action
and perception. In doing so, we touch upon related perspectives, the neural underpinnings of active inference,
and the opportunities for future development. Key steps in this development include the formulation of pre­
dictive coding models and related theories of neuronal message passing, the use of sequential models for
planning and policy optimization, and the importance of hierarchical (temporally) deep internal (i.e., generative
or world) models. Active inference has been used to account for aspects of anatomy and neurophysiology, to offer
theories of psychopathology in terms of aberrant precision control, and to unify extant psychological theories.
We anticipate further development in all these areas and note the exciting early work applying active inference
beyond neuroscience. This suggests a future not just in biology, but in robotics, machine learning, and artificial
intelligence.

1. Introduction neuroscientists interested in the neural circuits supporting predictions

and prediction errors (Bastos et al., 2012; Parr & Limanowski, Rawji,
Psychologists and neuroscientists are increasingly entertaining the et al., 2021; Parr & Friston, 2018; Walsh et al., 2020); psychologists
idea of the brain as a “prediction machine”, which learns an internal (i. interested in how we deal with uncertainty and cognitive effort during
e., generative) model of the lived world – and of the consequences of its decision-making (Parr et al., 2023; Rens et al., 2023), modelers inter­
actions – to make sense of sensations, predict how the current situation ested in the mechanisms of action-perception, exploration-exploitation
will unfold (i.e., learning and perception), and to act in a purposeful and higher cognition (Friston, FitzGerald, et al., 2017; Friston, Lin, et al.,
manner (i.e., action selection, exploration-exploitation, planning, et 2017; Pezzulo et al., 2015, 2018), clinicians interested in understanding
cetera). This idea appears in several guises, including the Bayesian brain, aberrant behavior in psychopathology (Maisto et al., 2021; Van den
the predictive brain, predictive processing, predictive coding, active inference Bergh et al., 2017), roboticists interested in self-supervised learning of
and the free energy principle, to name a few. world models and goal-directed behavior (Ahmadi & Tani, 2019; Tani­
Here, we critically review the origins, scope and impact of this idea, guchi et al., 2023), and neurophilosophers (Clark, 2015; Hohwy, 2013).
in fields like psychology and neuroscience. For conceptual clarity, we This breadth of application is appealing, but risks creating a frag­
focus specifically on active inference: a normative theory of sentient mented picture and some uncertainty about its original commitments
behavior that formalizes the “predictive brain” idea and provides a first- and conceptual implications. The aim of this brief manuscript is to help
principle account of its computational and neuronal processes (Parr researchers using (or interested in) predictive coding and active infer­
et al., 2022). ence to “connect the dots” and orient themselves within a growing
While active inference is still relatively young, it has a growing literature. Despite distinct lines of work — that emphasize different
impact across various disciplines. It is increasingly used by (for example) aspects of active inference — these applications all rest on the same core

* Correspondence to: Institute of Cognitive Sciences and technologies, National Research Council, Via S. Martino della Battaglia 44, 00185 Rome, Italy.
E-mail address: [email protected] (G. Pezzulo).

https://doi.org/10.1016/j.biopsycho.2023.108741
Received 18 July 2023; Received in revised form 5 December 2023; Accepted 29 December 2023
Available online 3 January 2024
0301-0511/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

principles. To foreground these core principles, we will look at the his­ out sensations that provide evidence for their continued existence.
torical and conceptual origins of active inference—to illustrate how its Intuitively, sensing our body temperature to be around 37 ◦ C offers more
core principles were introduced; then consider briefly how the scope of evidence that we are still alive than body temperatures far from this
active inference has expanded into several disciplines—and finally look value. The concept of autopoiesis gave birth to enactive approaches in
to future developments. Given the brevity of this treatment, we cannot philosophy (Maturana & Varela, 1980). From another angle, it has been
provide a full introduction to active inference. Rather, we provide an postulated that a central imperative for living organisms is maintenance
overview of the narrative in (Parr et al., 2022), which interested readers of physiological homeostasis (i.e., correction of deviations from
can consult. preferred physiological states through reflexive actions) and the regu­
In the next section, we briefly discuss the conceptual (and historical) lation of basic imperatives (Cannon, 1929) – but more modern theories
roots of active inference in early views of prediction and action-based emphasize that physiological regulation is fundamentally anticipatory
cognition. We then review some key developments of active inference, (i.e., allostatic) (Sterling, 2012). Various researchers have proposed that
by focusing on landmark papers that explain how it stems from a single closed-loop adaptive regulation (and not stimulus-response) is key to
principle (namely, free energy minimization). We next consider its scope understanding not just physiology but (potentially) all cognitive pro­
across perception, action, planning, etc. This brief review helps us make cessing (Cisek, 1999; Pezzulo & Cisek, 2016).
the point that active inference provides a unifying perspective on several Finally, another root is the idea that cognitive processes, such as
cognitive topics and theories and across levels of understanding, from learning, perception and decision-making, require an active engagement
conceptual to neural. Finally, we briefly highlight some promising of organisms with the environment. One early example of this action-
research directions that could expand the scope of active inference – and oriented perspective is the view of Gibson that perceiving things con­
potentially its impact on psychology and neuroscience. sists in seeing what to do or not to do with them, i.e., perceiving affor­
dances (Gibson, 1979). More recently, various researchers proposed the
2. The conceptual and historical roots of active inference necessity of a “pragmatic turn” in cognitive science and neuroscience –
and the need to recognize the importance of action as part and parcel of
Active inference has roots in various early theories in cognitive sci­ our cognition (Buzsaki, 2019; Cisek & Kalaska, 2010; Cisek &
ence (and beyond, in fields that would not necessarily use the label Pastor-Bernier, 2014; Engel et al., 2016; Lepora & Pezzulo, 2015;
“cognitive”). One root is the idea that the brain carries a small-scale O’Regan & Noe, 2001), rather than just a way to report “central” de­
model of the environment and uses it to mentally simulate what-if ac­ cisions, as assumed in conventional (serial) theories.
tions, instead of (or before) acting on the environment (Craik, 1943). Interestingly, each of these ideas implies a shift from reactive to
This idea is foundational in cognitive science. For example, (Tolman, predictive, enactive views of the brain. While a reactive brain waits for
1948) proposed that humans, rodents and other animals find their way incoming stimuli, a predictive and active brain predicts external events
in a maze by first learning a mental model or “cognitive map”, rather (e.g., predictive coding) and actively gathers evidence (i.e., active
than by considering which of their navigation actions were previously sensing and active learning) to make sense of the world. While a reactive
rewarded the most, as assumed by behaviorist formulations. brain selects actions based on the past and present (e.g., the history of
Another root is the idea of (Helmholtz, 1866) that perception is an reinforcement and the current cue), a predictive brain actively imagines
(unconscious) inference based on an internal generative model – that its preferred future and then makes this happen by acting (e.g., acts in a
uses recurrent (top-down and bottom-up) counter-streams of processing, goal-directed manner). While a reactive brain maintains homeostasis, a
rather than bottom-up transduction of external sensations into internal predictive brain acts to anticipate needs and performs anticipatory
representations (and later actions). This idea was later developed in regulatory (or allostatic) actions.
psychology (Gregory, 1968, 1980) and computational neuroscience; All these (and other) views contributed to raising the importance of
giving rise to the “Bayesian brain” hypothesis (Doya et al., 2007) and to predictive and enactive views of the brain and of cognition. However,
formulations of predictive coding as a possible neurobiological imple­ each of these perspectives were somewhat disconnected from one
mentation of perception-as-inference in the brain (Friston, 2005; Rao & another and linked to different research traditions, which are sometimes
Ballard, 1999). Beyond perception, other cognitive functions were later seen as conflicting with one another (e.g., the Helmholtzian and the
described in terms of inference, i.e., planning-as-inference (Botvinick & Gibsonian traditions). One benefit of active inference is that it helps
Toussaint, 2012). unify and thereby advance these traditions, as we explain in the
Yet another “root” is the idea of cyberneticists (Miller et al., 1960; following Sections.
Powers, 1973; Wiener, 1948) that goal-directed action proceeds by
firstly setting up a desired state or observation (e.g., feeling warm), then 3. The normative perspective of active inference – and how it
monitoring the discrepancy – now referred to as a “prediction error” – has developed
between the preferred and sensed state (e.g., feeling excessively warm),
and then selecting a course of action that reduces this discrepancy – Active inference provides a normative perspective that unifies and
where “action” is a suitcase word and can include any means to exert advances the predictive and enactive views of brain and behavior. It
control over external stimuli; ranging from simple autonomic reflexes (e. does so by highlighting that several apparently disconnected accounts –
g. thermoregulation) to sophisticated plans (e.g., visiting one’s favorite identified by early theories – stem parsimoniously from the assumption
ice cream shop). A key result in this field – which coheres with the that living organisms obey a single imperative: namely, they act to
Helmholtzian perspective above – is the ‘Good regulator theorem’ of minimize their surprise,1 or more formally, their variational free energy.
(Conant & Ashby, 1970), which argues that effective regulatory systems The mathematics of variational free energy minimization is beyond
must [be a] model the environment they regulate. In a similar vein, in the scope of this article; we suggest to the interested readers to consult
psychology, ideomotor theory proposed that action control is essentially (Parr et al., 2022). Here, instead, we introduce the key concepts of the
anticipatory and that action are selected and controlled by their antic­ theory, by briefly reviewing (non-chronologically) selected landmark
ipated consequences or outcomes, not through stimulus-response papers and linking them to the early theories.
(Hoffmann, 2003; Hommel, 2003; James, 1890). Active inference starts from a simple consideration: that to maintain
Besides cybernetics, there are other influential views that highlight
the centrality of adaptive regulation for behavior and life itself. One
example is the idea that living organisms are autopoietic systems, which 1
Technically, surprise here refers to self-information (a.k.a., surprisal);
create the conditions for their own existence. More recently, this idea namely, the implausibility of some (sensory) outcome under a (generative)
has been framed as ‘self-evidencing’ (Hohwy 2016) – i.e., creatures seek model of how that outcome was generated.

2
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

Fig. 1. Generative model and generative process in active inference. This

Figure—reproduced from (Parr et al., 2022)—illustrates the structure of active
inferential theories of brain function. Our worlds evolve according to some
Fig. 2. Perception and action play complementary roles in the minimi­
dynamical process that generates observations (y) from hidden states (x * ). Our
zation of variational free energy. This Figure— reproduced from (Parr et al.,
internal models account for observations in terms of hypothetical hidden states
2022)—highlights the relationship between action and perception via free en­
(x). Our inferences about these states based upon our observations then drive
ergy (F). Perception involves minimizing free energy by changing our beliefs
actions (u) that intervene on the processes generating our sensations.
(Q) about states (x). This effectively minimizes the divergence (DKL) between
our beliefs and the probability of these states given sensory data (y). Action
their existence and integrity, all living organisms need to remain in a minimizes free energy through changing those parts of the free energy that
bounded set of characteristic states that basically define their place depend upon sensory data—notably, the evidence or probability of data under
within an ecological niche; for example, a fish cannot live out of water. our internal model.
Using the lexicon of Bayesian inference, being out of water for a fish is a
“surprising” state. Clearly a fish should avoid this surprise, and the idea
generalizes to suggest that living organisms must avoid surprising states
(Friston et al., 2010). If they did not, they would not be living organisms
for long. Another way of looking at this is that everything (including me)
is defined by being in some characteristic (attracting) set of states.
Conversely, I am defined by the kinds of states I cannot be in. These are
surprising states.
A computationally tractable solution to surprise minimization is the
minimization of an information-theoretic quantity – variational free
energy – which is a function of two things: a generative model (i.e., a
statistical model that describes how sensations are generated) and
observed sensory data. This implies that a living organism must be
equipped with a generative model – or in the lexicon of (Craik, 1943), a
small-scale model – to predict the sensations generated by the world
(and by the organism’s place in it). In Bayesian terms, a generative
model comprises two things: a prior over the hidden (i.e., unobserved) Fig. 3. The architecture of predictive coding. This Figure—reproduced from
variables of interest and a likelihood function that maps the hidden var­ (Parr et al., 2022)—shows the message passing between populations of neurons
iables to observables (Bishop, 2006). See Fig. 1 for a schematic illus­ under a predictive coding scheme as it might manifest in the layers of the ce­
tration of the organism’s generative model of the world and its relation rebral cortex (separated into superficial layers I-III, layer IV, and deep layers
with the generative process: the true environmental contingencies that V-VI). This shows predictions based upon expectations (μ) being subtracted
generate its observations, which is inaccessible to the organism. from ascending signals to compute errors (ε), which are used to update ex­
Put simply, an organism can minimize variational free energy by pectations. The subscripts indicate whether we are dealing with fast changing
aligning the predictions of its generative model and the data it observes. dynamical variables (x) or more slowly changing contextual variables (v) which
act to link together different hierarchical levels, with hierarchy indicated by the
In different settings, this minimization has been described in various
bracketed superscripts. As we ascend the hierarchy, the variables we deal with
ways, such as the minimization of surprise, of prediction errors, or of the
become slower, such that the contextual variables at one hierarchical level
discrepancy between the model and the world. All of these are equiva­ evolve over the same timescale as the dynamical variables at the level above.
lent to the minimization of variational free energy under specific sets of
assumptions.
minimization. In formal terms, they map to its two components: the
Interestingly, aligning the predictions derived from a generative
minimization of divergence and the maximization of evidence, see
model and data can be achieved in two ways: by changing the model
Fig. 2. Recognizing that action and perception can be unified within a
predictions and by changing the observed data. The former corresponds
single formal imperative – the minimization of free energy – is one of the
to revising the agent’s beliefs (used in the technical sense of probability
key innovations of active inference, which helps integrate and extend
distributions over hidden variables) if they do not explain the data well.
the early theories reviewed above.
This is exactly the inferential view of perception of (Helmholtz, 1866).
Regarding neural implementation, one of the most widely enter­
The latter corresponds to acting in the world to change the data that will
tained hypotheses – about how the brain might implement perceptual
be sampled next – to render them more like the organism’s prior pre­
inference – is predictive coding (Rao & Ballard, 1999). Fig. 3 shows the
dictions. This latter perspective on action – and on its dependence on
architecture of a predictive coding scheme as it might manifest in the
expected outcomes – is highly congruent with cybernetics (Miller et al.,
cerebral cortex. In this predictive coding network, inference is realized
1960; Powers, 1973; Wiener, 1948) and ideomotor theory (Hoffmann,
by propagating predictions and prediction errors through top-down and
2003; Hommel, 2003; James, 1890).
bottom-up pathways, respectively, and by minimizing prediction errors
In sum, changing beliefs about the causes of data (i.e., perception)
across all levels. Interesting, predictive coding can be derived as a spe­
and changing the data (i.e., action) are two aspects of free energy
cial case of variational free energy minimization (Friston, 2005).

3
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

the consequences of different policies, scoring the policies according to

how much they are expected to minimize free energy in the future and
then (with some simplifications) select the best policy.
This planning process induces a novel quantity – expected free energy
– that is the functional that active inference uses to evaluate (and assign
a prior to) policies and it is distinct from the notion of variational free
energy discussed so far (Friston et al., 2017). The notion of expected free
energy has been very useful in the development of active inference
models of things like (bounded) decision-making, planning,
exploration-exploitation and curiosity (Friston, Lin, et al., 2017; Parr &
Pezzulo, 2021; Schwartenbeck et al., 2019). This is because this notion is
richer than the common optimization objectives used in other formal
frameworks (e.g., economic theory and reinforcement learning). This is
because expected free energy considers jointly a pragmatic imperative
(utility maximization) and an epistemic imperative (information gain, or
the resolution of the uncertainty). Indeed, as Fig. 4 illustrates, it is
Fig. 4. Expected free energy and the way it can be mapped to different possible to map expected free energy to various other formal notions (e.
formal notions (e.g., Bayesian surprise, Risk-sensitive control, Expected g., Bayesian surprise, Risk-sensitive control, Expected utility theory), by
utility theory) by removing one or more terms, denoted with numbers. removing one or more of its terms.
This Figure—reproduced from (Parr et al., 2022)—expresses expected free
Active inference is a general scheme that can be applied to address
energy in terms of beliefs about trajectories (indicated by the tilde ~). The
various cognitive processes. Crucially, the functioning of active infer­
additional symbols here, not in previous figures, are the π for policies and the C
for preferences. Note that some terms (including term 1) are expressed in terms ence is the same across all problems: what differs is the generative
of expectations—i.e., averages under the subscripted distribution. model, which is task specific. This implies that by designing the
appropriate generative models, it is possible to address a variety of
cognitive tasks with the same approach – and to pass from the normative
While predictive coding is a model of perception, it can be readily
perspective of active inference to specific implementations that have
extended to encompass the role of action in the minimization of free
energy (described above). The move from predictive coding to active biological plausibility (Friston, Parr, et al., 2017; Parr & Friston, 2018).
Here, a worked example may be helpful. To illustrate some of the
inference can be realized by equipping predictive coding networks with
simple motor reflexes. In this perspective, the motor system works by principles we have outlined so far, we will consider how we might go
about developing a model for a ubiquitous task in cognitive neuro­
generating proprioceptive predictions (in the same way standard pre­
science—a delay period oculomotor task. This is a relatively simple task
dictive coding generates exteroceptive predictions) — and not motor
that can be performed by humans—and some animals—and that is
commands, as conventionally proposed – and these proprioceptive
designed to probe working memory function (Funahashi et al., 1989).
predictions are realized through the motor reflexes (Adams et al., 2013).
The task sequence is as follows. First, a cross is presented on screen and
Subsequently, this theory was extended to also model autonomic
our subject maintains fixation on this cross. A target then appears at one
control (Barrett & Simmons, 2015; Pezzulo, 2014; Seth et al., 2012). The
of several possible locations towards the periphery of the screen, but our
general idea is that autonomic control might work by generating inter­
subject still maintains fixation. The target then disappears and, after a
oceptive predictions (i.e., homeostatic setpoints) and then fulfilling
‘delay period’, a stimulus appears to signify that the subject should make
them through autonomic reflexes, in much the same way motor control
a saccadic eye movement to the location of the target. Successful per­
might work by generating proprioceptive predictions and then fulfilling
formance of this task relies upon retaining a memory of the target
them through motor reflexes. This development of active inference helps
location during the delay and response phases.
connect it with theories of allostatic control (Sterling, 2012) and paves
To model this task, we must consider the data available to the sub­
the way to a better understanding of our ability to model and control the
ject. In this case, these are the visual stimuli and proprioceptive inputs,
internal milieu, not just the external environment. This stream of
and whether the correct action was chosen. To do so, we need to take
research underwrote novel approaches to psychopathology – as deficits
account of the causes of these data. The causes of proprioceptive data are
of interoceptive processing (Paulus et al., 2019).
simply the direction in which our subject’s eyes are pointing. Visual
So far, we have discussed active inference using generative models
outcomes, depend upon a combination of (1) gaze direction, (2) the
that characterize processes that unfold in continuous time (e.g., pre­
intended target location, and (3) the current stage of the task (i.e., the
dictive coding networks) and use continuous variables (i.e., the formal
fixation, target presentation, delay, or response stage). For each of these
framework of dynamical systems and state-space models). However,
three variables, we must then specify how we expect them to evolve
many cognitive problems can be characterized at a distinct level: as
throughout the task. The gaze direction will transition from one step to
(sequences of) discrete decisions. These include problems that require
the next based upon the decisions our subject makes. The intended
the selection of discrete responses during psychology experiments, the
target location will be fixed (although initially unknown) throughout
targets for saccades, or navigational trajectories in discretized environ­
the task. The task stage evolves predictably through a sequence of steps.
ments (Friston et al., 2017; Friston, Lin, et al., 2017). These problems
Together, these beliefs about the way in which data are generated and
can be modeled in active inference, using generative models that use
the dynamics of the causes allow our subject to predict what will be
discrete variables (and the formal framework of Partially Observable
observed next, and to update these beliefs when these predictions are
Markov Decision Processes).
violated.
In addition to the two aforementioned components (priors and
As outlined above, active inference equips models with prior beliefs
likelihood function), the generative models for active inference in
about the relative plausibility of different choices based upon their
discrete time often include a third component: the transition function,
relative expected free energies. In this model, the key part of the ex­
which describes the way in which hidden states change depending upon
the agent’s actions (or sequences of actions, called policies). Crucially, pected free energy is a preference for receiving the ‘correct’ feedback
outcome which is only available during the response phase of the task
these generative models have temporal depth and afford a novel capa­
(see (Mirza et al., 2016) for a similar setup in the context of scene cat­
bility that was not available in simpler models: namely, planning. In
egorisation, in which the main role of the expected free energy is to
simple terms, planning involves using the generative model to predict
promote information seeking). It is this that prompts a saccade to the

4
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

at different timescales. One example is the model of active listening

processes, in which (for example) lower hierarchical levels deal with
words and higher levels deal with sentences (Friston et al., 2021).
Another example is a model of hierarchical action recognition that
recognizes actions at different levels, from low level kinematics to
higher level goals and intentions (Proietti et al., 2023). It is also possible
to use hierarchical models to model hierarchies of control, in which
lower-to-higher levels deal with autonomic imperatives (e.g., ensure a
correct basic temperature) in increasingly complex ways (e.g., from
thermoregulation to the goal-directed plan to buy water before a long
run) (Pezzulo et al., 2015; Tschantz et al., 2021). These developments –
from simple to more sophisticated (e.g., hierarchically and temporally
deep) generative models has extended the range of cognitive models that
have been addressed using active inference over the years.
Another interesting realization is the fact that it is possible to derive a
biologically motivated “process theory” for active inference in discrete
time, by interpreting the specific operations (variational updates)
required to minimize free energy as signals that are computed or
exchanged across neurons (Friston et al., 2017). This is important
because it permits crossing levels of explanation – from normative to
mechanistic and neuronal – and to use active inference to simulate
neuronal activity that would ensue from the performance of cognitive
tasks (Friston, Parr, et al., 2017; Parr & Friston, 2018).
Another important development of active inference regards precision
control and its role in psychopathologies. In predictive coding, variables
are encoded as Gaussian distributions and precision simply refers to the
inverse of the variance of a distribution (Friston, 2005). Precision con­
trol refers to a mechanism that optimizes the precision of (the distri­
bution of) each variable of an organism’s generative model. It is
important since it regulates the relative importance of top-down pre­
dictions and bottom-up prediction errors across the hierarchy. This is
because prediction errors that are assigned greater (lower) precision
have greater (lower) impact on the belief updating and the ensuing
inference. Veridical inference requires the precision of (the distribution
of) each variable to be optimized, to reflect the signal-to-noise ratio of
Fig. 5. A simulated oculomotor delay period task. This figure, taken from sensory signals – therefore highlighting a link between precision control
(Parr & Friston, 2019b) (published under a CC BY 4.0 license), shows simulated and attention as gain control (Parr & Friston, 2019a) – or the importance
performance of a simple working memory task under active inference. Although of an organism’s prior preferences – reflecting the fact that an organ­
simple, this task calls for planning (of our next saccade), recall (of the target
ism’s innate drives or goal states can be encoded as highly precise priors
location), and movement execution. The upper left images show a series of
(Pezzulo et al., 2015). Interestingly, when precision control fails, it can
frames taken from the simulation, as if we were observing our participant’s
eyes. The black arrows link these behavioural responses to the view of the produce excessively rigid forms of inference (when priors fail to be
stimulus screen from the time of the target (red) presentation to the response updated in the light of novel evidence) or excessive sensitivity to stimuli
phase. A series of black dots show the (cumulative) trajectory of gaze direction. (when belief revision follows the sensory input and its random fluctu­
Because this model is formulated to have both continuous (prediction-error ations too closely—i.e., it overfits). These forms of aberrant inference,
minimising) and discrete (sequential planning) parts, we can plot the trajectory which depend sensitively on predicted precision, have been adopted to
both in terms of position and velocity (lower left) and in terms of the sequence explain several psychopathological conditions, such as delusions,
of actions taken. depression, psychosis, and many others (Barrett et al., 2016; Corlett &
Fletcher, 2015; Edwards et al., 2012). In turn, these theories also speak
remembered target location. Finally, the predicted action must be to aberrant neuromodulation, since the precision of (the distribution of)
executed. This depends upon resolving the error between the anticipated different variables might be encoded by different neuromodulators, e.g.,
proprioceptive information given the inferred saccade and current acetylcholine for the precision of the likelihood, noradrenaline for the
proprioceptive input. The result is the sequence of steps shown in Fig. 5. precision of transitions, dopamine for the precision of policies, etc. (Parr
The oculomotor control example illustrates how active inference can & Friston, 2018).
be concretely applied to study cognitive tasks, by designing (or learning) Yet another development regards the analysis of generative models
the appropriate generative models. Generative models represent formal during sleep or other ‘offline’ periods. It has long been hypothesized that
hypotheses about how cognitive tasks are accomplished – hypotheses learning generative models benefits from alternating on-line and off-line
that can be validated with empirical data. A useful illustration of the periods (Hinton et al., 1995). While on-line generative modelling max­
design principles to realize (or train) generative models for different imises accuracy (under complexity constraints), during off-line activity –
cognitive problems is provided in (Parr et al., 2022). This treatment in the absence of sensory data to “explain away” – model optimisation
makes a distinction between generative models in continuous time (that can focus on minimising complexity; for example, by removing redun­
are useful to address motor control tasks) and discrete time (that are dant parameters (Friston, Lin, et al., 2017; Pezzulo et al., 2021). From a
useful to address decision and planning tasks) and explains how these neuronal perspective, generative modelling during offline periods could
two types can be combined to form so-called hybrid or mixed generative be associated with (generative) replay activity in the hippocampus, the
models, in which discrete-time models are placed on top of prefrontal cortex and other brain areas; but these links remain to be fully
continuous-time models. Furthermore, the generative models of active established (Foster, 2017; Schwartenbeck et al., 2023; Stoianov et al.,
inference can be extended hierarchically, to model processes that unfold 2022).

5
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

Finally, an interesting development regards the realization of active Friston, 2010). Oscillatory dynamics that are ubiquitous (and that often
inference in which the free energy minimization extends “beyond the occur in synchrony) both within and across brain area might be signa­
skull”, to model the ways multiple active inference agents engage in tures of temporal prediction and of the exchange of top-down and
cooperative or competitive tasks (Friston & Frith, 2015; Maisto et al., bottom-up information across hierarchical levels of the brain’s genera­
2023) or construct their own niches (Constant et al., 2022). These and tive model (Arnal & Giraud, 2012).
other works illustrate that the concept of free energy minimization can Second, active inference suggests that cognitive functions – usually
readily extend to multi-agent settings – including settings that go addressed in isolation – might be instead better understood by appealing
beyond the standard scope of cognitive science, such as morphogenesis to a unique process theory. For example, in prominent computational
(Friston & Levin, Sengupta, et al., 2015) and autopoiesis (Friston, 2013) neuroscience theories, perception and action are two separate functions
– and hence potentially shed light on the relations between multiple with different objectives and neural substrates. According to Bayesian
nested levels of (self-)organization, from individual to social and cul­ decision theory (Robert, 2007), the goal of perception is to provide an
tural levels. accurate estimate of the agent’s state, whereas the goal of action se­
In sum, we have highlighted various developments of active infer­ lection is to maximize its expected utility. The former process is a
ence, which encompass the complementary roles of perception and ac­ precondition for the latter, implying an outdated, serial view of cogni­
tion in minimizing an organism’s variational free energy (and ensuring tive processing. Active inference holds that perception and action
that it successfully avoids “surprising” and characteristic states), the cooperate to minimize free energy, by minimizing divergence and
proposal of biologically plausible architectures for continuous time maximizing evidence, respectively (Parr et al., 2022). Another example
predictive coding and action control, the realization of generative is the fact that in 20th-century cognitive science, working memory was
models for discrete decisions that afford planning and the minimization considered as a separate storage that can be assessed by other compo­
of expected free energy, the hierarchical extension of these models, the nents when needed; therefore, imposing a separation between infor­
importance of precision control, and beyond. For each of these topics, mation storage and information processing. In contrast, active inference
we have cited some selected papers that the interested readers might models of hierarchical perception and action (Friston et al., 2021; Pez­
want to consult for more detailed information. Clearly, this is not an zulo et al., 2018) treat memory of the previous state as intrinsic to the
exhaustive list, but each of these developments has been useful to belief updating under generative or world models, across multiple
develop models of increasingly complex cognitive and social functions; timescales, which is in keeping with 21st-century accounts of working
see (Parr et al., 2022) for a more exhaustive treatment of active memory (Hasson et al., 2015).
inference. Third, active inference has the potential to unify different “levels of
understanding” of cognitive processes. Marr famously introduced a
4. The benefits of unification distinction between computational, algorithmic and neural imple­
mentation levels and argued that progress can be made within each level
In the previous Section, we saw that the scope of active inference and by connecting different levels (Marr, 1982). Establishing links be­
touches several domains of psychology and neuroscience. Here, we tween theories that operate at different levels is often challenging.
foreground a benefit of this rapid expansion: namely, unification. Active inference helps establish firm relations across levels of descrip­
Arguably, a main goal of cognitive psychology and neuroscience is tion. Rather than Marr’s tripartite distinction, in active inference it is
explaining behavior and its neural foundations, in a comprehensive (if more common to appeal to a distinction between normative theory and
not a unified) way. Yet, to ensure methodological rigor, these disciplines process theory (Friston et al., 2017). Free energy minimization is the
usually adopt restricted laboratory settings that tend to isolate cognitive normative objective of living organisms, whereas predictive coding and
functions and obfuscate their relations (Maselli et al., 2023). Consider variational message passing are process-level theories that describe how
for example a mundane task that we solve almost every day: crossing a the brain might support free energy minimization. Importantly, as
busy road. Even this relatively simple task engages several cognitive shown by (Friston, 2005), under certain assumptions predictive coding
processes in a coordinated manner, such as perception (of the situation), can be directly derived by the minimization of variational free energy,
memory (of past street crossing episodes), planning and action selection connecting the two levels of explanation. A similar case can be made for
(of the best route), motivation (and the “why” of crossing), attention (to the variational message passing schemes proposed to support discrete
select the most relevant stimuli), etc. These processes are often studied active inference in neural circuits (Friston et al., 2017).
in isolation using different paradigms leading to a proliferation of hy­ Fourth, unification endows existing constructs with validity, via the
pothesis and theories that assign each of them a distinct computational application of active inference across domains. One example is the
objective (and perhaps brain area) – therefore determining a very development of theories of interoceptive inference and autonomic
fragmented theoretical landscape. control (Barrett & Simmons, 2015; Pezzulo, 2014; Seth et al., 2012) by
Active inference proceeds the other way around: it starts from a analogy with the functioning of action control (Adams et al., 2013). In
single principle and asks how far one can go with it. And to what extent this perspective, autonomic control works exactly like action control –
it is possible to derive from that principle empirically testable hypoth­ namely, it aims to minimize a discrepancy between a predicted and a
eses about behavior and its cognitive and neural mechanisms? This sensed signal – except that the “signal” refers to interoceptive streams
approach brings the benefits of unification, in at least six ways. rather than proprioceptive streams. Another example can be found in
First, active inference assumes that everything, from perception to computational psychiatry, where numerous accounts of psychopathol­
action selection and learning ultimately serves to minimize variational ogy appeal to a single mechanism: namely, aberrant precision control.
free energy. A consequence of this is that one can align the (sometimes Fifth, active inference has the potential to reconcile (or at least to
vague) conceptual terms used in psychology with crisp formal terms of contextualize) theoretical perspectives that have long been considered
free energy minimization. For example, one can assign things like at odds in psychology, neuroscience and philosophy. One example is the
attention to precision control. At the neuronal level, the fast updates – Helmholtzian view that perception constitutes an inference about the
mediated by synaptic activity – might correspond to inferential pro­ entities of the external world that cause our sensations (Helmholtz,
cesses that minimize free energy at a fast time scale, whereas the slower 1866) and the Gibsonian view that perceiving consists in seeing action
updates – at the level of synaptic efficacy – might correspond to learning opportunities and affordances, not reconstructing a model of the
processes that minimize free energy at a slower timescale. Precision external reality within the brain (Gibson, 1979). This apparent dialectic
dynamics might correspond to the activity of neuromodulators, which could be dissolved by considering that there are multiple ways to design
finesse the inference at multiple levels, for example, by increasing the generative models; specifically, a relevant distinction is between
post-synaptic gain of sensory or prediction error-units (Feldman & generative models that explicitly model the ways external states produce

6
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

Box 1
Glossary of technical terms.

Active Inference: A normative framework that elucidates the neural and cognitive processes underlying sentient behavior, beginning with first
principles. This framework posits that perception and action work in concert to minimize a shared functional known as variational free energy.
Expected Free Energy: This is the quantity that is used in active inference to score action sequences or policies (and then to select between
them). It takes into consideration both the pragmatic value of policies – or how close a policy’s expected outcomes are to the preferred outcomes
– and their epistemic value (or information gain) – or how much the policy is expected to reduce uncertainty.
Generative Model: A statistical model designed to explain the generation of observable content from unobservable, hidden (latent) causes. For
instance, it clarifies the process by which a visual object gives rise to an image on the retina. Generative models serve a dual purpose: they allow
the generation of novel, synthetic content and support the inference of hidden causes from observable data. From a technical standpoint,
generative models encode the joint probability distribution governing both observables and hidden causes.
Latent (or Hidden) Variable: An internal variable within a generative model, referred to as "latent" or "hidden" due to the fact that it cannot be
directly observed, but must be inferred.
Precision and precision-weighting: Precision denotes the inverse of variance or standard error, serving as a measure of the reliability or
certainty associated with sensory information. Precision-weighting refers to the fact that in predictive coding and active inference, prediction
errors are weighted by their respective precisions, therefore determining the extent to which sensory observations influence the process of
updating beliefs.
Predictive Coding: A computational framework in neuroscience that provides a possible neural implementation for the idea that perception
consists in a process of inference. In hierarchical predictive coding networks, inference is realized by minimizing (precision-weighted) pre­
diction errors across all hierarchical levels. In turn, this requires bidirectional loops between top-down processes (conveying predictions) and
bottom-up processes (conveying prediction errors).
Variational Free Energy: This is the functional (function of a function) that is minimized within the framework of active inference. It is also
widely utilized in utilized in probabilistic modeling, statistical inference and machine learning. In its simplest instantiation, it corresponds to a
summation of prediction errors, which quantifies the deviation of observed data from the predictions of the generative model. More formally,
variational free energy serves as an upper bound on the negative logarithm of the evidence, which is the probability of observed data given a
model.

sensations (a.k.a., environmental models) or the ways actions produce the requisite interactions. An open question for future research is
sensations (a.k.a., sensorimotor models) (Sims & Pezzulo, 2021; Pezzulo whether the enactive and embodied approach of active inference has the
et al., 2023). Some active inference studies use generative models that potential to complement and advance the development and deployment
include explicit beliefs about entities in the external world that cause of Generative AI.
sensations, such as one’s location in space (Friston et al., 2017). Other
active inference studies use generative models that only consider the 5. Opportunities for the future
sensory consequences of one’s action, such as touch sensations that
follow whisking at a given amplitude, but not explicit beliefs about It’s Difficult To Make Predictions, Especially About the Future. Niels
objects ‘out there’ (Mannella et al., 2021). The latter generative models Bohr.
adhere more closely to the notions of affordance (Gibson, 1979) and of The compass of active inference is expanding rapidly, but the land­
sensorimotor contingency (O’Regan & Noe, 2001), despite the fact they scape of future opportunities may be even ampler. Here, we focus on
still entail inferential dynamics. Besides this specific topic, there is a some of the developments that we consider most promising and most
vivid debate in philosophy that concerns the most appropriate way to likely in the near future.
consider active inference, in relation to internalist (Hohwy, 2013), The first and perhaps most obvious direction for the future regards a
externalist (Clark, 2013) or enactivist theories (Bruineberg et al., 2018). deeper empirical scrutiny of active inference. A question that is some­
Finally, and importantly, the integrative perspective of active infer­ times asked of active inference is whether any empirical findings could
ence could be valuable in characterising of sentient behaviour – offer evidence for or against the framework. This can be a vexed ques­
considered here to be the capacity to infer states of the world and to act tion to answer as it constitutes a category error. A framework is not in
upon it with a sense of purpose (Friston, Da Costa, et al., 2023). This itself a hypothesis. It is a way of formulating hypotheses. The relation­
operational definition is satisfied by active inference when, and only ship between active inference and empirical psychology is that we can
when the generative model includes the consequences of action formalize psychological theories in terms of the generative models that
(mathematically, when the generative model includes priors over pol­ underwrite neurophysiological and behavioural responses. Equipped
icies based upon expected free energy). This notion of sentience is does with a proposed model, the framework can be used to express a hy­
not have any phenomenological commitments and is probably best read pothesis, to predict the behaviour expected under that hypothesis, and
as ‘basic sentience’ in the sense of (Clark, 2023). to fit to measured data to formally compare alternative hypotheses. In
Recently, there has been a proliferation of advanced Generative AI other words, while active inference is an application of the free energy
systems that process language, images and videos with very high accu­ principle – which is a principle (i.e., method) rather than a theory
racy. However, in most cases, these systems learn passively from large (Friston, 2010) – theories tested under the active inference framework
predefined datasets and disregard agency – and the possibility to act (e.g., those considered in this article) make specific empirical pre­
upon the world with a purpose – to develop genuine understanding dictions that can (and need to) be empirically validated. One example of
(Pezzulo et al., 2023). Active inference suggests a different path to un­ this is the oculomotor delay period model shown in Fig. 5, which
derstand and simulate sentient behaviour, which focuses on the devel­ generate empirically testable predictions about oculomotor perfor­
opment of grounded world (i.e., generative) models, by actively mance as a function of varying delay periods (Parr & Friston, 2019b).
engaging with the environment and by predicting the consequences of Various empirical studies are already addressing the empirical

7
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

predictions of predictive coding, such as how top-down and bottom-up Barrett, L. F., Quigley, K. S., & Hamilton, P. (2016). An active inference theory of
allostasis and interoception in depression. Philosophical Transactions of the Royal
dynamics support predictions and prediction errors, respectively
Society B: Biological Sciences, 371(1708), 20160011. https://doi.org/10.1098/
(Walsh et al., 2020). However, active inference makes a number of rstb.2016.0011
specific predictions about (for example) the way the motor system works Bastos, A. M., Usrey, W. M., Adams, R. A., Mangun, G. R., Fries, P., & Friston, K. J.
(Shipp et al., 2013) and the way higher cognitive functions are imple­ (2012). Canonical microcircuits for predictive coding. Neuron, 76(4), 695–711.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.
mented (Pezzulo et al., 2018) that differ from mainstream theories and Botvinick, M., & Toussaint, M. (2012). Planning as inference. Trends in Cognitive Sciences,
could be increasingly scrutinized by future studies. 16(10), 485–488. https://doi.org/10.1016/j.tics.2012.08.006
A second interesting direction for the future is assessing to what Bruineberg, J., Kiverstein, J., & Rietveld, E. (2018). The anticipating brain is not a
scientist: The free-energy principle from an ecological-enactive perspective. Synthese,
extent active inference – and more broadly, the free energy principle – 195(6), 2417–2444. https://doi.org/10.1007/s11229-016-1239-1
can help us understand the evolution of complex neural circuits and life Buzsaki, G. (2019). USA. The brain from inside out. Oxford University Press.
forms from simpler ones. Active inference suggests a possible path from Cannon, W. B. (1929). Organization for physiological homeostasis. Physiological Reviews,
9(3), 399–431.
the simple mechanisms that supported prediction and control in our Cisek, P. (1999). Beyond the computer metaphor: behaviour as interaction. Journal of
earlier evolutionary ancestors to the more sophisticated abilities of our Consciousness Studies, 6(11–12), 11–12.
species (Pezzulo et al., 2022), but a comprehensive account of the Cisek, P. (2019). Resynthesizing behavior through phylogenetic refinement. Attention,
Perception, & Psychophysics, 81(7), 2265–2287. https://doi.org/10.3758/s13414-
evolution and “phylogenetic refinement” (Cisek, 2019) of living organ­ 019-01760-1
isms remains to be fully developed (Friston et al., 2023; Friston, Fried­ Cisek, P., & Kalaska, J. F. (2010). Neural mechanisms for interacting with a world full of
man, et al., 2023). action choices. Annual Review of Neu, 33, 269–298. https://doi.org/10.1146/
annurev.neuro.051508.135409
A third interesting direction for the future regards the realization of
Cisek, P., & Pastor-Bernier, A. (2014). On the challenges and mechanisms of embodied
advanced artefacts, such as AIs and robots, based on active inference. decisions. Philosophical Transactions of the Royal Society B.
There have already been several successful robotic implementations of Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of
active inference, but the full potential of the theory has not yet been cognitive science. Behavioral and Brain Sciences, 36(03), 181–204. https://doi.org/
10.1017/S0140525X12000477
reached (Ahmadi & Tani, 2019; Lanillos et al., 2021; Priorelli et al., Clark, A. (2015). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford
2023; Taniguchi et al., 2023). Interestingly, some of the central concepts University Press, Incorporated.
of active inference, such as the importance of generative models and Clark, A. (2023). The experience machine: How our minds predict and shape reality.
Conant, R. C., & Ashby, W. R. (1970). Every good regulator of a system must be a model
self-supervised, predictive learning, are becoming central in mainstream of that system. International Journal of Systems Science, 89–97.
research in AI, as testified by the recent successes in generative AIs such Constant, A., Clark, A., Kirchhoff, M., & Friston, K. J. (2022). Extended active inference:
as large language models. This creates an important opportunity, since Constructing predictive cognition beyond skulls. Mind & Language, 37(3), 373–394.
https://doi.org/10.1111/mila.12330
(apart for their obvious technological impact), state-of-the-art AI sys­ Corlett, P. R., & Fletcher, P. C. (2015). Delusions and prediction error: Clarifying the
tems can be precious in advancing our understanding of living organ­ roles of behavioural and brain responses. Cognitive Neuropsychiatry, 20(2), 95–105.
isms, providing that they incorporate appropriate (design) principles https://doi.org/10.1080/13546805.2014.990625
Craik, K. (1943). The Nature of Explanation. Cambridge University Press.
(Pezzulo et al., 2023). Doya, K., Ishii, S., Pouget, A., & Rao, R. P. N. (Eds.). (2007). Bayesian Brain: Probabilistic
Approaches to Neural Coding (1st ed..,). The MIT Press. http://www.amazon.com/
Funding and acknowledgements exe℅bidos/redirect?tag=citeulike07-20&path=ASIN/026204238X.
Edwards, M. J., Adams, R. A., Brown, H., Pareés, I., & Friston, K. J. (2012). A Bayesian
account of “hysteria. Brain: A Journal of Neurology, 135(Pt 11), 3495–3512. https://
This research received funding from the European Union’s Horizon doi.org/10.1093/brain/aws129
2020 Framework Programme for Research and Innovation under the Engel, A. K., Friston, K. J., & Kragic, D. (2016). The Pragmatic Turn: Toward Action-
Oriented Views in Cognitive Science. MIT Press.
Specific Grant Agreements No. 945539 (Human Brain Project SGA3) to
Feldman, H., & Friston, K. J. (2010). Attention, uncertainty, and free-energy. Frontiers in
GP and KF and No. 952215 (TAILOR) to GP; the European Research Human Neuroscience, 4, 215. https://doi.org/10.3389/fnhum.2010.00215
Council under the Grant Agreement No. 820213 (ThinkAhead) to GP; Foster, D. J. (2017). Replay comes of age. Annual Review of Neuroscience, 40, 581–602.
Friston, K. (2005). A theory of cortical responses. Philosophical Transactions of the Royal
the PNRR MUR projects PE0000013-FAIR and IR0000011–EBRAINS-
Society of London B Biological Sciences, 360(1456), 815–836. https://doi.org/
Italy to GP; for the Wellcome Centre for Human Neuroimaging (Ref: 10.1098/rstb.2005.1622
205103/Z/16/Z) to KF, a Canada-UK Artificial Intelligence Initiative Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews
(Ref: ES/T01279X/1) to KF. The funders had no role in study design, Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787
Friston, K. (2013). Life as we know it. Journal of The Royal Society Interface, 10(86).
data collection and analysis, decision to publish, or preparation of the https://doi.org/10.1098/rsif.2013.0475
manuscript. Friston, K., & Frith, C. (2015). A Duet for one. Consciousness and Cognition, 36, 390–405.
The authors did not use generative AI technologies for preparation of Friston, K., Parr, T., & de Vries, B. (2017). The graphical brain: Belief propagation and
active inference. Network Neuroscience (Cambridge, Mass ), 1(4), 381–414. https://
this work. doi.org/10.1162/NETN_a_00018
Friston, K., Daunizeau, J., Kilner, J., & Kiebel, S. J. (2010). Action and behavior: A free-
energy formulation. Biol Cybern, 102(3), 227–260. https://doi.org/10.1007/s00422-
Declaration of Competing Interest 010-0364-z
Friston, K., Levin, M., Sengupta, B., & Pezzulo, G. (2015). Knowing one’s place: A free-
The authors declare no conflict of interest. energy approach to pattern regulation. Journal of The Royal Society Interface, 12
(105), 20141383. https://doi.org/10.1098/rsif.2014.1383
Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active
Data availability Inference: A Process Theory. Neural Computation, 29(1), 1–49. https://doi.org/
10.1162/NECO_a_00912
Friston, K., Lin, M., Frith, C. D., Pezzulo, G., Hobson, J. A., & Ondobaka, S. (2017). Active
No data was used for the research described in the article. Inference, Curiosity and Insight. Neural Computation, 29(10), 2633–2683. https://
doi.org/10.1162/neco_a_00999
References Friston, K., Sajid, N., Quiroga-Martinez, D. R., Parr, T., Price, C. J., & Holmes, E. (2021).
Active listening. Hearing Research, 399, Article 107998. https://doi.org/10.1016/j.
heares.2020.107998
Adams, R. A., Shipp, S., & Friston, K. J. (2013). Predictions not commands: active
Friston, K., Friedman, D.A., Constant, A., Knight, V.B., Parr, T., & Campbell, J.O. (2023).
inference in the motor system. Brain Structure & Function, 218(3), 611–643. https://
A variational synthesis of evolutionary and developmental dynamics. https://doi.org/
doi.org/10.1007/s00429-012-0475-5
10.3390/e25070964.
Ahmadi, A., & Tani, J. (2019). A novel predictive-coding-inspired variational RNN model
Friston, K., Da Costa, L., Sakthivadivel, D. A. R., Heins, C., Pavliotis, G. A., Ramstead, M.,
for online prediction and recognition. Neural Computation, 31(11), 2025–2074.
& Parr, T. (2023). Path integrals, particular kinds, and strange things (arXiv:
https://doi.org/10.1162/neco_a_01228
2210.12761). arXiv. https://doi.org/10.48550/arXiv.2210.12761
Arnal, L. H., & Giraud, A.-L. (2012). Cortical oscillations and sensory predictions. Trends
Funahashi, S., Bruce, C. J., & Goldman-Rakic, P. S. (1989). Mnemonic coding of visual
in Cognitive Sciences, 16(7), 390–398. https://doi.org/10.1016/j.tics.2012.05.003
space in the monkey’s dorsolateral prefrontal cortex. Journal of Neurophysiology, 61
Barrett, L. F., & Simmons, W. K. (2015). Interoceptive predictions in the brain. Nature
(2), 331–349. https://doi.org/10.1152/jn.1989.61.2.331
Reviews Neuroscience, 16(7), 419–429. https://doi.org/10.1038/nrn3950

8
G. Pezzulo et al. Biological Psychology 186 (2024) 108741

Gibson, J. J. (1979). The ecological approach to visual perception. Lawrence Erlbaum Pezzulo, G., & Cisek, P. (2016). Navigating the Affordance Landscape: Feedback Control
Associates, Inc. as a Process Model of Behavior and Cognition. Trends in Cognitive Sciences, 20(6),
Gregory, R. L. (1968). Perceptual illusions and brain models. Proceedings of the Royal 414–424. https://doi.org/10.1016/j.tics.2016.03.013
Society of London Series B Biological Sciences, 171(1024), 279–296. Pezzulo, G., Rigoli, F., & Friston, K. J. (2015). Active Inference, homeostatic regulation
Gregory, R. L. (1980). Perceptions as Hypotheses. Philosophical Transactions of the Royal and adaptive behavioural control. Progress in Neurobiology, 136, 17–35.
Society of London B, Biological Sciences, 290(1038), 181–197. https://doi.org/ Pezzulo, G., Rigoli, F., & Friston, K. J. (2018). Hierarchical Active Inference: A Theory of
10.1098/rstb.1980.0090 Motivated Control. Trends in Cognitive Sciences, 0(0). https://doi.org/10.1016/j.
Hasson, U., Chen, J., & Honey, C. J. (2015). Hierarchical process memory: Memory as an tics.2018.01.009
integral component of information processing. Trends in Cognitive Sciences, 19(6), Pezzulo, G., Zorzi, M., & Corbetta, M. (2021). The secret life of predictive brains: What’s
304–313. https://doi.org/10.1016/j.tics.2015.04.006 spontaneous activity for? Trends in Cognitive Sciences, 25(9), 730–743. https://doi.
Helmholtz, H. von (1866). Concerning the perceptions in general. In J. P. C. Southall org/10.1016/j.tics.2021.05.007
(Ed.), Treatise on physiological optics (Vol. 3). Dover. Pezzulo, G., Parr, T., & Friston, K. (2022). The evolution of brain architectures for
Hinton, G. E., Dayan, P., Frey, B. J., & Neal, R. M. (1995). The “wake-sleep” algorithm for predictive coding and active inference. Philosophical Transactions of the Royal Society
unsupervised neural networks. Science, 268(5214), 1158–1161. of London Series B, Biological Sciences, 377(1844), Article 20200531. https://doi.org/
Hoffmann, J. (2003). Anticipatory behavioral control. In M. V. Butz, O. Sigaud, & 10.1098/rstb.2020.0531
P. Gerard (Eds.), Anticipatory Behavior in Adaptive Learning Systems: Foundations, Pezzulo, G., Parr, T., Cisek, P., Clark, A., & Friston, K. (2023). Generating Meaning: Active
Theories, and Systems (pp. 44–65). Springer-Verlag. Inference and the Scope and Limits of Passive AI.
Hohwy, J. (2013). The predictive mind. Oxford University Press. Pezzulo, G., D’Amato, L., Mannella, F., Priorelli, M., Van de Maele, T., Stoianov, I. P., &
Hommel, B. (2003). Planning and representing intentional action. TheScientificWorld Friston, K. (2023). Neural representation in active inference: using generative
JOURNAL, 3, 593–608. models to interact with–and understand–the lived world. arXiv preprint arXiv:
James, W. (1890). The Principles of Psychology. Dover Publications. 2310.14810.
Lanillos, P., Meo, C., Pezzato, C., Meera, A. A., Baioumy, M., Ohata, W., Tschantz, A., Powers, W. T. (1973). Behavior: The Control of Perception. Aldine.
Millidge, B., Wisse, M., Buckley, C. L., & Tani, J. (2021). Active Inference in Robotics Priorelli, M., Pezzulo, G., & Stoianov, I. P. (2023). Deep kinematic inference affords
and Artificial Agents: Survey and Challenges (arXiv:2112.01871). arXiv. https://doi. efficient and scalable control of bodily movements. Proceedings of the National
org/10.48550/arXiv.2112.01871 Academy of Sciences, 120(51), Article e2309058120.
Lepora, N. F., & Pezzulo, G. (2015). Embodied Choice: How Action Influences Perceptual Proietti, R., Pezzulo, G., & Tessari, A. (2023). An active inference model of hierarchical
Decision Making. PLoS Comput Biol, 11(4), Article e1004110. https://doi.org/ action understanding, learning and imitation. Physics of Life Reviews, 46, 92–118.
10.1371/journal.pcbi.1004110 https://doi.org/10.1016/j.plrev.2023.05.012
Maisto, D., Donnarumma, F., & Pezzulo, G. (2023). Interactive Inference: A Multi-Agent Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional
Model of Cooperative Joint Actions. IEEE Transactions on Systems, Man, and interpretation of some extra-classical receptive-field effects. Nat Neurosci, 2(1),
Cybernetics: Systems, 1–12. https://doi.org/10.1109/TSMC.2023.3312585 79–87. https://doi.org/10.1038/4580
Maisto, D., Barca, L., Van den Bergh, O., & Pezzulo, G. (2021). Perception and Rens, N., Lancia, G. L., Eluchans, M., Schwartenbeck, P., Cunnington, R., & Pezzulo, G.
misperception of bodily symptoms from an active inference perspective: Modelling (2023). Evidence for entropy maximisation in human free choice behaviour.
the case of panic disorder. Psychological Review, 128(4), 690–710. https://doi.org/ Cognition, 232, Article 105328. https://doi.org/10.1016/j.cognition.2022.105328
10.1037/rev0000290 Robert, C. P. (2007). The Bayesian choice: From decision-theoretic foundations to
Mannella, F., Maggiore, F., Baltieri, M., & Pezzulo, G. (2021). Active inference through computational implementation (Vol. 2). Springer.
whiskers. Neural Networks: The Official Journal of the International Neural Network Schwartenbeck, P., Passecker, J., Hauser, T. U., FitzGerald, T. H., Kronbichler, M., &
Society, 144, 428–437. https://doi.org/10.1016/j.neunet.2021.08.037 Friston, K. J. (2019). Computational mechanisms of curiosity and goal-directed
Marr, D. (1982). Vision: A Computational Investigation into the Human Representation and exploration. eLife, 8, Article e41703. https://doi.org/10.7554/eLife.41703
Processing of Visual Information. Henry Holt and Co., Inc.. http://portal.acm.org/ Schwartenbeck, P., Baram, A., Liu, Y., Mark, S., Muller, T., Dolan, R., Botvinick, M.,
citation.cfm?id=1096911. Kurth-Nelson, Z., & Behrens, T. (2023). Generative replay underlies compositional
Maselli, A., Gordon, J.R., Eluchans, M., Lancia, G.L., Thiery, T., Moretti, R., Cisek, P., & inference in the hippocampal-prefrontal circuit. Cell. https://doi.org/10.1016/j.
Pezzulo, G. (2023). Beyond simple laboratory studies: Developing sophisticated models to cell.2023.09.004
study rich behavior. Seth, A. K., Suzuki, K., & Critchley, H. D. (2012). An Interoceptive Predictive Coding
Maturana, H. R., & Varela, F. J. (1980). Autopoiesis and Cognition: The Realization of Model of Conscious Presence. Frontiers in Psychology, 2. https://doi.org/10.3389/
Living. D. Reidel Pub. fpsyg.2011.00395
Miller, G. A., Galanter, E., & Pribram, K. H. (1960). Plans and the Structure of Behavior. Shipp, S., Adams, R. A., & Friston, K. J. (2013). Reflections on agranular architecture:
Holt, Rinehart and Winston. Predictive coding in the motor cortex. Trends in Neurosciences, 36(12), 706–716.
Mirza, M. B., Adams, R. A., Mathys, C. D., & Friston, K. J. (2016). Scene Construction, https://doi.org/10.1016/j.tins.2013.09.004
Visual Foraging, and Active Inference. Frontiers in Computational Neuroscience, 10, Sims, M., & Pezzulo, G. (2021). Modelling ourselves: What the free energy principle
56. https://doi.org/10.3389/fncom.2016.00056 reveals about our implicit notions of representation. Synthese. https://doi.org/
O’Regan, J. K., & Noe, A. (2001). A sensorimotor account of vision and visual 10.1007/s11229-021-03140-5
consciousness. Behavioral and Brain Sciences, 24(5), 883–917. Sterling, P. (2012). Allostasis: A model of predictive regulation. Physiology & Behavior,
Parr, T., & Friston, K. J. (2018). The Anatomy of Inference: Generative Models and Brain 106(1), 5–15.
Structure. Frontiers in Computational Neuroscience, 12. https://www.frontiersin.org/ Stoianov, I., Maisto, D., & Pezzulo, G. (2022). The hippocampal formation as a
articles/10.3389/fncom.2018.00090. hierarchical generative model supporting generative replay and continual learning.
Parr, T., & Friston, K. J. (2019a). Attention or salience? Current Opinion in Psychology, 29, Progress in Neurobiology, 217, Article 102329. https://doi.org/10.1016/j.
1–5. https://doi.org/10.1016/j.copsyc.2018.10.006 pneurobio.2022.102329
Parr, T., & Friston, K. J. (2019b). The computational pharmacology of oculomotion. Taniguchi, T., Murata, S., Suzuki, M., Ognibene, D., Lanillos, P., Ugur, E., Jamone, L.,
Psychopharmacology, 236(8), 2473–2484. https://doi.org/10.1007/s00213-019- Nakamura, T., Ciria, A., Lara, B., & Pezzulo, G. (2023). World Models and Predictive
05240-0 Coding for Cognitive and Developmental Robotics: Frontiers and Challenges (arXiv:
Parr, T., & Pezzulo, G. (2021). Understanding, Explanation, and Active Inference. 2301.05832). arXiv. https://doi.org/10.48550/arXiv.2301.05832
Frontiers in Systems Neuroscience, 15. https://www.frontiersin.org/articles/10.3389/ Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological Review, 55, 189–208.
fnsys.2021.772641. Tschantz, A., Barca, L., Maisto, D., Buckley, C.L., Seth, A.K., & Pezzulo, G. (2021).
Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active Inference: The Free Energy Principle Simulating homeostatic, allostatic and goal-directed forms of interoceptive control
in. Mind, Brain, and Behavior. MIT Press. using Active Inference. bioRxiv, 2021.02.16.431365. https://doi.org/10.1101/20
Parr, T., Limanowski, J., Rawji, V., & Friston, K. (2021). The computational neurology of 21.02.16.431365.
movement under active inference. Brain, 144(6), 1799–1818. https://doi.org/ Van den Bergh, O., Witthöft, M., Petersen, S., & Brown, R. J. (2017). Symptoms and the
10.1093/brain/awab085 body: Taking the inferential leap. Neuroscience and Biobehavioral Reviews, 74(Pt A),
Parr, T., Holmes, E., Friston, K. J., & Pezzulo, G. (2023). Cognitive effort and active 185–203. https://doi.org/10.1016/j.neubiorev.2017.01.015
inference. Neuropsychologia, 184, Article 108562. https://doi.org/10.1016/j. Walsh, K. S., McGovern, D. P., Clark, A., & O’Connell, R. G. (2020). Evaluating the
neuropsychologia.2023.108562 neurophysiological evidence for predictive processing as a model of perception.
Paulus, M. P., Feinstein, J. S., & Khalsa, S. S. (2019). An Active Inference Approach to Annals of the New York Academy of Sciences, 1464(1), 242–268. https://doi.org/
Interoceptive Psychopathology. Annual Review of Clinical Psychology, 15(1), 97–122. 10.1111/nyas.14321
https://doi.org/10.1146/annurev-clinpsy-050718-095617 Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the
Pezzulo, G. (2014). Why do you fear the bogeyman? An embodied predictive coding Machine. The MIT Press.
model of perceptual inference. Cognitive, Affective & Behavioral Neuroscience, 14(3),
902–911. https://doi.org/10.3758/s13415-013-0227-x