Andy Clark presents a theory of predictive processing
which can be summarised as: no representation without prediction.[1]
This construes conscious mental contents as reflections of a predictive
processing that aims to optimise its accuracy by constantly checking for error
against incoming sensory input and representing the most likely hypothesis. One
major question for predictive processing is: what exactly are these hypotheses?
This essay will explore the place of language in predictive processing,
following Lupyan and Clark; and in particular, see how this might clarify the
otherwise underdetermined mental ontology implied by the predictive model.
Lupyan and Clark explicitly extended this model to include also words and
larger verbal constructions which create a “second system” of representation
which is projected into the world of perception.[2]
This system does not alter the basically determinate appearance of phenomenal
experience, rather it reveals aspects and creates objects both concrete and
abstract alike. The major benefit of this second system is not only this
structuring, which is also evident in the case of pre-linguistic schemas such
as tone perception, but rather that it entails a flexible set of predictive
priors which can be creatively combined and transmitted between individuals. We
will further describe how this aspect of predictive processing can be understood
as scaffolded via a cultural inheritance from our predecessors. Finally, this
notion of notion of an ‘artificial context’ constructed from linguistic priors
and scaffolded via a cultural inheritance will be illustrated with reference to
Kuhn’s account of scientific paradigms. To conclude, the predictive model will
itself be seen as part of the second system—whereby its indeterminate mental
ontology will be traced to the corresponding indeterminacy of ‘hidden layers’
in the neural network model from which it is derived.
It is
generally accepted that the brain exhibits hierarchical characteristics in its
inter-relation of mind and world. Where predictive processing differs is in its
reversing the typical relationship, wherein the contents of experience are
usually thought to flow from world to mind; instead it holds that the mind is
constructed by top-down processes.[3]
Sensory data is relevant only insofar as it registers as an error which forces
the model to cycle to an alternative hypothesis that has been rendered more
likely in light of this new piece of information. This cycling behaviour is
seen in the case of binocular rivalry as presented by Clark.[4]
This entails presenting a different image to each eye: say the image of a house
to the left; and the right, a cat. The experience which results is one in which
these two images cycle between one another relatively smoothly. The idea, says
Clark, is that the brain computes an equal probability of each but whenever one
is presented the other registers as an error signal and then selected only to
be replaced as a result of the same process. This continues back and forth
because the brain recognises that the two cannot be simultaneously present.
But this
argument from binocular rivalry can be compared another case outlined by Clark,
that of a blurry percept in the distance which could be either a dog or a fox.
There the image doesn’t alternate between hypotheses, our experience remains
that of a “unitary-coherent whole.”[5]
This wholeness is characteristic of immediate perceptual experience under
ordinary circumstance—whereas the case of binocular rivalry as outlined above
is obviously unusual. Clark also presents an even simpler instance of the same
phenomenon as observed in the dog and fox example.[6]
This is the image presented below, wherein the central percept can be
interpreted as either B (when read from left to right as A, B, C) or 13 (when
read from top to bottom as 11, 12, 13). And yet, whichever way we interpret it,
the percept itself clearly remains determinate:
What these
two cases seem to indicate is that that perception and understanding reflect
separable levels of predictive processing. We can accept that in perception
there is no representation without prediction, and yet in the meaning
attributed to symbols (as in the cases, as dog/fox or B/13) there is a line
between this unitary percept and the indeterminate understanding we may
attribute to it. Even here we find that our understanding is a unitary whole.
The underlying percept, however, is not transformed by our respective
understandings of it. Instead we can see this as an example of cognitive
penetration wherein our understanding is read into the percept. Clark explains
this as reflecting the various levels of predictions at work in the human
brain.[8]
This can be
further supported with reference to Tallon-Baudry and Bertrand’s finding of
neural oscillations correlated with object construction.[9]
They found a consistent pattern of activity—specifically, gamma oscillations in
the range of 40Hz as measured by MEG and EEG—is correlated with the perception
of objects and their understanding. This activity has also been linked to the
association of words with their meaning.[10]
Altogether these findings appear to reflect a neurological function which
operates in the cross-modal binding of knowledge in perception. This is
precisely what the predictive processing theory suggests we should find. We can
see in this activity a sign of the construction of a predictive representation
exactly as expected. The below images illustrate how this neural activity
further reflects not only the recognition of obvious percepts but also the
construction of illusory triangles: “There was no statistical difference in the
gamma range between the responses to the illusory and real triangles. Induced
gamma could thus reflect the spatial binding of the elementary features of the
picture into a coherent representation of the triangle.”[11]
We might
here turn for a moment to examine cognitive penetration in particular. One
difficulty in this area is what exactly counts as ‘cognitive’ and what aspects
of perception are ‘penetrated’—or even what this means. Here we will be
particular in our description, leaving aside cognitive penetration as a much
broader concept and focusing on a specific form. While one response may be to
say this particular form is not a case of cognitive penetration at all, that is
no difficulty for our argument. Either way, it is important to clarify the
specific phenomenon here under consideration. What we have been describing thus
can be distinguished from broader accounts of cognitive penetration in two main
ways. Firstly, we do not understand this form to entail any alteration of, for
instance, the visual percept itself. Instead what we describe is a cognitive
structure imposed on the conceptual indeterminacy of perceptual experience.
This entails, for instance, possibilities for identification and other
action—from as simple as referring to the aspect in language to purposefully
interacting with some particular part of a tool. The above
case of ‘seeing’ an illusory triangle also falls under this and illustrates how
prior knowledge of forms penetrates our perception without actually altering
the percept. Instead we perceive something like the superimposed phantasmic
form of, for instance, a triangle or other such abstract entity.
Where this
first difference applies to the ‘penetration,’ the second involves what is
meant by ‘cognitive.’ Macpherson, for instance, seems to have quite a narrow
idea of which cognitive aspects count for the purposes of identifying an
instance of cognitive penetration—specifically, “belief, knowledge or
concepts.”[13]
They suggest, for instance, that the lexical processing of linguistic forms may
not count as cognitive penetration insofar as these “do not depend on the
semantic processing of the lexical forms.”[14]
Their thought seems to be that cognitive penetration requires the specific
involvement of the aforementioned cognitive modules: beliefs, knowledge, concepts.
Instead, we are concerned with a sub-personal form of cognitive penetration
wherein the ‘cognitive’ component refers to structures derived from prior
knowledge, whether innate or learned. This is evident in the case of the
percept which can be read as either B or 13, wherein learned schemas of letters
and numbers ‘penetrate’ perception to the extent that they superimpose a
meaningful form on the otherwise indeterminate perceptual experience. We will
take cases like these as paradigmatic for the phenomenon which is here under
investigation.
Indeed,
cognitive penetration may be an inappropriate word for the phenomenon.
Something like ‘embedded meaning’ or ‘superimposed structure’ might better
capture this phenomenon. Whatever we call this, it seems certain that there is
some predictive construction of expected structures based on perceptual cues.
This predictive processing allows the mind to fill in details which are not
apparent in the bare perceptual data so as to infer hidden structures in the world—as,
for instance, the illusory triangle in the image above. As Clark has pointed
out, this is not a matter of conscious choice but occurs at an unconscious or
sub-personal level.[15]
It emerges through the interaction between generative models produced by the
cognitive mechanisms of predictive processing and error signals propagated
upwards from our sensory apparatus. The mental output reflects the interplay of
these elements in a way which we have here characterised as particular type of
‘cognitive penetration.’[16]
There are two main categories in which this specific phenomenon is evident:
percepts, bounded structures superimposed on bare perceptual experience, and
concepts, wherein a percept is further identified by a communicable name—we
will begin by addressing the former before turning to language.
Infants
in their first months after birth discriminate a wide range of speech
contrasts, both native and non-native. A process of perceptual reorganization
takes place over the first year, such that discrimination of most non-native
speech sounds deteriorates.[17]
While
infants younger than ten months, whatever their language exposure, are equally
able to discern tone-contrasts common to Hindi, the ability of English-learning
infants fell off between ten and twelve months—whereas that of Hindi-learning
infants remained stable.[18]
This perceptual reorganisation depends on the prior existence of an innate
capacity for learning, known as the Baldwin effect.[19]
This can be understood, on the predictive processing model, as an innate
capacity for tone distinctions which is then formed in line with the speech
sounds to which the infant is exposed. We can see this process of acquiring of
language-specific tonal schemas as learning a flexible set of generative models
which constitute the basic aural forms of the language—akin to when one learns
an alphabet from which many words can be made.
Here we
might note, moreover, that the perception of a relevant tone distinction in these
studies requires an observable indicator of awareness—e.g., turning their head
when they hear the relevant sound. Already we can start to see how perceptual
discrimination might entail meaning in the form of possible action. A further
example can be drawn from individuals who have absolute pitch—in other words,
those few who “can name or identify the pitch for a tone without any external
reference.”[20]
But people with absolute pitch more quickly identify certain pitches than
others, for instance, “the pitches C# and F# are more quickly identified than
D# and G#.”[21]
These variations have been found to align with the relative frequency of these
pitches in a large sample of music, wherein “pitches like C# and F# [were found
to] occur more frequently than pitches like D# and G#.”[22]
These
cases, absolute pitch and infant tone perception, can each be understood as
consistent with the predictive processing model. We would expect that more
frequently occurring percepts would be more quickly processed than uncommon
ones and this is precisely what is found. The case of absolute pitch differs
from that of infant tone perception in that it involves the explicit
identification by name of a specific aspect of perceptual experience. While the
infant tone perception studies all necessarily involve an action which
effectively amounts to identification, this is more readily seen as a primitive
form of communication rather than language proper. The identification of
pitches, in contrast, clearly involves something of a language—albeit of a
musical sort. All in
all, these cases are in agreement with predictive processing. In tone
perception and absolute pitch alike, familiar perceptual forms were
internalised as predictive schemas which can then be deployed in our ordinary
perception, cognition, and activity.
For our purposes, however, the most important aspect
here is that these involve naming and identifying the relevant percept. Lupyan
and Clark note, for instance, that while immediate percepts are always
particular:
… words and larger verbal
constructions are special kinds of perceptual input. While perceptual
experiences of, for example, vehicles are always experiences of specific
vehicles, the word “vehicle” is categorical.[23]
This leads them to suggest that language endows the
mind with a flexible set of contexts or hypotheses on which our minds may draw
to structure our predictive representation of the world. Language can thus be
understood as offering metabolically cheap generative models upon which our
ordinary predictive processes can be scaffolded. We can see language, on this
view, as involving the coupling of a percept in immediate experience with a
coherent concept as identified by a specific name. Take the following image,
for instance, how many specific colours can you identify?
Some may be
able to distinguish more or less specific colours, but this does not entail
that they perceive altogether different images. We can describe it as instead
involving predictive structuring of perceptual experience. This can be understood, like the
prior example of A, B, C or 12, 13, 14, as involving a higher-level generative
model which leaves the underlying percept the same while also projecting a
meaningful structure upon it. This exercise can be compared to the experience
of tasting wine, wherein a subject with some expertise in wine tasting may be
able to discern qualities that fade into the background for an unexperienced
subject. The same might be found in comparing, for instance, graphic designers
or artists to the general populace on the example of colour as above.
But where are the boundaries between the identified
colours? It is difficult to say there are any clear lines of demarcation in the
above image, though we can nevertheless identify specific colours: red, green,
purple, orange, etc. This possibility is precisely what is expected on the
predictive processing account of language. To begin, Lupyan and Clark’s line
runs counter to the standard view of language—namely, that “although different
languages provide their speakers with different ways of talking about things,
these differences have nothing to [do] with how we think about or perceive
things.”[24]
Instead, Lupyan and Clark suggest a theory derived from the cognitive
penetrability of perceptual experience made possible by predictive processing:
Language not only functions as a
means of communicating our thoughts but plays an active role in shaping them.
Rather than passively reflecting the joints of nature, words and larger
constructions help to carve joints into nature.[25]
This idea
resembles that of the Sapir-Whorf hypothesis, and indeed one of Clark’s own
examples can be rejigged to make this point. The following image has been repeated
from above, but this time suppose you are only familiar with a base-5 number
system: 0, 1, 2, 3, 4,
10, 11, 12, 13, 14, 20, etc.
While the left to right line will read the same as
before (A, B, C) the top to bottom line, while superficially identical, now
entails an altogether different meaning. For this individual reading in base-5
from top to bottom the meaning is now equivalent, when translated to base-10,
to the series: 7, 8, 9. This can be seen to support the Sapir-Whorf hypothesis,
insofar as alternative mathematical languages give rise to incompatible
understandings of an identical perceptual experience. While we will not here
examine Sapir and Whorf’s argument in any detail, it is worth noting the
similarity between the two and, moreover, that a predictive processing model
seems capable of providing a mechanism that could readily underlie this
hypothesis.
Eskimo, for instance, are said to have a much greater
repertoire of words for various shades of snow than can be found in English.[27]
Lupyan and Clark suggest this array of specific verbal constructions provides
them with the capacity to better distinguish between shades of snow. This makes
sense insofar there is little use for such an extensive array of words unless
there is also an underlying perceptual structuring that provides the
discriminatory capacity required for their meaningful use.
This example also supports the notion that errors are
relative to the environmental context in which embodied activity occurs.
Similar to how Clark sees predictive processing as involving organism-dependent
priors, we might suggest in line with the Sapir-Whorf hypothesis that
environmental differences will also result in differences between linguistic
communities within a species. For us it is of little importance whether we can
discern between various shades of snow, whereas the Eskimo may have good reason
for desiring greater discernment in order to predict and communicate further
predictive models for climatological conditions which could bear directly on
their survival. We can further supplement this line with findings from
evolutionary game theory, specifically on the relation between evolutionary
dynamics and veridical percepts:
Natural selection can send
perfectly, or partially, true perceptions to extinction when they compete with
perceptions that use niche-specific interfaces which hide the truth in order to
better represent utility. Fitness and access to truth are logically distinct
properties. More truthful perceptions do not entail greater fitness.[28]
This line indicates, in other words, that our set of
hypotheses are likely to be specifically human and determined by our particular
evolutionary history and environmental context. Some of these
organism-dependent schemas have been further reified as concepts and words,
which—via the mechanism of cognitive penetration that we have outlined—are
embedded in coherent aspects of perceptual experience and thereby coupled with
their corresponding percepts.
Taken together, we can see that the language of a
group can be expected to vary according to its utility in ordinary activity
given their context and circumstances. This is precisely what has been found in
comparing the Eskimo words for shades of snow with the relatively blurry
structure of those acculturated in environments where snow is rare or even
absent.[29]
What matters here, as Clark repeatedly emphasises, is the possibility for
perception and activity which these structural schemas allow.[30]
There are, moreover, two centrally important aspects to the specifically
linguistic components of this predictive structuring of experience. First, that
our linguistic reality provides a set of flexible priors which allow us to more
quickly adapt to change in the material environment. Second, that these
linguistic priors are communicable and may thereby save individuals and even
whole generations from needing to encounter painful errors.
This leads to a picture in which language is a central
mechanism in our adaptive activity in the world, one which aids our capacity to
distinguish and respond to signals in the perceptual environment and, most
significantly, allows us to communicate this knowledge with others in our
linguistic community:
Exposure
to language (whether shared or self-produced) thus becomes a potent and
fundamentally unified means of exploring and exploiting the full potential of
our own acquired knowledge about the world—a kind of artificial “second system”
enabling us to take full advantage of our own knowledge as well as the
knowledge of others.[31]
There are thus two levels of representation in the
predictive processing model, that of the unitary-coherent whole and the aspects
that can be grasped as abstract entities—this second level is our
understanding, not all of which is conscious. Our discussion is limited to this
second system with a focus on its linguistic aspect. This is particularly
important, as noted in the above excerpt, insofar as the communicability of
linguistic priors allows us to draw on the knowledge of others. Following this
line, therefore, we will now turn to focus on the ways in which this predictive
model of linguistic reality aligns with an image that extends beyond the bounds
typically thought to circumscribe the mind.
Clark and Lupyan thus present an account of predictive
processing and language in which the internalised schemas of linguistic
understanding form a predictive structure which is then coupled with
identifiable percepts. We can understand this example as reflecting the way in
which cognitive development entails acquiring a set of culturally-determined
schemas that superimpose structure upon our experience of the world. On the
predictive processing view, these schemas can be understood as higher-level
priors which provide an efficient and flexible set of hypotheses for
structuring the world. This aligns with Kruschke’s argument for the parsimony
of hypothesis space:
Entertaining an infinite space of
hypothetical values does not imply the need for an information processor of
infinite capacity, for infinite belief distributions can be represented with
small sets of values.[32]
And indeed, this possibility is exactly what Lupyan
and Clark have also suggested:
Words (and larger verbal
constructions) become not simply ways to communicate our preexisting thoughts
but highly flexible (and metabolically cheap) sources of priors throughout the
neural hierarchy.[33]
Most importantly, this is a metabolically cheap method
and further aligns with the notion that predictive processing seeks to minimise
the use of energy.[34]
We should note, moreover, that the ‘free energy principle’ has been used to
explain the error minimisation imperative that is often taken as central to
predictive processing accounts of the mind. If this is the case, then we might
plausibly expect scenarios in which these two aims could contract one another.
One way this might be possible is through what
Sterelny calls the ‘scaffolded mind.’[35]
This view is particularly relevant for our current purposes, insofar as
language and other larger constructions deployed by our predictive processing
system are an inter-generational enterprise. Language allows not only a set of
flexible priors but, as Lupyan and Clark emphasise, these priors are also
communicable between individuals and can be transmitted across generations. We
do not, for instance, reinvent the meaning of words anew each generation—though
they may vary slightly across generations, within a lifetime, or even between
locations. Instead these meanings are preserved and passed on, hence the idea
of the mind as scaffolded by the language and linguistic constructions we
inherit from our predecessors:
… the cognitive competence of
generation N+1 individually and collectively depends on cognitive provisioning
by generation N. The most critical, mind-and- brain-shaping environmental
supports for cognition are these cumulatively built, collectively provided
tools for thinking, tools that are provided to many or all of a generation by
many or all of the previous generation.[36]
This notion of the ‘scaffolded mind,’ taken together
with the Lupyan and Clark’s characterisation of language as a set of flexible
and communicable priors, provides convincing account for how linguistic models
by which predictive processing structures much of our representation of the
world may constitute a cultural inheritance.
We can see this, moreover, as coherent with something
like a ‘free energy principle’ but at the level of a linguistic community or
species; it allows the inter-generational transmission of cultural and
linguistic priors between individuals and across generations, which can thus be
acquired far more cheaply than would be possible were each individuals required
to encounter or observe errors on their own. And yet here we may also note the
possible tension which was foreshadowed earlier. The ‘scaffolded mind’
hypothesis is similar to what has been described as the ‘ratchet effect’ or
‘cumulative culture.’[37]
This is typically seen as a peculiarly human characteristic, one which explains
our cultural and technological advancement. All this has been possible because,
rather than starting afresh, each generation has been able to build on the
efforts of their predecessors. One factor which has been suggested as crucial
in this regard is observational learning or, more specifically, imitation.
Indeed, some have even described this trait in humans as
‘over-imitation’—wherein we will copy even unnecessary steps that, for
instance, a chimpanzee would not.[38]
Imitation can be contrasted with emulation, wherein the aim is to replicate the
purpose of behaviour without necessarily copying its precise form. Humans, in
contrast, tend to copy the form of behaviour even where certain steps seem
plainly superfluous. The idea is that this over-imitation has been vital for
ensuring the fidelity with which we transmit knowledge from one generation to the
next—and hence, our capacity for cumulative culture.[39]
Our journey to this point begun with the predictive
model presented by Clark, then followed the line which Lupyan and Clark draw
from this basic model to the linguistic ‘second system.’ The
fulcrum of this system is the phenomenon of ‘cognitive penetration’ wherein the
structure of our understanding is embedded in or superimposed upon our
experience of the world. We have further described this as a ‘scaffolded’
system, whereby the transmissibility of language and culture enables each
generation to build upon the knowledge of their predecessors. As a result, our
linguistic reality is historically-determined—which is another sense in which
cognition can be accurately characterised as ‘situated.’ Indeed, that minds are somehow
situated is emphasised throughout the predictive model—and in particular,
within the language extension provided by Lupyan and Clark. While this
‘situated cognition’ is something broader, we will instead enumerate here the
various ways in which the mind can be said to be situated on this account. For
one, there is the fact that we have yet to find any actual instance of a
disembodied mind. Instead everywhere the mind is found not as any disembodied
universal but as a particular referent which corresponds to the embodiment of
an existing human being. That the mind thus corresponds, whatever the nature of
this relation, to a physical body here suggests the second way in which it is
situated. This is that minds are always situated at a determinate point in time
and space, and though they may move from here to there in time this history is
always carried with them.[40]
Thirdly, we carry not only our own history but that of our predecessors insofar
as the ‘second system’ is scaffolded by the transmissibility of linguistic
priors between individuals.[41]
And finally, all these individual particularities give rise to the ‘artificial
context’ which conditions the representation and activity of an existing human
mind.[42]
All this can be seen, for instance, in the activity of
individuals within the scientific enterprise and, more specifically, the
progress of scientific paradigms as outlined by Kuhn.[43]
The work of science is obviously scaffolded, insofar as generations build on
the work of their predecessors. This work takes place within a ‘scientific
paradigm’ which can be seen as a large-scale structure within the second system
of practicing scientists. Kuhn describes, for instance, the Ptolemaic and
Copernican models of astronomy as paradigms. These determine the
representational structure of the field which is being investigated—defining
central concepts, relevant questions, evidential standards, etc. All this
provides the basic framework for standard scientific practice and can be
understood as predictive hypotheses or larger verbal constructions, in line
with the notion of an “artificial context” as in Lupyan and Clark.[44] The
predictive aspect of this also explains a crucial fact: that science proceeds
first by the positing of hypotheses and only then by the testing for errors.
This activity can be seen to reflect the predictive model of mind precisely,
insofar each is an essentially predictive activity entailing specific
hypothetical representations which condition activity. Of course, what we are
concerned with here is not the specific hypothesis which the scientist
understands his experiment as testing. Instead we are concerned with the
‘scientific paradigm’ or ‘second system’ in the mind of this particular
scientist which entails the predictive priors which condition their scientific
activity.
Here we must further differentiate, therefore, between
the technical knowledge that is produced by scientists and the practical
knowledge that structures their activity, including the production of technical
knowledge—writing books, giving lectures, discussing problems, etc. These two
aspects are obviously separable only in abstract insofar as to write a paper,
for instance, on even the most technical of physical phenomenon takes as basic
the practical knowledge of language necessary to do so. But when we are
speaking of scientific paradigms, we do not mean these in their technical or
textual form as in a history of ideas. Instead we mean the practical knowledge
that underlies scientific activity—as much for communication as experiment—and
which is internal to the mind of a scientist situated within that paradigm.
This difference can be seen most clearly in what Kuhn describes as a
‘scientific revolution’ or ‘paradigm shift’ wherein the movement is not only
technical but rather entails a radical reorganisation of the predictive priors
which constitute the paradigm upon which scientific practice had thus far been
scaffolded:
Examining the record of past
research from the vantage of contemporary historiography, the historian of
science may be tempted to exclaim that when paradigms change, the world itself
changes with them. Led by a new paradigm, scientists adopt new instruments and
look in new places. Even more important, during revolutions scientists see new
and different things when looking with familiar instruments in places they have
looked before.[45]
Predictive processing thus readily accounts for the
mechanism by which this second system of linguistic priors is represented by
our minds—whereby the predictive system is seen, in other words, as the engine
of thought. And yet nevertheless this seems to tell us little about the actual
vehicle of language or how it might work. This may partly stem from the nature
of Clark’s theoretical enterprise, which is derived from the neural network
model of machine learning. It is an account, in other words, of how the
mind—understood as a computer—deals with uncertainty in the world, wherein
machine learning is taken as the paradigmatic instance of this computational
understanding of mind. Some understanding of neural networks, therefore, may be
helpful in properly grasping the predictive model. But here we encounter an
unexpected problem:
… [a neural network is] a type of
artificial intelligence (AI) that is modelled on the brain, and that promised
to be better than standard algorithms at dealing with complex real-world
situations. Unfortunately, such networks are also as opaque as the brain.
Instead of storing what they have learned in a neat block of digital memory,
they diffuse the information in a way that is exceedingly difficult to
decipher.[46]
This is what computer scientists working with
artificial neural networks have described as the ‘black box’ problem. While we
can more or less understand the basic computational structure, its internal
activity is obscured with only inputs and outputs being susceptible to any
clear representation. This is particularly troubling insofar as the cognitive
science from which predictive processing emerges was initially motivated by a
desire to overcome another black box, that of behaviourism.
That the predictive processing model shares in the
opacity of the neural networks from which it is derived is reflected in the
indeterminacy with which hypotheses are characterised by theorists. They are
described variously as ‘predictive priors’ or ‘generative models’—or broadly,
as an “entire spectrum” or “infinite space of hypothetical values.”[47]
One way to understand this is to see the neural network model as providing the
paradigm which underlies predictive processing. Through this metaphorical
framing of the mind we can thus map our concrete understanding of computers—in
this case, specifically of artificial neural networks—so as to structure our
understanding of the more abstract domain of the human mind.[48]
And yet as much as this highlights aspects which are coherent with the
structure of an artificial neural network, in that the mind also uses prior
knowledge to extract relevant information from noisy data, it also hides those
aspects of the underlying domain of mind where the metaphorical structuring
maps poorly. One such underdetermined aspect is the predictive ontology of
mind, which in the case of artificial networks is limited to obvious inputs and
outputs with many ‘hidden layers’ between the two.[49]
Hence the indeterminate mental ontology of predictive processing may stem from
the analogous indeterminacy that characterises the hidden layers of artificial
neural networks.
Nevertheless, this is not an insurmountable problem
for predictive processing. The line taken by Lupyan and Clark, for instance,
goes some way to determining at least one aspect of this mental ontology. As we
have outlined, language can thus be seen as providing a flexible set of
predictive priors. The transmissibility of language, moreover, also indicates
the source of at least some of the generative models on which predictive
processing depends. This solves the problem of whether, in Clark’s top-down
predictive model, the entire catalogue of generative models needs to be innate.
While some innate aptitude for language is still required, the developmental
acquisition of a set of scaffolded priors susceptible to creatively recombination
is surely more plausible than the almost Platonic notion that we are endowed at
birth with an infinite spectrum of hypotheses corresponding to everything in
existence. Some such acquisition as Lupyan and Clark allow seems necessary that
we might, for instance, attain the computational metaphor which serves as a
basic paradigm for cognitive science. There is reason to suspect, however, that
this addition of communicable linguistic priors will require some adjustment to
the basic predictive model insofar as the acquisition of language clashes with
Clark’s top-down picture of predictive processing. And yet more broadly, this
emphasis on linguistic priors further suggests a fruitful line of research
whereby the cognitive science of predictive processing might be combined with,
for instance, the cognitive linguistics of metaphor to yield a deeper
understanding of this second system.[50]
Castelvecchi, Davide. "Can we open the black box
of AI?." Nature News 538, no. 7623 (2016): 20.
Clark, Andy, and David Chalmers. "The extended
mind." analysis 58, no. 1 (1998): 7-19.
Clark, Andy. Surfing uncertainty: Prediction,
action, and the embodied mind. Oxford University Press, 2015.
Clark, Andy. "Beyond the 'Bayesian blur':
predictive processing and the nature of subjective experience." Journal
of Consciousness Studies 25, no. 3-4 (2018): 71-87.
Hurford, James R. "The evolution of the critical
period for language acquisition." Cognition 40, no. 3
(1991): 159-201.
Huron, David. Sweet anticipation: Music and
the psychology of expectation. MIT press, 2008.
Kay, Paul, and Willett Kempton. "What is the
Sapir‐Whorf hypothesis?." American anthropologist 86, no.
1 (1984): 65-79.
Kuhn, Thomas S. The structure of scientific
revolutions. University of Chicago press, 2012.
Lakoff, George, and Mark Johnson. Metaphors we
live by. University of Chicago press, 2008.
Lakoff, George, and Mark Johnson. Philosophy
in the flesh: The embodied mind and its challenge to western thought. Vol.
640. New York: Basic books, 1999.
Lewis, Hannah M., and Kevin N. Laland.
"Transmission fidelity is the key to the build-up of cumulative
culture." Philosophical Transactions of the Royal Society B:
Biological Sciences 367, no. 1599 (2012): 2171-2180.
Lupyan, Gary, and Andy Clark. "Words and the
world: Predictive coding and the language-perception-cognition
interface." Current Directions in Psychological Science 24,
no. 4 (2015): 279-284.
Mark, Justin T., Brian B. Marion, and Donald D.
Hoffman. "Natural selection and veridical perceptions." Journal
of Theoretical Biology 266, no. 4 (2010): 504-515.
Mattock, Karen, and Denis Burnham. "Chinese and
English infants' tone perception: Evidence for perceptual
reorganization." Infancy 10, no. 3 (2006): 241-265.
Piaget, Jean. The construction of reality in
the child. Vol. 82. Routledge, 2013.
Prinz, Jesse. "Ways of mindmaking." Andy
Clark and His Critics (2019): 222.
Sterelny, Kim. "Minds: extended or
scaffolded?." Phenomenology and the Cognitive Sciences 9,
no. 4 (2010): 465-481.
Tallon-Baudry, Catherine, and Olivier Bertrand.
"Oscillatory gamma activity in humans and its role in object
representation." Trends in cognitive sciences 3, no. 4
(1999): 151-162.
Tennie, Claudio, Josep Call, and Michael Tomasello.
"Ratcheting up the ratchet: on the evolution of cumulative
culture." Philosophical Transactions of the Royal Society B:
Biological Sciences 364, no. 1528 (2009): 2405-2415.
Tomasello, Michael. "Cultural transmission: A
view from chimpanzees and human infants." Journal of cross-cultural
psychology 32, no. 2 (2001): 135-146.
Werker, Janet F., and Richard C. Tees.
"Cross-language speech perception: Evidence for perceptual reorganization
during the first year of life." Infant behavior and development7,
no. 1 (1984): 49-63.
Wiese, Wanja, and Thomas Metzinger. "Vanilla PP
for philosophers: A primer on predictive processing." (2017).
[1]
Clark, Surfing Uncertainty.
[2]
Lupyan & Clark, ‘Words and the world: predictive coding and the language-perception-cognition
interface.’
[3]
Clark, Surfing Uncertainty.
[4]
Clark, ‘Beyond the Bayesian blur,’ p. 78–80.
[5]
Ibid, p. 73.
[6]
Ibid, p. 82–83.
[7]
Image sourced from Lupyan & Clark, ‘Words and the world,’ p. 282.
[8] Clark, ‘Words and the world,’ p. 281.
[9] Tallon-Baudry & Bertrand, ‘Oscillatory gamma activity in humans and its role in object representation.’
[10]
Ibid, p. 157.
[11] Ibid, p. 156.
[12]
Image sourced from Tallon-Baudry & Bertrand, ‘Oscillatory gamma activity in
humans and its role object representation,’ p. 156.
[13]
Macpherson, p. 578.
[14]
Macpherson, p. 578.
[15]
Clark, Surfing Uncertainty.
[16]
Lupyan & Clark, ‘Words and the world,’ p. 281.
[17]
Mattock & Burnham, ‘Chinese
and English infants’ tone perception,’ p. 241.
[18]
Werker & Tees,
‘Cross-language speech perception.’
[19]
Huron, p. 62.
[20]
Huron, p. 64.
[21]
Ibid.
[22]
Ibid.
[23]
Lupyan & Clark, ‘Words and the world,’ p. 283.
[24]
Lupyan & Clark, ‘Words and the world,’ p. 282.
[25]
Ibid, p. 282.
[26]
Image sourced from Lupyan & Clark, ‘Words and the world,’ p. 282.
[27]
Lupyan & Clark, ‘Words and the world,’ p. 282.
[28]
Mark, Marion, & Hoffman, ‘Natural selection and veridical perception,’ p.
513.
[29]
Lupyan & Clark, ‘Words and the world,’ p. 282.
[30]
Ibid, p. 283.
[31]
Ibid, p. 283.
[32]
Kruschke in Clark, ‘Beyond the Bayesian blur,’ p. 73.
[33]
Lupyan & Clark, p. 283.
[34]
Wiese & Metzinger,
‘Vanilla PP for philosophers,’ p. 12.
[35]
Sterelny, ‘Minds: extended or
scaffolded?’
[36]
Ibid, p. 479.
[37] Tennie, Call & Tomasello, ‘Ratcheting up the ratchet: on the evolution of cumulative culture.’
[38] Tomasello, ‘Cultural transmission: A view from chimpanzees and human infants.’
[39]
Lewis & Laland, ‘Transmission fidelity is the key to the build-up of
cumulative culture.’
[40]
This is evident, for instance, in the role of prior knowledge in predictive
processing—we are thus situated within our learning history.
[41]
Sterelny, ‘Minds: extended or scaffolded?’
[42]
We can see this in how predictive priors structure our experience, or more
specifically in the linguistic relativity described by Lupyan and Clark—as in
the Sapir-Whorf Hypothesis.
[43]
Kuhn, The Structure of
Scientific Revolutions.
[44]
Lupyan & Clark, ‘Words and the World,’ p. 283.
[45]
Kuhn, p. 111.
[46]
Castelvecchi, ‘Can we open the black box of AI?’
[47]
Kruschke in Clark, ‘Beyond the Bayesian blur,’ p. 73.
[48]
Lakoff & Johnson, Metaphors
We Live By, p.
[49]
Castelvecchi, ‘Can we open the
black box of AI?’
[50]
See Lakoff & Johnson, Philosophy
in the Flesh.