148 Sodobna pedagogika/Journal of Contemporary Educational Studies
Danaja Rutar
Developing higher cognition through 
predictive processing
Abstract: Predictive processing is an influential theoretical framework in cognitive science that 
promises to provide a unifying account of all perception, action and cognition. Whilst it has already 
successfully explained many facets of perception and action, crucially, it is lacking a coherent theoret-
ical account of cognitive development and higher cognition and, even more ambitiously, the develop-
ment of higher order thought. In this paper, I investigate to what extent predictive processing in its 
current form is even suitable for explaining development and higher thought. I suggest that as things 
stand, predictive processing cannot sufficiently explain the development of higher order as the inter-
nal models it proposes lack two crucial ingredients that characterise higher thought: compositionality 
and generativity . I propose how these two core features can be addressed within predictive processing. 
Lastly , I elaborate on the reciprocal relationship between pedagogical practices and predictive process-
ing: pedagogical practices are crucial for model building and predictive processing offers a theoretical 
framework for understanding psychological mechanisms involved in pedagogical settings.
Keywords: predictive processing, development of higher thought, compositionality, generativity , ped-
agogical implications
UDC: 159.95
Scientific article
Danaja Rutar, PhD., Molenveldlaan 222, 6523 RP , Nijmegen, Netherland; 
e-mail: rutar.danaja@gmail.com
Let./Vol. 74 (140)
Issue 3/2023
pp. 148–166
ISSN 0038 0474

149 
Introduction
Predictive processing (PP from now onwards) is a key theoretical framework 
in cognitive science that promises to explain the entirety of cognition, percep-
tion and action with one set of core mechanisms (Friston 2009, 2010; Clark 2013, 
2015). PP is built upon Bayesian theories of the brain, and as such, it postulates 
that the brain is a »probabilistic machine« that is continuously predicting the 
world and updating its generative model based on the discrepancy between what 
it predicts and what it perceives (Clark 2013, 2015). PP has been successfully ap-
plied to many different aspects of cognitive functioning, such as perception (Kok 
et al. 2013; Muckli et al. 2015; Summerfield and de Lange 2014), action under-
standing (Urgen and Miller 2015), planning and navigation (Kaplan and Friston 
2018), communication (Friston and Penny 2011), learning (Friston et al. 2017; 
Kwisthout et al. 2017; Rutar et al. 2023; Rutar et al. 2022; Smith et al. 2020) and 
even mentalising (Koster-Hale and Saxe 2013; Heil et al. 2019). Whilst PP has 
been applied to a great many domains, the majority of success has been achieved 
with regard to perceptual (as opposed to higher cognitive) processes in adults. 
If PP claims to provide a unifying theory for all aspects of human function-
ing – perception, action and cognition – accounting for the topics that are poorly 
addressed in PP , such as development and higher cognition, should be the priority 
of PP research. Investigation of higher cognition might prove to be specifically 
thorny, however, as PP was initially designed as a data compression method for 
studying information processing on the retina (Srinivasan et al. 1982). PP has 
been primarily used to try to explain low-level perceptual processes, and it is not 
trivial how and if it can scale up to higher cognition. In particular, it is not clear 
whether the kinds of internal models that underlie perceptual processes can suffi-
ciently account for higher cognitive processes (Clark 2013, p. 201): »What [...] the 
local approximations to Bayesian reasoning look like as I depart further and fur-
ther from the safe shores of basic perception and motor control? What new forms 
of representation are then required, and how do they behave in the context of the 
hierarchical predictive coding regime?« 
As for motivating the investigation of the developmental origins of PP , any 
Rutar

150 Sodobna pedagogika/Journal of Contemporary Educational Studies
account of cognition must be developable; otherwise, claims about adult cognition 
have weak explanatory value. Namely, the plausibility of adult capacities is rather 
limited if no explanation is provided as to how such capacities came to be (Ward 
et al. under review). This holds truer still if a theory aims to provide a unifying 
account of human functioning, such as that of PP .
The PP community is slowly starting to recognise the importance of address-
ing development and higher cognition. For example, in recent years, evidence that 
predictive machinery is in place in young children has started to pile up (Kayhan 
et al. 2019; Kayhan et al. 2019; Nagai 2019; Zhang et al. 2019), claims have been 
made that PP could be compatible with existing developmental findings (Köster 
et al. 2020) and first attempts at delineating developmental PP have begun to 
emerge (Ward et al. under review). Similarly, research considering higher aspects 
of cognition have started to take off (Heil et al. 2019; Neacsu et al. 2022; Rutar et 
al. 2022; Rutar et al. 2023; Smith et al. 2020). 
However, what we are still lacking is a mechanistic account of higher-order 
cognition and a theory of development or, even more ambitiously, a theory of the 
development of higher cognition. I will treat development and higher cognition 
(from the PP perspective) in synchrony. In what follows, I articulate the require-
ments that internal models and the learning mechanisms in PP will need to fulfil 
to allow for the development of higher thought. In fleshing out said requirements, 
I will consider two questions: What is the nature of internal models that support 
higher order thought? And what learning mechanisms underlie the development 
and refinement of these internal models? I start by presenting the core tenets of 
the PP theory and then address each of the questions in turn: in relation to the 
first question, I discuss how the models that support higher cognition are compo-
sitional (Fodor 1975; Fodor and Pylyshyn 1988) and in relation to the second ques-
tion, I discuss generativity as a crucial aspect of internal models that allows for 
further development and learning (Fodor 1975; Fodor and Pylyshyn 1988). Each 
of these two characteristics, compositionality and generativity , are discussed from 
the point of view of PP as well. Lastly, I explore how pedagogical practices are det-
rimental for model building and how PP can offer a useful theoretical framework 
for understanding psychological processes involved in a pedagogical setting.
What is PP?
Generative models compute predictions, hypotheses and prediction error 
According to PP , the brain functions as a hierarchically organised generative 
model that computes increasingly abstract hypotheses about the world. More spe-
cifically, at each level, the brain computes hypotheses about the hidden causes of 
the input from the layer below (Clark 2015; Friston 2010; Kwisthout et al. 2017; 
Friston et al. 2016). The generative model also continuously makes predictions 
about observable outcomes in terms of discrete events (Clark 2013; Friston et al. 
2016; Hohwy 2013; Kwisthout et al. 2017). More informally, the model predicts 
Rutar

151 
what it is going to experience next. For example, in the context of playing Don’t 
get angry! game predictions will pertain to the outcome of a dice roll (numbers 
from one to six), and when talking with a friend, predictions will pertain to the 
friend’s emotions (happy , sad, angry). Across the entirety of the generative model, 
predictions from the layers above and the sensory input from the layers below are 
compared – the difference between the two is computed and is called prediction 
error (Clark 2013, 2015; Friston and Kiebel 2009; Hohwy 2013). According to 
PP , the ultimate goal of any living system is to minimise prediction error in the 
long run (Friston 2009, 2010). Lower layers of the model are involved in process-
ing concrete perceptual details, whereas the higher layers code for more abstract 
aspects of events that can be more temporally and spatially remote (Clark 2013; 
Hohwy 2013). For example, predictions at lower layers will contain information 
about patterns of light on the retina, and complementary predictions at higher 
layers will contain information about which object might have caused those pat-
terns of light on the retina.
Minimising prediction error
There are two ways in which a cognitive system can minimise prediction 
error: through perceptual inference and active inference (Kwisthout et al. 2017). 
Perceptual inference or Bayesian model updating pertains to changing the current 
probability distribution over the hypotheses at different layers of the hierarchy 
until there is no prediction error left. Bayesian model updating is an iterative 
process that happens every time the brain receives new sensory evidence (Bayes 
1763/1958; Oaksford and Chater 2009; O’Reilly et al. 2013). With every new piece 
of evidence, the hypothesis changes slightly, which is reflected in the new estimate 
of the mean and variance of the hypothesis. This process leads to more accurate 
predictions in the future (Clark 2013; Friston 2010; Kwisthout et al. 2017). For 
example, suppose that you meet a new friend and after a period of meeting up 
with them, you start noticing a pattern: your friend is always approximately 10 
minutes late for your appointment. What happens at the level of the model is that 
the probability distribution over the hypothesis that predicts your friend coming 
at the agreed-upon time is slowly changing. Over time, the hypothesis that your 
friend will be on time becomes less and less likely; thus, when your friend is yet 
again 10 minutes late, this occurrence generates less prediction error. Alterna-
tively, prediction error can be minimised through active inference, which means 
sampling new sensory evidence to bring it more in line with predictions, which 
also leads to more accurate predictions in the future (Friston 2010; Kwisthout et 
al. 2017; Friston et al. 2016). For example, a dancer makes a jump and accidentally 
lands on a co-dancer. The failed jump generates a proprioceptive prediction error 
between where she expected to land and where she ended up landing. In light 
of this prediction error, she changes her position to where she intended to land, 
bringing her position closer to the predicted one.
Developing higher cognition through predictive processing

152 Sodobna pedagogika/Journal of Contemporary Educational Studies
PP account of higher-order cognition 
Having presented the core features of PP , I now move to the core part of the 
paper. In the following, I will investigate to what extent PP is as it stands able to 
explain the development of higher cognition. In addressing this question, I will 
investigate two sub-questions: What is the nature of internal models that support 
higher-order thought? And what learning mechanisms allow for the development 
and refinement of these internal models? Discussing models and learning require-
ments in relation to these questions will provide the basis for further evaluation 
of the extent to which PP satisfies these requirements.
What is the nature of internal models that support higher-order thought?
Higher thought is compositional
A core feature of the human (both adult and children) mind is that it is capa-
ble of thinking about anything, imagining things that it has never seen, reimaging 
events that have happened long ago, solving novel problems, generating new be-
liefs, coming up with new concepts and even new languages, to mention just a few 
examples of this remarkable ability. One common denominator of these higher 
cognitive processes is that they are compositional – compositional processes rely 
on a limited number of simple elements (e.g. simple concepts) that can be flexibly 
combined and recombined to produce more complex elements (e.g. complex con-
cepts) according to certain rules. In other words, compositionality means that the 
meaning of complex conceptual structures is a function of the meaning of more 
basic elements and the way they are combined (Fodor 1975; Fodor and Pylyshyn 
1988). A simple, concrete example of compositionality is the following case: the 
meaning of the sentence »Mary loves John« is a compositional function of the 
grammar and meaning of the units in the sentence, »Mary«, »John« and »loves«, 
the way in which these are combined into the sentence. Another complementary 
feature of higher cognition is systematicity, which denotes that there exist system-
atic relations between cognitive elements (Fodor and Pylyshyn 1988). While it is 
an important area of debate, I will not investigate systematicity further here due 
to space constraints and will assume that compositionality is systematic. 
If human thought is compositional, then internal models that underlie 
thought processes are compositional as well. Language of thought (LOT) (Fodor 
1975; Fodor and Pylyshyn 1988; Piantadosi 2021; Piantadosi et al. 2012, Piantado-
si 2016) is an influential theory in cognitive science that successfully empirical-
ly and computationally investigates the nature of internal models that support 
compositionally structured thought processes. According to LOT , humans possess 
a symbolically structured internal model that operates analogous to language 
(Fodor 1975; Piantadosi 2021). The mind composes complex thoughts by reas-
sembling parts of the internal model expressed in a symbolic language of thought 
(Fodor 1975; Fodor and Pylyshyn 1988; Piantadosi 2021; Piantadosi et al. 2016).
Rutar

153 
If the mind is compositional, what are its most basic elements? One way 
to think about this is in terms of primitives, that is, basic cognitive operations 
and composition laws that are available to a cognitive system before the process 
of compositional (hypothesis) formation begins (Piantadosi et al. 2012). In other 
words, primitives are the key cognitive operations that are available at the onset 
of learning and that allow for the acquisition of concepts and conceptual systems. 
For example, in the context of learning how to count, Piantadosi (2012) suggests 
that a plausible set of primitives that children possess at the start of this process 
consists of three primitive functions – singletons, doubletons, tripletons – which 
test whether a set contains one, two, or three elements. Other primitive oper-
ations in the context of counting could also be next, previous and equal, which 
provide functions for moving forwards and backwards on the count list and check-
ing if two numbers are equal (Piantadosi et al. 2012). Some other, more general 
primitives are Boolean operators, such as and, or, not and if … then (Piantadosi et 
al. 2012). Whilst some primitives are used across cognitive domains, some other 
primitives are more domain specific, such as in the case of counting.
To sum up, internal models that support higher cognition need to be com-
positional and contain a number of primitives that allow for the construction of 
more complex compounds. For the purposes of this paper, the crucial question is 
whether internal models in PP are compositional.
Internal models in PP are not richly compositional
At the beginning, I discussed how generative models in PP consist of hy-
potheses that are arranged at different hierarchical layers. What exactly do these 
hypotheses encode? They encode the hidden causes of the sensory patterns in the 
co-occurring feature observations. More formally, hypotheses have a prior distri-
bution and a likelihood distribution that describes how well each hypothesis cap-
tures the observed sensory pattern. Thus, differences between hypotheses pertain 
to the differences in the likelihoods of sensory patterns occurring under different 
hypotheses. Given such formalisation of hypotheses in PP , it is not immediately 
clear how they could account for the compositional character of thought. To use 
an enhanced version of an example referred to earlier, »Mary loves John because 
they share core values«, we understand the meaning of this sentence instanta-
neously, without having to rely on observing the co-occurrence of features or as-
sociations that code for this particular occurrence at some level. We understand 
this sentence because we understand the meaning of »love« and »values« in some 
abstract sense, long dissociated from learned associations between sensory obser-
vations. PP does not specify how the encoded co-occurring feature observations 
translate to primitive operations, what these might be in the first place, and how 
the primitives combine to result in new concepts, abstractions and laws. Thus, 
hypotheses as postulated by PP might not be sufficiently compositional and hence 
unable to account for logically structured higher thought. 
Coming from a similar angle, Williams (2020) argues that hypotheses in PP 
Developing higher cognition through predictive processing

154 Sodobna pedagogika/Journal of Contemporary Educational Studies
are only weakly compositional. As with any statistical model, weak composition-
ality is achieved by the possible assignments of probabilities to values and assign-
ment of values to hypothesis variables where different assignments correspond to 
different facts about the world. Generative models construed this way have the 
expressive power of propositional logic (Russell 2010). To understand the limita-
tions of propositional logic, it should be contrasted with a more expressive formal 
system, first order (i.e. predicate) logic (Haaparanta 2009), as follows (Williams 
2020, p. 1768): »In contrast to factored representations, the ontology of first-or-
der logic comprises not just facts but objects and relations, thereby representing 
,the world as having things in it that are related to each other, not just variables 
with values‘ (Russell and Norvig 2010, p. 58). As such, first-order logic decompos-
es a propositional representation (e.g. Hume is a philosopher) into subjects (e.g. 
Hume), n-place predicates (e.g. is a philosopher), and quantifiers (e.g. there is an 
x such that x is Hume and x is a philosopher). The resulting gains in expressive 
power are famously enormous.«
Williams’ (2020) conclusion is that higher order thought is richly composi-
tional and that the minimal formal system that can sufficiently capture this kind 
of rich compositionality is predicate logic. Generative models in PP , in contrast, 
can be described by propositional logic, which can only account for weak composi-
tionality. Therefore, generative models in PP are currently not sufficiently rich to 
fully account for higher order thought.
Could PP account for the compositionality of thought? 
In the following, I discuss how compositionality might be addressed within 
PP , and in particular, what compositionality precursors might look like in infant 
generative models. In the cognitive science community, it is a relatively non-con-
tentious assumption that generative models do not need to be learned from 
scratch (Friston et al. 2021). That is, evidence exists that infants are born with 
some built-in evolutionary acquired cognitive architecture and functionality; they 
perform basic statistical computations (Kirkham et al. 2002), engage in probabil-
istic reasoning (Denison et al. 2013), have some expectations about the physical 
and the social world (Spelke and Kinzler 2007), can reason about hidden causes 
(Koster-Hale and Saxe 2013) and engage in model updating and prediction error 
minimisation (Kayhan et al. 2019; Zhang et al. 2019). 
Additionally, infants less than one year old (Lewkowicz et al. 2018; Werchan 
et al. 2015) are already able to learn hierarchical rules and can engage in the 
precursors of analogical thought (Gentner and Hoyos 2017), which might be in-
terpreted as an indicator of infants possessing some compositional capacity from 
very early on. Based on these empirical findings, it is plausible to assume that the 
basic structure of the compositional hierarchical generative model is already in 
place in infants. What might cognitive architectures that support compositional 
higher thought look like?
Rutar

155 
Initial state of the generative model 
I propose that initially, a simple distinction might exist between a lower level 
and a higher level of a generative model and gradually through development, the 
model will become populated with more and more complex hypotheses at different 
hierarchical levels. I suggest that the higher level of minimally compositional in-
fant models contains some core primitive functions – within PP , these would corre-
spond to simple hypotheses. Such hypotheses would equip infants with a general 
representational capacity and would need to be sufficiently rich to create mean-
ingful, more complex hypotheses at lower levels of the hierarchy when combined. 
Whilst the higher levels of the model would specify available primitives, the lower 
levels would represent the compositions of these primitives into concrete hypoth-
eses that could explain the causes underlying the incoming sensory sensations. 
Similarly, coming from a general Bayesian tradition (though not PP), Per-
fors (2012) suggests that within (adult) generative models, a distinction exists 
between a so-called latent hypothesis space and an explicit hypothesis space. The 
former contains general primitive functions that can be combined into more com-
plex hypotheses – as such, it represents the space of all possible hypotheses that 
can be represented. These simple hypotheses in the latent hypothesis space can 
be combined to create more complex hypotheses for predicting and explaining the 
incoming sensory evidence – such hypotheses belong to the explicit hypothesis 
space. The crucial distinction between the two kinds of hypothesis spaces or levels 
is in the type and function of the hypotheses they possess. Hypotheses in the latent 
hypothesis space represent basic and general building blocks for building concrete 
hypotheses that can then be actively manipulated by a cognitive system, for ex-
ample, for predicting and evaluating sensory evidence in the explicit hypothesis 
space (Perfors 2012).
Possible primitives in initial generative models
Next, I suggest two kinds of primitives – individual primitives and systems of 
primitives – that might be present in the highest levels of the earliest generative 
models.
One set of primitives that infants could possess would be involved in con-
structing hypotheses for explaining a wide range of sensory observations across 
domains; thus, these primitives would be domain-general. An important condition 
that such primitives would need to fulfil is that they would allow for emerging 
compositionality that parallels that of humans. To my knowledge, developmental 
scientists are yet to map out a list of such domain-general primitive functions in 
infants although there are findings that show that eight-month-old infants can 
learn the same/different relation (Addyman and Mareschal 2010) as well as em-
pirical evidence that five-month-olds can perform simple addition and subtraction 
(Simon et al. 1995). These findings indicate the type of primitives that could be 
found in infant generative models. Another set of expressively powerful primitives 
Developing higher cognition through predictive processing

156 Sodobna pedagogika/Journal of Contemporary Educational Studies
that allow for rich compositionality was suggested by Piantadosi (2016) in the 
context of adult concept learning. He found that adult concept learning can be 
sufficiently explained only by a small set of primitives, such as »and«, »or«, »not« 
and »if and only if«. Note that these are just some ideas for the kind of primitives 
we might be looking for in the infant mind. 
The other set of theoretical ideas that could inspire what the nature of prim-
itive hypotheses might be comes from research on core knowledge (Spelke et al. 
1992; Spelke and Kinzler 2007). This research pertains to early emerging or pos-
sibly innate expectations about the world, such as expectations about object be-
haviour, agents, number, space and time (Spelke et al. 1992; Spelke and Kinzler 
2007). Core knowledge is general and abstract and applies to every member with-
in a domain and therefore can be thought of as laws and principles that govern 
the behaviour of entities in specific domains. Therefore, core knowledge is a kind 
of a start-up software or a start-up library of primitive operations and programs 
(combinations of primitive operations) built in by evolution (Ullman and Tenen-
baum 2020). An example of such a primitive program would be the principle of 
solidity – 2.5-month-old infants already expect that solid moving objects will not 
pass through another solid object (Renee Baillargeon 2004). Similarly , 2.5-month-
olds and 3.5-month-olds understand the principle of continuity (Renée Baillargeon 
2008) and object permanence (Renée Baillargeon and DeVos 1991), respectively. 
Infants expect that objects continue to exist when occluded and seem surprised if 
objects magically disappear.
For a side-by-side comparison of both ideas for primitives at the highest lev-
els of the generative model (as presented in this section) and how they might 
relate to complex hypotheses at the lower levels (as presented in The initial state 
of the generative model), see Figure 1.
Rutar

157 
COMPOSITIONAL BUILDING BLOCKS FOR HYPOTHESES 
Individual primitives Programs - collections of primitives
Domain general Mostly domain specific 
Single operations (e.g. and, or, if 
… then, add)
Laws governing specific domains 
(e.g. principle of continuity, object 
permanence, principle of goal-
directedness)
Abstract
Higher levels of 
generative model
(a)
HYPOTHESES
Composed from primitive hypotheses and programs 
Applied to concrete sensory observations
Actively used in prediction and explanation
Lower levels of 
generative model
(b)
Primitives and programs are used 
for generating concrete hypotheses. 
Figure 1. Higher-level primitives and lower-level concrete hypotheses. (a) Two types of primitives might 
co-exist in the highest levels of infant generative models: individual, general primitives and programs, 
which are combinations of primitives and explain the laws underlying individual cognitive domains. 
(b) Concrete hypotheses are formed at the lower levels of infant generative models: these hypotheses are 
formed based on the higher-level primitives and are constructed in response to concrete observations. 
Developing higher cognition through predictive processing
What learning mechanisms allow for the development and refinement of these in-
ternal models?
Learning is the result of generative capacity of the mind 
The first part of this paper discusses the centrality of compositionality to 
human thought (Fodor 1975; Fodor and Pylyshyn 1988; Piantadosi 2016). In fact, 
compositionality is usually presented together with another feature that is cen-
tral to higher cognition – productivity or generativity (Fodor 1975; Fodor and Py-
lyshyn 1988). Whilst compositionality characterises the nature of internal models 
that can support higher thought processes, productivity is specifically important 
in the context of learning and development. Productivity can most succinctly be 
characterised as the »infinite use of finite means«, which means that a cognitive 
system equipped with a productive capacity is, in principle, capable of producing 
an infinite amount of thoughts and concepts. An example of a productive system 
is language; English contains a finite number of words and because there is no 
upper limit on the length of sentences, there is also no upper limit on the number 
of unique sentences that can be formed. The capacity for a speaker’s sentence 
construction is productive, which means that they are able to, in principle at least, 
form an infinite number of unique sentences (Katz 2009).

158 Sodobna pedagogika/Journal of Contemporary Educational Studies
Do internal models in PP have generative capacity? 
Productive capacity necessitates that there exist some sort of learning mech-
anism that can generate new thought, that is, hypotheses. What might this mech-
anism be in the context of PP? As discussed at the beginning, the basic mode of 
learning in PP that pertains to model changes is Bayesian model updating, where 
the incoming sensory evidence is continuously changing the probability distribu-
tion over existing hypotheses. Thus, casting learning as Bayesian model updating 
is suitable when all the relevant hypotheses are already present in the model; 
however, it cannot explain how novel hypotheses are generated in the process of 
learning and development.
To address this exact problem, very recently another form of learning in PP 
has been formalised – structure learning (Friston et al. 2017; Kwisthout et al. 
2017; Neacsu et al. 2022; Rutar et al. 2022; Rutar et al. 2023; Smith et al. 2020) – 
and empirically differentiated from Bayesian model updating (Rutar et al. 2023). 
This type of learning, unlike Bayesian model updating, allows for changing the 
structure of the generative model by removing, adding, merging hypotheses, or 
establishing new connections between hypotheses. Consider the following exam-
ple from earlier (your friend always being fashionably late) that illustrates the 
difference between structure learning and Bayesian model updating: your friend 
continuously being late leads to gradual updating of the hypothesis (about your 
friend’s arrival) such that after some time, the hypothesis starts predicting the 
friend coming at the proposed time with only a low probability . Then one day , your 
friend tells you that she is always late because she got two small puppies that al-
ways demand her attention when she is about to head out. What accounts for the 
acquisition of a new belief of why your friend is late is structure learning – your 
model added a new hypothesis »puppies attention« that is causally connected to 
the hypothesis coding for »friend being late«.
What is fascinating for the purpose of this paper is that structure learning 
could potentially offer a formal basis for addressing the productivity of thought 
from the PP perspective. If compositional models provide primitive elements, then 
structure learning could specify ways of combining these primitives to trigger new 
thoughts, concepts and ideas. More specifically, what I am suggesting here is that 
primitives combined in a new way would result in a person coming up with a con-
cept that would correspond to a new hypothesis being added in a generative mod-
el via structure learning. Yet another example of structure learning could be to 
combine the primitives in a way that results in a more detailed hypothesis. Each 
new and meaningful combination of primitives would correspond to a distinct 
structural change in a model, such as adding a new hypothesis, removing an old 
one, merging existing ones, etc. Manipulating the hypothesis space via structure 
learning and hence manipulating the expressive power of the generative model 
could then be a PP way of satisfying the productivity condition. 
The same mechanism – structure learning – could also explain developmen-
tal processes. For example, in the previous section, I suggested that initially, in-
fants have a very crude structure of a generative model in place with the simplest 
Rutar

159 
hypotheses – primitives – on the highest level of the model. Development with 
respect to structure learning could be characterised as a gradual population of the 
generative model with hypotheses at different levels. More specifically, one aspect 
of the emerging generative model might be important in the context of develop-
ment: establishing hypotheses at different levels of granularity. Granularity, or 
the level of detail of hypotheses, pertains to the amount of information present 
in one hypothesis (Kwisthout et al. 2017). For example, a hypothesis with a high 
level of detail would be that when I go to a dog shelter, I will see a brown puppy 
Labrador, and a hypothesis with a low level of detail would be that I will see a 
Labrador. Learning different levels of granularity of hypotheses might underlie 
cognitive processes such as category learning in children, in which more abstract 
categories would correspond to hypotheses with a lower level of detail and learn-
ing sub-categories would pertain to more detailed hypotheses. As children’s gen-
erative models mature, they might start additionally considering the prediction 
error–information trade-off when choosing the appropriate level of detail of a hy-
pothesis. The trade-off pertains to the fact that the more detailed the hypothesis, 
the more information it can yield if correct, but such a hypothesis can also gener-
ate more prediction error if wrong (Kwisthout and van Rooij 2015). For example, 
if I am not sure that the dog shelter allows for puppies, then predicting that I will 
see a brown puppy Labrador is costlier (as my prediction will be likely wrong and 
hence generate prediction error) than predicting that I will see a Labrador. Devel-
opment thus pertains not only to generating increasingly more detailed hypothe-
ses but also to choosing hypotheses with the appropriate level of detail.
Predictive processing in a pedagogical context
In this final chapter, I sketch a reciprocal relationship between pedagogical 
practices and PP as a theoretical framework.
Pedagogical practices play a crucial role in model building 
The focal point of this paper has been to provide PP theoretical foundations 
for studying the development of higher cognition and to elucidate the learning 
processes involved in this development. What is at the core of learning and devel-
opment is »change« or »transition« within one’s internal model. Learning can be 
characterised as moving from a place of not knowing to a place of knowing (more), 
and development is about a transition from a less developed to a more developed 
cognitive capability. Whilst in this paper I have, so far, focused exclusively on the 
theoretical characterisation of this change in an individual, crucially , learning and 
developmental changes most often happen in (in)formal learning environments. 
Thus, pedagogical practices, which create learning environments, play an import-
ant role in an individual’s model building. Below, I present three ways in which 
individual learning and teaching are intertwined and point out how internal mod-
el building might benefit from pedagogical guidance.
Developing higher cognition through predictive processing

160 Sodobna pedagogika/Journal of Contemporary Educational Studies
Goal setting: Learning is a directed activity in which a learner tries to get 
from an initial point of not knowing to a more advanced point of knowing. Often-
times, this is achieved through explicit goal setting. Well-defined goals can specify 
the structure of the problem as well as the solution (Magid et al. 2015). Such struc-
turing makes learning more manageable as it constrains a student’s hypothesis 
space, that is, the space of all hypotheses that a student might consider in solving 
a problem (Magid et al. 2015). Guidance from teachers can help shape students’ 
learning (by the above-mentioned constraining of the hypothesis space) and in so 
doing specify where in the learning space learning progress should happen. It is as 
if well-defined goals carve out the path for learning in a student’s model.
Scaffolding: If goal setting defines the starting and end point of learning, 
another crucial ingredient is finding the right steps to get from the starting to the 
end point. Model building is a complex cognitive activity in which the structured 
sequence of steps matters a lot. Well-structured learning environments will have 
a greater potential to incur well-connected and meaningful structural changes in 
pupil’s internal models (Vosniadou et al. 2001). Namely, when a model is in its 
immature state and hence a student does not possess a lot of knowledge in one do-
main yet, the student cannot just ‘jump’ to a highly advanced, complex model, but 
rather needs to take systematic, ordered steps. And this is exactly where teachers 
play a highly valuable role, as they can a) identify the starting level of pupils and 
b) scaffold their learning process, that is, challenge them just a bit above their cur-
rent level of competence and help them get to the next competence level through 
guided instruction and probing questions (Vygotsky and Cole 1978). If goal-set-
ting carves out the learning space, then continuous and gradual scaffolding guides 
children systematically through this space.
Feedback: Evidence about the role of explicit feedback in teaching is mixed 
and its efficacy can depend on how and when it is delivered and what is corrected 
(see Ellis 2009, for an overview). I suggest that in the context of model building, 
teacher’s feedback might play a crucial role as it can reinforce successful cognitive 
strategies and promote meta-cognitive learning skills in students. Teachers can, 
in delivering their feedback, elaborate why certain ways of solving problems are 
better than others, and they can also encourage pupils to actively reflect on their 
own problem-solving approaches. In the previous section, I argued that teachers 
can implicitly scaffold learning for students via guided instruction, whereas feed-
back could be regarded as an explicit form of scaffolding.
Predictive processing as a theoretical tool for pedagogical practices
Above, I show how pedagogical practices are constitutive of model building 
(in PP). Now, the reverse relationship also holds true. Namely, PP is a useful the-
oretical tool for understanding learning and teaching processes. Often, references 
to the »active student«, »active learning«, »student-oriented approach«, etc. are 
made only at the level of phraseology and blanket assertions. To substantiate these 
claims, it is important to elaborate on these concepts precisely and shed light on 
Rutar

161 
the neural and cognitive mechanisms that underpin them. This is exactly where I 
see the contribution of a) the PP framework and specifically of b) my theoretical 
framework (for characterising the development of higher cognition through PP). 
Regarding a), according to PP , humans are continuously probabilistically 
predicting their world (Clark 2013; Friston et al. 2016; Hohwy 2013), gathering 
new evidence so as to reduce prediction error and associated uncertainty (Clark 
2013, 2015; Friston and Kiebel 2009; Hohwy 2013), and updating their beliefs 
according to new evidence (Oaksford and Chater 2009; O’Reilly et al. 2013). Re-
cent experiments with cognitively inspired robots (which are in line with human 
experiments, e.g. Poli et al. 2020) have also found that the most effective algo-
rithm for self-organisation and structuring of the learning curriculum in novel 
environments is the maximisation of »learning progress« (Oudeyer and Smith 
2016). That is, when the robot was equipped with the algorithm that maximised 
learning progress, it avoided activities that were too hard to learn as well as those 
that were too easy to learn, instead exploring activities that yielded the biggest 
learning progress. When the learning progress of the current activity started to 
plateau, the robot switched to the activity that offered the second biggest learn-
ing progress. All these results suggest that humans are intrinsically highly active 
learners that are continuously adjusting their internal models, seeking out novel 
information and actively structuring their learning environments.
In one sense, any mental act – such as gathering information, updating and 
minimising prediction error – is an activity. However, a different kind of activity 
is also involved in learning, that is, higher cognitive processes. Such cognitive 
activities play a constitutive role in effectively executing and regulating the learn-
ing progress through the use of abstraction, generalisation, induction, deduction, 
classification, analogy, etc. (Ritchhart et al. 2011). Therefore, regarding b), my PP 
theoretical advance on the development of higher cognition is not only interest-
ing from a theoretical point of view (expanding the explanatory scope of PP), but 
also provides the cognitive and possibly neural basis (as PP is neuroscientifically 
sound) of higher cognitive processes at work in learning. Whereas cognitive ac-
tivities described under a) proceed without any instruction or guidance, higher 
cognitive activities described under b), crucially, rely on teachers’ support in the 
form of scaffolded instructions, questioning and supportive guidance.
Conclusion 
PP promises to provide a unifying account of all human functioning, from 
perception and action to cognition. Whilst PP has been successfully applied to 
various aspects of human functioning (Friston and Penny 2011; Heil et al. 2019; 
Kaplan and Friston 2018; Rutar et al. 2023; Summerfield and de Lange 2014) its 
application to higher order cognition and development has been slower. Some re-
searchers have shown, for example, that the basic PP machinery, that is, Bayesian 
model updating (Kayhan et al. 2019; Kayhan et al. 2019) and prediction error 
minimisation (Zhang et al. 2019), is already in place in infants. Likewise, Köster 
Developing higher cognition through predictive processing

162 Sodobna pedagogika/Journal of Contemporary Educational Studies
(2020) has proposed ways in which PP is compatible with multiple developmental 
findings. In addition, researchers have recently started evaluating higher cogni-
tion from the PP perspective (Heil et al. 2019; Neacsu et al. 2022; Rutar et al. 
2022; Rutar et al. 2023; Smith et al. 2020). In short, in recent years, the founda-
tions have been laid for rigorously examining the empirical and modelling basis of 
the PP account of development and higher cognition. However, what we are still 
lacking is a mechanistic account, an overarching theory of higher-order cognition 
and a theory of development (though see Ward et al. under review, as a first at-
tempt at constructing a theory of development in PP), or even more ambitiously, 
a theory of the development of higher cognition.
In this paper, I provided some initial elements that might be important to 
consider in developing a full-blown theory of the development of higher order 
thought in PP . I suggested that two questions should be considered in this context 
– What is the nature of internal models that support higher-order thought? And 
what learning mechanisms allow for the development and refinement of these in-
ternal models? – and proposed that internal models that support higher cognition 
are compositional and the learning mechanism that further refines and develops 
internal models needs to possess generativity (Fodor 1975; Fodor and Pylyshyn 
1988). I further argued that PP could be complemented with LOT to account for 
the compositionality condition and that structure learning could be a good formal-
ism to capture the generativity of thought.
Future work should investigate the suggested leads further and, most impor-
tantly, evaluate the compatibility between conceptual contributions here and the 
mathematical/formal basis of PP . Additionally, if within PP the core aim of cogni-
tive systems is to minimise prediction error in the long run, the obvious question 
in the context of higher cognition is how higher thought processes such as think-
ing, imagining and learning new concepts fulfil that goal. Importantly, I have also 
argued that considering the role of education and teaching will be important in 
the context of understanding how infants’ internal models develop. This last point 
calls for a tighter link between cognitive science, educational policies and teaching 
practice – how exactly findings from this research should inform educational prac-
tices should be investigated in the future. Many open questions remain; however , I 
believe that constructing a PP theory of the development of higher cognition will 
be worthwhile.
References
Addyman, C. and Mareschal, D. (2010). The perceptual origins of the abstract same/differ-
ent concept in human infants. Animal cognition, 13, pp. 817–33.
Baillargeon, R. (2004.) Infants’ reasoning about hidden objects: Evidence for event-general 
and event-specific expectations. Developmental science, 7, issue 4, pp. 391–414.
Baillargeon, R. (2008). Innate ideas revisited: For a principle of persistence in infants’ 
physical reasoning. Perspectives on psychological science, 3, issue 1, pp. 2–13.
Baillargeon, R. and DeVos, J. (1991). Object permanence in young infants: Further evi-
dence. Child development, 62, issue 6, pp. 1227–46.
Rutar

163 
Bayes, T . (1958). Studies in the history of probability and statistics: Ix. Thomas Bayes’ 
essay towards solving a problem in the doctrine of chances. Biometrika, 45, issue 3, 
pp. 296–315.
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cog-
nitive science. Behavioral and brain sciences, 36, issue 3, pp. 181–204.
Clark, A. (2015). Surfing Uncertainty: Prediction, action, and the embodied mind. Oxford: 
Oxford University Press.
Denison, S., Reed, C. and Xu, F . (2013). The emergence of probabilistic reasoning in very 
young infants: Evidence from 4.5-and 6-Month-Olds. Developmental psychology, 49, 
issue 2, p. 243.
Ellis, R. (2009). Corrective feedback and teacher development. L2 Journal, 1, issue 1.
Fabry, R. E. (2017). Predictive processing and cognitive development. In: T . Metzinger and 
W . Wiese (Eds.). Philosophy and predictive processing: 13. Frankfurt am Main: MIND 
Group.
Fodor, J. A. (1975). The language of thought. Vol. 5. New York: Harvard University Press.
Fodor, J. A. and Pylyshyn, Z. W . (1988). Connectionism and cognitive architecture: A critical 
analysis. Cognition, 28, issue 1/2, pp. 3–71.
Friston, K. J. (2009). The free-energy principle: A rough guide to the brain? Trends in cog-
nitive sciences, 13, issue 7, pp. 293–301.
Friston, K. J. (2010). The free-energy principle: A unified brain theory? Nature reviews 
neuroscience, 11, issue 2, pp. 127–38.
Friston, K. J., FitzGerald, F ., Rigoli, F ., Schwartenbeck, P . and Pezzulo, G. (2016). Active 
inference and learning. Neuroscience & biobehavioral reviews, 68, pp. 862–79.
Friston, K. J., Lin, M., Frith, C. D., Pezzulo, G., Hobson, J. A. and Ondobaka, S. (2017). 
Active inference, curiosity and insight. Neural computation, 29, issue 10, pp. 2633–83.
Friston, K. J. and Kiebel, S. (2009). Predictive coding under the free-energy principle. Phil-
osophical transactions of the royal society B: Biological sciences, 364, issue 1521, pp. 
1211–21.
Friston, K., J. Moran, R. J., Nagai, Y., Taniguchi, T., Gomi, H. and Tenenbaum, J. (2021). 
World model learning and inference. Neural networks, 144, pp. 573–590. 
Friston, K. J. and Penny, W. (2011). Post hoc Bayesian model selection. Neuroimage, 56, 
issue 4, pp. 2089–99.
Gentner, D. and Hoyos, C. (2017). Analogy and abstraction. Topics in cognitive science, 9, 
issue 3, pp. 672–93.
Haaparanta, L. (2009). The development of modern logic. New York: Oxford University 
Press. 
Heil, L., Colizoli, O., Hartstra, E., Kwisthout, J., van Pelt, S., van Rooij, I. and Bekkering, 
H. (2019). Processing of prediction errors in mentalizing areas. Journal of cognitive 
neuroscience, 31, issue 6, pp. 900–912.
Hohwy, J. (2013). The predictive mind. Oxford: Oxford University Press.
Kaplan, R. and Friston, K. J. (2018). Planning and navigation as active inference. Biologi-
cal cybernetics, 112, issue 4, pp. 323–43.
Katz, M. (2009). The language of thought hypothesis. Internet encyclopedia of philosophy, 
pp. 1–10.
Kayhan, E., Heil, L., Kwisthout, J., van Rooij, I., Hunnius, S. and Bekkering, H. (2019). 
Young children integrate current observations,priors and agent information to pre-
dict others’ actions. PloS One, 14, issue 5, p. e0200976.
Kayhan, E., Meyer, M., O’Reilly, J. X., Hunnius, S. and Bekkering, H. (2019). Nine-month-
old infants update their predictive models of a changing environment. Developmental 
cognitive neuroscience, 38, p. 100680.
Developing higher cognition through predictive processing

164 Sodobna pedagogika/Journal of Contemporary Educational Studies
Kirkham, N. Z., Slemmer, J. A. and Johnson, S. P . (2002). Visual statistical learning in in-
fancy: Evidence for a domain general learning mechanism. Cognition, 83, issue. 2, p. 
B35–42.
Kok, P ., Brouwer, G. J., van Gerven, M. A. J. and de Lange, F . P . (2013). Prior expectations 
bias sensory representations in visual cortex. Journal of neuroscience, 33, issue 41, 
pp. 16275–84.
Köster, M., Kayhan, E., Langeloh, M. and Hoehl, S. (2020). Making sense of the world: In-
fant learning from a predictive processing perspective. Perspectives on psychological 
science, 15, issue 3, pp. 562–71.
Koster-Hale, J. and Saxe, R. (2013). Theory of mind: A neural prediction problem. Neuron, 
79, issue 5, pp. 836–48.
Kwisthout, J., Bekkering, H. and van Rooij, I. (2017). To be precise, the details don’t mat-
ter: On predictive processing, precision, and level of detail of predictions. Brain and 
cognition, 112, pp. 84–91.
Kwisthout, J., and van Rooij, I. (2015). Free energy minimization and information gain: 
The devil is in the details. Cognitive neuroscience, 6, issue 4, pp. 216–18.
Lewkowicz, D. J., Schmuckler, M. A. and Mangalindan, D. M. J. (2018). Learning of hier-
archical serial patterns emerges in infancy. Developmental psychobiology, 60, issue 3, 
pp. 243–55.
Magid, R. W ., Sheskin, M. and Schulz, L. E. (2015). Imagination and the generation of new 
ideas. Cognitive development, 34, pp. 99–110.
Muckli, L., De Martino, F ., Vizioli, L., Petro, L. S., Smith, F . W ., Ugurbil, K., Goebel, R. and 
Yacoub, E. (2015). Contextual feedback to superficial layers of V1. Current biology, 25, 
issue 20, pp. 2690–95.
Nagai, Y. (2019). predictive learning: Its key role in early cognitive development. Philo-
sophical transactions of the royal society B, 374, issue 1771, p. 20180030.
Neacsu, V ., Mirza, M. B., Adams, R. A. and Friston, K. J. (2022). Structure learning enhanc-
es concept formation in synthetic active inference agents. Plos one, 17, issue 11, p. 
e0277199.
Oaksford, M. and Chater, N. (2009). Précis of bayesian rationality: The probabilistic ap-
proach to human reasoning. Behavioral and brain sciences, 32, issue 1, pp. 69–84.
O’Reilly , J. X., Schüffelgen, U., Cuell, S. F ., Behrens, T . E. J., Mars, R. B. and Rushworth, M. 
F . S. (2013). Dissociable effects of surprise and model update in parietal and anterior 
cingulate cortex. Proceedings of the national academy of sciences, 110, issue 38, pp. 
E3660–69.
Oudeyer, P . and Smith, L. B. (2016). How evolution may work through curiosity of driven 
developmental process. Topics in cognitive science, 8, issue 2, pp. 492–502.
Perfors, A. (2012). Bayesian models of cognition: What’s built in after all? Philosophy com-
pass, 7, issue 2, pp. 127–38.
Piantadosi, S. T . (2021). The computational origin of representation. Minds and machines, 
31, pp. 1–58.
Piantadosi, S. T., Tenenbaum, J. B. and Goodman, N. D. (2012). Bootstrapping in a lan-
guage of hought: A formal model of numerical concept learning. Cognition, 123, issue 
2, pp. 199–217.
Piantadosi, S. T . (2016). The logical primitives of thought: empirical foundations for compo-
sitional cognitive models. Psychological review, 123, issue 4, p. 392–424.
Poli, F ., Serino, G., Mars, R. B. and Hunnius, S. (2020). Infants rtailor their attention to 
maximize learning. Science advances, 6, issue 39, p. eabb5053.
Ritchhart, R., Church, M. and Morrison, K. (2011). Making thinking visible: How to pro-
mote engagement, understanding, and independence for all learners. San Francisco: 
John Wiley & Sons.
Rutar

165 
Russell, S. J. (2010). Artificial intelligence a modern approach. New Jersey: Pearson Edu-
cation, Inc.
Rutar, D., Colizoli, O., Selen, L., Spieß, L., Kwisthout, J. and Hunnius, S. (2023). Differen-
tiating between Bayesian parameter learning and structure learning based on behav-
ioural and pupil measures. PloS one, 18, issue 2, p. e0270619.
Rutar, D., de Wolff, E. van Rooij, I. and Kwisthout, J. (2022). Structure searning in pre-
dictive processing needs Revision. Computational brain & behavior, 5, issue 2, pp. 
234–43.
Simon, T . J., Hespos, S. J. and Rochat, P . (1995). Do infants understand simple arithmetic? 
A replication of wynn (1992). Cognitive development, 10, issue 2, pp. 253–69.
Smith, R., Schwartenbeck, P ., Parr, T. and Friston, K. J. (2020). An active inference ap-
proach to modeling structure learning: Concept learning as an example vase. Fron-
tiers in computational neuroscience, 14, issue 41.
Spelke, E. S., Breinlinger, K., Macomber, J. and Jacobson, K. (1992). Origins of knowledge. 
Psychological review, 99, issue 4, p. 605.
Spelke, E. S. and Kinzler, K. D. (2007). Core knowledge. Developmental science, 10, issue 
1, pp. 89–96.
Srinivasan, M. V ., Laughlin, S. B. and Dubs, A. 1982. Predictive coding: A fresh view of in-
hibition in the retina. Proceedings of the royal society of London. Series B. Biological 
sciences, 216, issue 1205, pp. 427–59.
Summerfield, C. and de Lange, F . P . (2014). Expectation in perceptual decision making: 
Neural and computational mechanisms. Nature reviews neuroscience, 15, issue 11, 
pp. 745–56.
Ullman, T. D. and Tenenbaum, J. B. (2020). Bayesian models of conceptual development: 
Learning as building models of the world. Annual review of developmental psycholo-
gy, 2, pp. 533–58.
Urgen, B. A. and Miller, L. E. (2015). Towards an empirically grounded predictive coding 
account of action understanding. Journal of neuroscience, 35, issue 12, pp. 4789–91.
Vosniadou, S., Ioannides, C., Dimitrakopoulou, A. and Papademetriou, E. (2001). Design-
ing learning environments to promote conceptual change in science. Learning and 
instruction, 11, issue 4/5, pp. 381–419.
Vygotsky, L. S. and Cole, M. (1978). Mind in society: Development of higher psychological 
processes. Cambridge, MA: Harvard University press.
Werchan, D. M., Collins, A. G. E., Frank, M. J. and Amso, A. (2015). 8-Month-old infants 
apontaneously learn and generalize hierarchical Rrules. Psychological science, 26, is-
sue 6, pp. 805–15.
Williams, D. (2020). Predictive coding and thought. Synthese, 197, issue  4, pp. 1749–75.
Zhang, F ., Jaffe-Dax, S. Wilson, R. C. and Emberson, L. E. (2019). Prediction in infants and 
adults: A pupillometry study. Developmental science, 22, issue 4, p. e12780.
Developing higher cognition through predictive processing

166 Sodobna pedagogika/Journal of Contemporary Educational Studies
Danaja RUTAR (Nizozemska)
RAZVIJANJE VIŠJIH KOGNITIVNIH PROCESOV SKOZI NAPOVEDNO  
PROCESIRANJE
Povzetek: Napovedno procesiranje je vplivna teoretična paradigma v kognitivni znanosti, ki obljublja 
poenotenje vseh vidikov človeškega delovanja: percepcije, motoričnega delovanja in kognicije. Teor-
ija je sicer že uspešno razložila številne vidike zaznavanja in motoričnega delovanja, a ji, ključno, 
manjka koherentna teoretična razlaga kognitivnega razvoja in višjih kognitivnih procesov ter še bolj 
ambiciozno, razlaga razvoja višjih kognitivnih procesov. V tem prispevku raziskujem, v kolikšni meri 
je teorija prediktivnega procesiranja v sedanji obliki sploh primerna za razlago razvoja in višjih kogni-
tivnih procesov. Predlagam, da glede na sedanje stanje napovedno procesiranje ne more dovolj razložiti 
razvoja višjih kognitivnih procesov, saj mentalnim modelom, ki jih predlaga, manjkata dve ključni ses-
tavini, ki sta sicer pomembni za karakterizacijo višjih kognitivnih procesov: kompozicionalnost in gen-
erativnost. V tem članku raziskujem, kako bi lahko ti dve ključni funkciji obravnavali z vidika teorije 
prediktivnega procesiranja. Predlagam tudi, kakšna je vloga izobraževalnih in pedagoških praks v tem 
kontekstu. Nazadnje pa tudi razložim vzajemno razmerje med pedagoškimi praksami in napovednim 
procesiranjem: pedagoške prakse so ključnega pomena za izgradnjo mentalnih modelov, napovedno 
procesiranje pa nudi teoretični okvir za razumevanje psiholoških mehanizmov prisotnih v pedagoških 
kontekstih.
Ključne besede: napovedno procesiranje, razvoj višjih kognitivnih procesov, kompozicionalnost, gen-
erativnost, pedagoške implikacije 
Elektronski naslov: danaja.rutar@gmail.com
Rutar