Content uploaded by Suspended Reason
Author content
All content in this area was uploaded by Suspended Reason on May 14, 2020
Content may be subject to copyright.
Predictive Hermeneutics
Suspended Reason and Tom Rutten
The first step in the direction of truth is to understand the
frame and scope of the intellect itself, to comprehend the
act itself of comprehension. [...] The first step in the
direction of beauty is to understand the frame and scope
of imagination, to comprehend the act itself of esthetic
comprehension.
—Joyce, A Portrait of the Artist as a Young Man
If I am to listen to your esthetic philosophy give me at
least another cigarette.
—ibid.
ABSTRACT
Predictive processing, a meta-theory from cognitive
science and computational neuroscience, proposes the
mind as a hierarchical system which predicts reality as it
unfolds (Clark 2016, Hohwy 2013, Friston 2006). This
prediction occurs at levels ranging from the “merely”
sensory to the highly conceptual. Here, we extend this
thesis in order to understand the hermeneutic process as
it relates to textual and artistic encounters. We argue that
one of the foundational functions of contemporary
artistic and literary production is to inform and exploit
the mind’s predictive system. We further show how this
conceptualization of the art encounter as a cognitively
rich interaction between the artwork and mind, coupled
with the predictive processing framework of cognition,
helps explain a host of traditional literary, aesthetic, and
art historical values, including ambiguity,
defamiliarization, and reversal.
1
TABLE OF CONTENTS
1 Into the predictive mind
1.1 Historical precedents
1.2 From predictive coding to Bayesian error minimization
1.3 HBMs as instantiations of hierarchical predictive structures
1.4 Predictive coding and the schema
1.5 A brief note on art and predictive coding
2 Predictive hermeneutics
2.1 The artwork as puzzle
2.2 Bayesian formulation of a hermeneutic encounter
2.3 Interpretation and intentionality
3 Art, prediction, and subversion
3.1 Schematic subversion
3.2 The artwork as a joke
3.3 Charged subjects and formal gestures
3.4 Hermeneutic revision and effect ideas
4 Characteristics of schema-subversive art
4.1 Art as superstimulus
4.2 Truthiness and adherence to model
4.2.1 Resonance
4.2.2 Credibility
4.3 Surprisal as situationally valenced
4.5 Art as high gain or “weighted”
5 Further applications to aesthetic theory
5.1 Accessibility
5.2 Ideal vs. implied readers
5.3 Hyperpriors
5.4 Defamiliarization
5.5 Ambiguity
5.5.1 Temporal decay in probability mapping
5.5.2 Ambiguity of priority and grounding
5.6 Predictive coding and music
5.7 Mass culture and mutual error minimization
5.8 Design as the antithesis of art
5.9 Future directions?
6 Glossary
7 Works Cited
2
§1 Into the predictive mind
1.1 Historical precedents
The schematism by which our understanding deals
with the phenomenal world... is a skill so deeply
hidden in the human soul that we shall hardly
guess the secret track that Nature here employs.
— Kant, Critique of Pure Reason
We can think of the inferential or “predictive” mind as an
elaborate feedback system of induction and deduction:
pulling patterns out of data and interpreting data based on
previously pulled patterns (i.e. acquired knowledge).
Though the problem of inference is “as old as recorded
Western thought,” (Griffiths, Kemp, and Tenenbaum, n.d.)
we attempt here to sketch out a handful of precedents, of
particular relevance to this paper, for the
cognitive-scientific predictive mind, taken from
philosophy, psychology, hermeneutics, and aesthetics.
Both Kant and Piaget worked prominently with the term
schema, referring to a pattern of thought that organizes
information (categorically, relationally, etc.) into a mental
structure or “interpretive framework” that influences
attention and learning (Swanson 2016; Beni 2017). The
schema anticipates predictive processing models by
arguing that an organism’s understanding of reality
requires a reconciliation of top-down models with
bottom-up data. In Driven By Compression Progress
1
(2009), Schmidhuber notes a forerunner of his “cognitive
compression” work (an alternate formulation of predictive
coding ) in Piaget's theories of explorative learning and
2
assimilation, where “new inputs [are] embedded in old
schemas” (Schmidhuber 2009). Psychologist Alison
Gopnik has also taken up this frame, arguing that
children’s cognitive development occurs in a manner
strikingly similar to shifts in scientific theory, and can be
understood as the accumulation—in an approximately
1 Kant’s philosophical work may anticipate predictive processing in ways
beyond the concept of top-down modeling. Swanson (2016) notes that the
concepts of hyperpriors (see §5.6) and generative models, among others, also
stem from the German philosopher. Fazelpour and Thompson (2015) even
go so far as referring to cognitive science’s predictive mind as a “Kantian
brain.” Additionally, there is some speculation of an indirect influence by
Kant on the development of predictive processing theories, specifically via
Hermann von Helmholtz’s inference machine. For discussion, see (Swanson
2016; Beni 2017) and (Swanson 2016; Beni 2017).
2 See Schmidhuber (2009) Appendix A1, “Predictors vs. Compressors”
Bayesian manner—of informal theories about the world as
a structured causal system (Gopnik 1996).
Psychological set, and its sub-concept perceptual set,
originated in the 1950s and 60s as a description of how
situational interpretations can be highly influenced by
expectations, and how perception is best understood as an
“active process involving selection, inference, and
interpretation” (Hill 2001). The perceiver has expectations
and allocates attention deliberately (referred to as the
selector function of perception). He also knows how to
classify, understand, and name selected data, as well as
draw inferences from them (interpreter function). Factors
influencing perceptual set include expectations, emotions,
motivation, and culture. One of the more famous
experiments on perceptual set involves the numbers
12-13-14 written vertically in criss-cross with the
horizontally written figures A-13-C. readers exposed to the
vertical figures interpreted the middle figure as the
numeral “13”; readers exposed to the horizontal array
interpreted the middle figure as the letter “B” (Bruner and
Leigh Minturn 1955). Perceptual set was anticipated by the
Gestalt psychology of the early 20th century, which argued
that the mind actively organizes incoming perceptions,
through their interrelation, into gestalt wholes.
3
Around the same time, the emerging field of cybernetics
had begun “framing purposive behavior” in systems of all
kinds as “governed by feedback,” the seeking out of
novelty with which to fuel said feedback system (Turner
and Larson 2015). In literary theory meanwhile,
hermeneutic work was undergone by thinkers like
Hans-Georg Gadamer, Wolfgang Iser, and Ernst
Gombrich, work which helps inform our own schematic
theorizing of art encounters. As noted by L. Kesner,
Gadamer’s hermeneutical scenario highlights the “ongoing
and dynamic” quality of a textual encounter, whereby parts
inform whole informs part and the whole history of the
reader is activated in the exchange (Kesner 2014).
Gombrich (1960), meanwhile, emphasized the role of the
schema in both the production and reception of visual
works.
3 Parallels to Gestalt theory are one of many bridges (alongside similarities
between machine learning processes and theories of predictive coding in
human minds) to Peli Grietzer’s “Theory of Vibe,” which analogizes
modernist art to the representational process of a trained autoencoder
(Grietzer 2017).
3
In §1.2, we explain the fundamentals of predictive coding,
prediction error minimization (PEM), and its speculated
extension via concepts of Bayesian inference. Section 1.3,
which focuses on the interfacing of PEM and hierarchical
Bayesian models, will be especially of pertinence to those
with research interest in high level predictive processing
and the Bayesian brain. Non-technical readers, primarily
interested predictive processing as a lens into art
interactions, may benefit from skipping to §2. A glossary
section at the conclusion of the paper provides a references
for Bayesian and predictive terms which may be
unfamiliar.
1.2 From predictive coding to error minimization
Predictive models of perception originated in
computational neuroscience in the late 90s a way of
explaining low-level sensory processing such as in the
visual and auditory fields (Rao and Ballard 1999).
However, cognitive philosopher Andy Clark and
neuroscientist Karl Friston have proposed that predictive
processing and error minimization explains cognition at all
levels of the brain and, through Friston’s free energy
principle, constitute a fundamental property of life (Clark
2013; K. J. Friston 2013; Clark 2017).
Here, we proceed from several stipulations of these
theories: one, that the predictive dynamics theorized of
lower-level sensory processing also operate at higher
levels of cognition such as conceptual learning, and two,
that the inference systems central to lower-level predictive
coding are similarly in play at higher levels of cognition.
Justification for this extension of predictive coding from
the merely sensory into the conceptual can be found in
(Hohwy 2013; Clark 2013; K. J. Friston 2013; Clark
2017), under the term predictive error minimization, Clark
(Hohwy 2013; Clark 2013; K. J. Friston 2013; Clark
2017), (Hohwy 2013; Clark 2013; K. J. Friston 2013;
Clark 2017) under the handle predictive processing, and
(Hohwy 2013; Clark 2013; K. J. Friston 2013; Clark 2017)
as free energy. As noted in (Clark 2016), it has also been
dubbed active inference in theories emphasizing its
extension to motor control, and hierarchical layered
4
4 Through movement and motor control, a predictive agent can actively
control what its senses will encounter and use that information to confirm,
deny, or refine hypotheses about its current situation. This situates the
agent in a causal loop from senses to action and vice versa reminiscent of the
cybernetic loop described in §1.1. For a particularly good exposition of this
dynamic see (Godfrey-Smith 2016), chapter 4.
coding by those emphasizing the organizational hierarchies
believed to structure the different levels of prediction
work.
In this paper, we’ll use the term prediction error
minimization (PEM) as an umbrella term for
cognitive-scientific models (such as Hohwy, Clark, and
Friston’s) postulating that the brain actively actively uses
predictive structures, organized hierarchically from
low-level sensory to high-level conceptual domains, to
anticipate future events. Moreover, PEM theorizes that,
through a self-evaluation of its own anticipatory response,
the mind builds accurate models of the world by reducing
the amount of error in its predictions of sense data. For our
purposes, predictive error minimization is congruous with
Clark’s conception of hierarchical predictive processing.
We can understand PEM as a “systematic bridge linking
three of our most promising tools for understanding mind
and reason: cognitive neuroscience, computational
modelling, and probabilistic Bayesian approaches to…
evidence and uncertainty” (Clark 2013). PEM conceives of
the mind as a system organized in a generally hierarchical
structure which constantly aims to predict sensory input
and learns from the incongruencies between its guesses
and the sensory input it experiences.
To predict the incessant stream of sense-data, the system
aims to capture in its own hierarchical structure the
statistical structure of reality, or the set of causes which
produce the incoming sense-data. Because sense-data can
take complex forms and arise from complicated networks
of causes (e.g. social interactions and cinema), the mind’s
hierarchical predictive system must be able to encode
high-level knowledge (i.e. store high-level patterns which
explain and therefore compress multiple phenomena ) to
5
make predictions (Clark 2013). Predictions at a given level
of the hierarchical structure are in turn anticipated at
5 By storing the high-level patterns or probability distributions that explain a
host of observed phenomena, a predictive system can avoid storing the
complete set of observed phenomena themselves. Additionally, this system
at certain points need only record the error between sensory inputs and its
prediction about those inputs, rather than the inputs in their entirety. From
Surfing Uncertainty: “As long as there is detectable regularity, prediction
(and hence this particular form of data compression) is possible. It is the
deviations from what is predicted that then carry the ‘news’, quantified as
the difference (the ‘prediction error’) between the actual current signal and
the predicted one. This affords major savings on bandwidth, an economy
that was the driving force behind the development of the [predictive
coding] techniques by James Flanagan and others at Bell Labs during the
1950s” (Clark 2016).
4
higher levels, cycling back. At the lowest level of the
predictive structure, the system aims to anticipate sensory
data. These predictions provide a probabilistic evaluation
of the sense-data sampled at each moment using
knowledge represented across all layers.
6
As described in Clark (2013, 2016) this predictive
structure is hierarchical as well as bidirectional. Two-way
connections between levels in the hierarchy crucially
enable the system to learn from its prediction errors.
Downward (or forward, in the statistics literature)
connections between layers make successive predictions
about, i.e. define a probability distribution over, lower
layers. These successive predictions cascade to the bottom
7
of the hierarchy and thus predict the incoming sense data.
Information represented at an arbitrary level of the
hierarchy is passed to and incorporated by successive
lower levels such that the lowest level of the hierarchy
uses an integrated version of all the system’s knowledge to
predict sense-data. Upward (or backward) connections
allow the structure to propagate information about
prediction errors to higher levels of the hierarchy [see Fig.
1]. Error information from predictions is used to adjust the
hierarchy to better predict future stimuli. This learning
process, iterated continuously as the system makes
predictions and encounters prediction errors, allows for the
6 A structure that is able to predict the sense-data well, i.e. assign a high
likelihood to it, is able to “explain away” the sense-data. One must only
sample from this probability distribution to generate a “virtual version” of
the sensory data (Clark 2013).
7 Sometimes, the meanings of “forward” and “backward” are reversed in
relevant literature, such that “forward” refers to the propagation of
prediction error and “backward” refers to the formulation of predictions, as
in Rao & Ballard (1999).
evolution of relationships between layers. These
relationships, where a layer of the hierarchy acts as a set of
prior beliefs for the layers connected to it from below,
describe the knowledge—including background
information and relationships between concepts retained
from previous experience—the system uses to make a
particular prediction about sense data. The forward
connections between layers direct a flow of concepts used
to explain the observed stimulus (as well as concepts at
subsequently lower levels of the hierarchy), and backward
connections enable updates to the structure and the
adjusting of the priors (i.e. concepts) relevant to
understanding the stimulus.
One biologically plausible, computationally tractable
8
suggestion for the learning mechanism uses Bayesian
principles. Under the Bayesian framework, layers of the
hierarchy act as statistical priors for successive layers
below it. Prediction involves the formation of a probability
distribution over the sample-space of events which might
occur in a particular stream of reality (i.e. umwelt)
experienced by the system. In Fig. 1, this distribution is
represented generally by P(X | HPS), where Xrefers to
sense data inputted to the system and HPS refers to the
hierarchical predictive structure (and its encoded
knowledge); the whole expression can be read as “the
probability distribution over Xconditioned on the HPS.”
Learning takes the form of Bayesian inference, which
requires calculating (or, more likely in a flesh-based
system, approximating) the posterior distribution, or
P(HPS | X). This distribution refers to the distribution over
possible hierarchical predictive structures given observed
sense-data X, and inference uses this quantity to update the
hierarchical predictive structure to best explain X(Clark
2016, 2013).
We are interested in how predictive processing and the
Bayesian framework can model a human observer’s
interaction with a work of art, namely the interpretation of
the artwork and the updated understanding of reality the
artwork inspires in the observer. The authors are well
aware that aesthetic theory is far flung from the areas of
cognitive science to which Bayesian models have been
applied. The Bayesian approaches from cognitive science
which most closely relate to our work here aim to
understand modes of cognition apparently distinct from
8 For a discussion on the biological plausibility of Bayesian principles see
Clark, 2016, p. 39-41.
5
cognition related to art. We use no experimental regime to
investigate our ideas here, and the type of cognition we
aim to address appears to be beyond the current
explanatory reach of cognitive science research. Rather,
we strive to introduce a vocabulary and mode of thinking,
synthesized from predictive processing and probabilistic
models of cognition, into aesthetic theories concerned with
the interpretation of texts, media, and art.
1.3 HBMs as instantiations of hierarchical predictive
structures
As described above, hierarchical predictive structures have
the basic requirements of prediction (through probabilistic
evaluation) and learning (through updates to the structure
according to mismatches between predictions and the
actual sense-data one encounters in the world). Bayesian
principles provide a framework and suite of algorithms to
reason about prediction and learning, or inference. What’s
more, recent research in cognitive science has studied a
class of Bayesian models, referred to as hierarchical
Bayesian models (HBMs), which learn and reason
9
probabilistically about reality by employing knowledge
encoded by the models’ structures. Though these models
from cognitive science do not assume the fundamental
predictive and anticipatory quality of the mind stipulated
in PEM, their foundations in Bayesian reasoning allow for
their importation to PEM à la Bayesian hypothesis (Clark
2013). If hierarchical predictive structures outline the
flows of information to anticipate the environment and
learn from incorrect predictions, Bayesian principles
provide a framework for contemplating their core
mechanisms, and HBMs can be considered a theoretical
instantiation of hierarchical predictive structures.
Generally hierarchical in organization, HBMs can define
probability distributions over structured symbolic forms
such as graphs. Structures of many different forms (e.g.
trees, clusters, spaces) can be represented as graphs, each
of which has been shown to effectively model certain types
9 Learning in complex hierarchical Bayesian models aims to distribute
information gained from observing evidence across the structure which aims
to predict, or provide a probability measure of, evidence. Learning from
evidence requires at the very least calculation of the posterior of the
probability distribution defined over the structure and often involves
message-passing algorithms. Previous work has shown the structure of
concepts governing relationships in a domain can be algorithmically learned
(Kemp, Perfors, and Tenenbaum 2004).
of cognition. Such graphical models are able to encode the
dependency structure of a system of random variables,
including rich networks of causal and correlative
relationships. For example, HBMs have been used to
represent reasoning for intuitive theories, or “a system of
related concepts, together with a set of causal laws,
structural constraints, or other explanatory principles, that
guide inductive inference in a particular domain” (Kemp
and Tenenbaum 2008); (Kemp and Tenenbaum 2008;
Kemp, Perfors, and Tenenbaum 2004). The mind is not
restricted to reasoning about one particular domain: it
consistently utilizes a multitude of intuitive theories, and
thus might invoke different networks of concepts (which in
turn can be modeled as HBMs) to reason about multiple
disparate sense-data streams.
Important to understanding art-encounters through the
PEM framework is how HBMs can represent concepts and
related principles which contribute to the explanation of
multiple phenomena. As an implementation of hierarchical
prediction structures, HBMs can effectively represent
background knowledge with prior distributions. Critically,
high-level priors can be used to represent knowledge, i.e.
relationships between concepts, that is relevant in various
contexts (Tenenbaum et al. 2011). Such priors can form
hypotheses about multiple distinct sequences of
sense-data. These hypotheses can be conceived of as
cascades of concepts in a rich network of relations that
explain observed phenomena.
HBMs are a way of we can express in formal terms the
rich processes of prediction and learning. We’ve seen how
they can be used to model intuitive reasoning—one might
imagine crafting an HBM which captures the core aspects
of engagement and reflection about an artwork. Below we
sketch a Bayesian model which encapsulates some of what
we consider to be fundamental dynamics in art encounters.
This model can be described in further detail as an HBM,
which we leave for future work. We stress the abilities that
the language and handles of Bayesian reasoning provide us
to ponder prediction and learning in the PEM framework.
We will elucidate further concepts of HBMs below
alongside our discussion of art encounters and PEM.
6
1.4 Predictive coding and the schema
We can understand schema informally in the way Piaget,
Kant, and others have used it: a mental framework which
is built and updated through processes of induction and
deduction. From psychology and hermeneutics come close
synonyms for schema/ta, including framework, worldview,
way of seeing, interpretive filter, and mental model
(Reason 2018f). Psychological and perceptual theories of
schemas function as a less-formalized cousin to the
cognitive-scientific concepts knowledge structure (in
Bayesian terms) and hierarchical generative model (in
PEM terms), which is to say the top-down probabilistic
models present in predictive and statistical systems.
10
We’ll use “schema” (or “schemas,” plural) here as a
handle, in the contexts of Bayesian cognitive models and
predictive error minimization respectively, to refer to
either knowledge structures (encoded by HBMs) or
hierarchical generative models (HGMs). This shared
terminology is meant to emphasize that though the
academic vocabularies of frameworks of PEM and
Bayesian cognitive models differ, in this paper’s theorized
system, HGMs and the Bayesian knowledge structure are
analogous; that is, following Hohwy (2013) and Clark
(2016), the theoretical conception of PEM used in this
paper speculates a Bayesian or Bayesian-approximate
predictive structure at all levels.
In other more speculative contexts throughout this paper,
“schema” is used as a means of conjecturing potential
alignments between the PEM and hermeneutic systems, or
as to ways artworks construct and exploit predictive
systems, without overstepping into over-speculative
mathematical formulations. We hope this stresses that the
observations of later sections are not contingent on the
specifically Bayesian attributes of cognition hypothesized
in §1.3 being the case (nor are the Bayesian attributes
contingent on the attributes of art encounters stipulated in
§3-5).
1.6 A brief note on art and predictive coding
Predictive coding has been applied previously to visual art,
such as in Van de Cruys and Wagemans (2011), Kesner
10 As an informal example of the ways a schematic worldview can be made
up of inferential probability distributions, consider the prototypical
conservative worldview, whose perception of high social precarity flows into
(and out of) all kinds of risk assessments and outcome probabilities.
(2014), and Kandel (2016). The focus, however, has
always been on visual processing, with higher-level
hermeneutic and conceptual processes largely ignored .
Kandel cites modern art, for instance, as “dismantling” the
literal visual perspective in order to force the brain to
“come up with a new logic of bottom-up processing”
(Kandel 2016). While such theories are congruous with our
own hypothesis, they constitute a mere subset of the types
of conceptual disruption this paper theorizes. Here, we
depart from these authors by looking at artworks from a
high-level, literary-theoretic lens (as opposed to a
lower-level, visual art perspective), surveying, first, how
hermeneutic processes rely on predictive processing, and
second, how works of art and literature actively set up and
then subvert high-level thematic, narrative, and formal
(e.g. stylistic) predictions in their audiences.
§2 Hermeneutics
Textual comprehension requires interpretation, a process
of sense-making requiring the extraction of a regularity or
logic from textual data. This is true at all levels of textual
meaning, from the ground on up: “The conceptual space
generated by reading a literary work is a ‘cluttered array’”
that “consists of colours, edges, forms, and textures that
are resolved into attended Gestalt objects” (Stockwell
2009). Elements are weighted by relevance and attention is
allocated according to the configuration of the text and the
schematic interests of the reader.
As argued by de Beaugrande in Introduction to Text
Linguistics, texts “make sense” in that they contain a
“continuity of senses” which are congenial to each other.
When texts do not make sense—are nonsensical—they
lack such a continuity. The property of “making sense” can
be called “coherence,” and the configuration which
7
underlies such coherence can be called a “textual world”
(De Beaugrande and Dressler 2016).
2.1 The artwork as puzzle
All works of art, including texts, are puzzles insofar as
they must be consciously or unconsciously “puzzled out”
in order to yield coherence (in Bourdieusian terms, they
must be “decoded” according to a soft grammar, which is
to say, “parsed”) (Bourdieu 1998). Texts are rarely
designed with a “correct answer,” though sometimes the
authorially intended meaning is presumed to fill this role.
Instead, it is the retrieval of any meaningful, non-arbitrary
reading whatsoever which requires decoding, be it of the
natural language in play (its set of connotations and
denotations), the intertextual history of the discipline, the
discursively determined values of the tradition within
which the work was made and released, etc. In most fiction
(and much of nonfiction) the knowledge required to
unconsciously puzzle out or “decode” is already possessed
by the majority of intended readers and can be applied
with little friction. (Indeed, the text is written specifically
with such audience-specific legibility in mind.) Such
readers are engaged in consistently but unconsciously
interpreting both the parts of the “puzzle” and their
interrelation, that is, how each part might fit together in a
(not “the”) coherent whole. New knowledge presented by
the work updates the prior conceptions, while prior
knowledge influences the understanding of new
conceptions. This process involves the recognition of—to
name but a few—sequential, logical, thematic,
mythological, psychological, and ideological patterns
within the work (Reason 2018a). We will consider such
encounters, in which interpretation is largely unconscious
and self-contained (that is, requiring no reference to
outside sources of information), as prototypical
hermeneutic encounters, whose structure is reformulated
into a predictive framework throughout §2.
2.2 Predictive formulation of a hermeneutic
encounter
In §2.1, we described the hermeneutic process as one in
which the observer (i.e. reader/viewer/audience member)
is constantly “engaged in consistently but unconsciously
interpreting both the parts of the ‘puzzle’ and their
interrelation, that is, how each part might fit together in a
(not ‘the’) coherent whole” (see Fig. 2). Here we will
formulate that process approximately in probabilistic
terms, representing a hermeneutic encounter graphically
through a hierarchy of probability distributions (Sharifian
1997).
Consider a set of events X= {x1, x2, …,xn}. Xis the set of
successive events in an art encounter, which may be more
narrative-based and determinately sequential (e.g. films
and literary texts) or more introspective and
indeterminately sequential (e.g. a static piece of visual art).
A reader’s interpretation of an artwork might refer to the
combination of events in set Xand a hierarchical
knowledge structure C(consisting of concepts and
relations between concepts) which are deployed to engage
with the set of events X. Cshould include all relevant
knowledge that exists beyond the artwork as well as
concepts, relations, and knowledge that apply only in the
world of the artwork. We can insert concepts Cand events
Xinto a graphical structure, where C and X are nodes in
the graph and the edges between them encode their
relationships. Using Bayes rule and principles from
graphical models, we can use the graphical structure to
conceive of a conditional probability distribution p(X | C)
over the events in Xgiven knowledge structure C (Koller,
Friedman, and Bach 2009). A “coherent whole” might
refer to a given structure spanning the sets of elements X
and Cwhich yields a high value for p(X | C), reflecting a
high likelihood of the art events given our understanding,
or graphical configuration, of related concepts. While
engaging with a work over time (e.g. reading sequential
chapters of a book, watching a film, interpreting elements
of a painting one after another), we constantly make
predictions about specific scenes and revise the possible
outcomes for future remaining events contained in X.
When interpreting a specific event of a work, xt, only
certain parts of the knowledge structure Cwill be
activated: call this subset C’, where C’
⊂
C. We also take
into account events previously experienced in the art
encounter; for example, we can use the events X’ = {x1, x2,
…, xt-1} to assess the possible outcomes for the next event
Xt, or for the remaining events in the artwork after Xt,
written X\ (Xt∪X’). Now, we can denote our predictions
about event Xtas p( Xt| X’, C’). So as we progress through
an artwork, at each timestep we expand, and contract, and
phase-shift the activated parts of our knowledge structure
8
C to understand the next event in the piece, the remaining
events of the piece, and the piece as a whole.
11
In addition to making top-down predictions about art
events, we also learn bottom-up from art events and update
our larger schemas based on information contributed by
the art events: we use the posterior distributions over
concepts given events from the art encounter, e.g. P(C | X)
or P(C’| X’), to revise the respective distributions over the
art events themselves, thereby revising our interpretation
of the artwork as a whole —a process markedly similar to
12
that of the “hermeneutic circle” gestured at in Fig. 2. As in
the hermeneutic circle, this process happens piecewise
during an encounter with work, where we calculate
posteriors over subsets of concepts from subsets of the
data p(C’ | X’); a scene tells us about a few but not all of
concepts relevant to a film. Learning algorithms from
Bayesian graphical models, such as message passing,
suggest how new observations can propagate throughout
and update the structure of concepts. When our knowledge
structures of the larger world, or the larger art tradition
(and not merely the world of the work at hand), are
updated in a significant way, schematic subversion has
taken place. We expand expand on schematic subversion
throughout §3.
By way of example, we can look at a few of the primary
concepts, C, in play when watching a Spaghetti Wester:
formal genre conventions that intersect with conventions
of the era, based on other Spaghetti Westerns, Westerns,
and previous films watched by the viewer; historical
knowledge about the American West; general knowledge
of human behavior put on display; and the physical and
operational possibility or “realism”—what we might call
the “soft physics” of the real and textual worlds. Crucially,
these are merely a handful of relevant informing concepts.
Priors will be based also on all action films watched by the
viewer, all Italian films watched by the viewer, all films of
any type watched by the viewer, all artworks consumed by
the viewer, etc, where there is decreasing relevance or
radial proximity of the concept to the work at hand.
Similarly, the viewer’s knowledge of human psychology
11 Each art form and tradition has its own patterns of entropic growth and
diminishment—rhyming poetry, for instance, narrows its realm of
possibility at the end of each line.
12 P(X|C), the probability of X given C, can be understood as deduction, the
process of understanding a particular instance through reference to a general
principle. P(C|X), the probability of C given X, can be understood as
induction, the process of understanding a general principle through
reference to a particular instance.
will come into play with respect to the analogic similarity
of the filmic scenario to learned or experienced
knowledge. Relevant concepts grow—but are also
constrained, and altered—as the film carries out its
runtime.
2.3 Interpretation and intentionality
Insofar as an observer sees a work of art as a form of
communication, attentional prioritization will take place on
somewhat the same grounds as, or with some degree of
symmetry to, what the director intends the viewer to
prioritize (since the essence of communication is the
ongoing attempt to understand what the interlocutor means
to convey through an utterance, and not merely what is,
technically, said). Audiences search for markers of
hierarchical importance in a communication, while artists
employ intuitive or else culturally established motifs of
emphasis to signal intentionality to observers. Insofar as
the artist knows element mof the artwork is a marker, the
audience knows mis a marker, and the director knows the
audience knows mis a marker (these being the conditions
for mutual knowledge), communication is possible.
Foregrounding an object, action, or utterance is one default
mode in film and visual art to communicate priority, e.g.
through volume, prominence, tonal stress differentiation,
size, position, or lens focus. Not only are foregrounded
events more likely to be naturally noticed, they also direct
attention and inform high-level conceptions of
communicative priority.
At the sentence level, commas and syntax are two
examples of textual markers which enable greater
hermeneutic accuracy by signaling intentionality—that is,
the specific way the words are intended to be interpreted,
or the “continuity of senses” postulated by de Beaugrande.
When the meaning of a sentence remains unclear,
13
perhaps due to ambiguous reference or grammatical
structure, the interpreter must formulate a guess based on
one, dynamics and structure from the real world, and two,
assessments of the speaker and his motives. Predictive
systems of inference allow us to accurately gauge the
intent in the ambiguous sentence She announced a
program to promote safety in trucks and vans (an example
taken from natural language processing). Is the
announcement made “in [read: inside] trucks and vans”?
13 Commas and syntax do not “mean” anything; they merely signal and
modulate relationships.
9
Did the speaker announce the program in order to promote
automobile safety, or is it the program which promotes
automobile safety?
In garden-path sentences, such as “The old man the boats,”
the initial assumptions of a first parse— that “old” is an
adjective modifying the noun “man”—must be updated to
identify “old” as a plural noun and “man” as in its verb
form. In the initial parse, the high probabilistic likelihood
of “the old man” signifying an elderly male makes that
initial interpretation so strong that the subsequent “the
boats” cannot immediately amend it, and a second or third
reading is required to verify that the dramatically unlikely
minority sense-meaning is, in fact, the one intended.
Finally, at the level of individual word senses, we can
theorize a subject who is aware of an approximate
probability distribution consisting of the discrete
likelihood of each individual sense-meaning—connotative,
denotative, rhetorical, and figurative—of a given utterance
within its context. Moreover, this probability distribution
is known, varyingly and approximately, among the larger
population of native speakers. Finally, the distribution,
similar to textual markers of priority, is mutually
known—that is, all members of the group are aware that
other members of the group will have similar hypotheses
about utterances, this being the step that enables
predictable communication, and thus allows us to
communicate at all. Language, as we’ve analyzed it so far
in §2.3, can be understood loosely, alluding not just to
natural languages but also the languages of cinema, visual
symbolism, musical moods or motifs, and so on.
At the phrasal level, predictions are made, for example, as
to whether an act of indirect speech (e.g. the veiled threat
It would be a shame if something happened to…) is meant
literally or more suggestively (Pinker, Nowak, and Lee
2008), and both interlocutors in the exchange must share
similar-enough models of the phrase’s rhetorical
probabilistic likelihood within the threatened party’s
linguistic schemas to communicate. Poetry, as we will see,
is engaged, among other things, is undermining and
“riffing on” the dominant and thus expected
sense-meanings of its language, requiring the poet to have
a well-tuned linguistic “metaschema”—a predictive model
of readers’ predictive models. Similarly, speakers and
writers engaged in rhetorical or informative
communication must acquire intimate knowledge of how
their words and grammars will be parsed by their
respective audiences, allowing them to sidestep
ambiguities or miscues to get across what they “really”
mean to say.
We propose that, as in speech and writing, much of artistic
production occurs under similar conditions, whereby an
artist’s expressions, selections, and framings are guided by
the active consideration (i.e. prediction) of how audiences
may interpret or anticipate a work from the grammatical to
the conceptual levels. However, as we’ll see, modernistic
and avant works appear to be marked by their specific
tendency to undermine or disrupt this anticipatory
structure, and accordingly the knowledge structure that
informed it.
§3 Art, Prediction, and Subversion
In §2, we considered how a hermeneutic encounter with an
artwork could be framed in terms of prediction by a
hierarchical probabilistic model. Specifically, we looked at
the ways that unconsciously “puzzling out” (i.e.
interpreting) a work gives rise to an explanation or
hypothesis that gives its parts a coherence of meaning. In
much of avant-garde writing, however—as well as in
specific literary traditions like modern and postmodern
poetry—elements like gappiness, homophonic slip, or
grammatical reversal require the reader to actively and
consciously puzzle out a work’s meaning through repeated
readings of the text, or through reference to external
sources (e.g. a dictionary, encyclopedia, or fellow literary
work). Though it’s outside the scope of this paper to define
what art “is,” we can still observe that there is a strong
correlation between how “artsy” (how “sophisticated,”
“literary,” or “avant-garde”) a work is perceived as within
the cultural field, and the extent to which it actively resists,
confounds, subverts, or problematizes the interpretive
attempts and procedures of its audiences (Reason 2018b;
Noë 2015). That is, the artsiness of art is a property
defined by, or at least strongly correlated with, the work's
confounding of, and resistance to, easy assimilation into
observers’ schemas. Tellingly, art which does not
14
14 Consider that the artistic traditions which gets described in terms of
formal advancement or dialectic—that is, the pushing of material, tonal,
formal and categorical boundaries in experimental and avant-garde fields
(Gordon and Poggioli 1968; Bensman and Gerver 1958; Crane 2015;
Clignet and Crane 1988)—can also be described in terms of the continuous
subverting expectations of observer assumptions, roughly equivalent to
10
complicate or subvert audience expectations is perceived
as formulaic, predictable, boring, or trite, and is accorded
less prestige as a result (Martorella and Crane 1988, 12);
(Bourdieu and Johnson 1993).
3.1 Schematic subversion
That is the danger with Kafka. Just when you think
you know him he makes a sharp turn and you end
up facing a wall.
—James Nulick, Valencia
Stockwell in Cognitive Poetics: An Introduction
summarizes “schema poetics,” a subdiscipline of cognitive
poetics concerned with readerly context and interpretive
lenses (within our framework, a reader’s probabilistic
assessments of the world). Stockwell divides discursive
modes into those which are schema-preserving, -adding,
-reinforcing, and -disrupting. Everyday discourse is
typically schema-preserving or schema-adding, in that it
tends to conserve and apply its interlocutors’ worldview.
Literary and artistic discourse, meanwhile, tends toward
the disruptive (Stockwell 2005).
In analyzing the information content of discourses (from
the everyday to the literary), Stockwell cites Robert-Alaine
de Beaugrande, who in 1980’s Text, Discourse, and
Process carves categories of discursive information into
first, second, and third orders. First-order informativity
entails low novelty and low surprise, and is therefore
schema-preserving. Second-order informativity presents
more “unusual” data and helps “develop schematic
knowledge by accretion,” adding nuance or detail but
stopping short of significantly altering high-level priors.
Third-order informativity, finally, involves “highly
unlikely” or unusual data with respect to the reader’s
incumbent schema, and can represent such a disruption as
to lead to its thorough restructuring.
For example, we might have a predictive structure partway
through a novel for “what kind of person” a literary
character is. This schema can be represented by a
subverting the work’s operational assumptions. Here, the avant-garde’s
trajectory (or “progress”) can be described not as the primary goal ofavant
producer but as a byproduct of a continuing quest to disrupt the
expectations of their audience. There is some precedent for this: Dada has
been described by Duchamp, for instance, as the “denial” of established
interpretive approaches (Duchamp 1956).
hierarchical model, resembling the one presented in §2,
which encodes in its structure the knowledge useful to
understanding the character and predicting their future
actions and outcomes. Said schema will require non-trivial
revision should we come upon surprising information
about that character’s actions or motivations later on in our
experience of the text (de Beaugrande 1982). Actions
disjunctive with our gestalt understanding, or new
information which alters our understanding of a character’s
motivations, may reshape, retroactively, our understanding
of the character’s previous actions.
We’ll refer to forced hermeneutic revision as a type of
schematic subversion (schema-disrupting, to extend
Stockwell’s terminology)—the process by which texts and
artworks methodically set up audience assumptions
through coherence, as outlined in §2 and §2.2, only to
systematically undermine many of those assumptions later
in the work. A variant of this strategy we term
opportunistic hermeneutic revision, in that it capitalizes on
implicit audience assumptions about how a category of
works “work” (e.g. what is expected of genre fiction, or of
music, or of poetry) that have been culturally built up and
are thus capable of being subverted forthright. A common
effect of poetry, for instance, is the subversion of dominant
sense-meanings of a word or phrase in favor of more
minority interpretations (see §2.3 Interpretation and
intentionality). From (Pierre Bourdieu, Bourdieu 1985):
“‘Pure’ poetry appears as the conscious and methodical
application of a system of explicit principles which were at
work, though only in a diffuse manner, in earlier writings.
Its most specific effects, for example, derive from games
of suspense and surprise, from the consecrated betrayal of
expectations, and from the gratifying frustration provoked
by archaism, preciosity, lexicological or syntactic
dissonances, the destruction of stereotyped sounds or
meaning sequences, ready-made formulae, idées reçues,
and commonplaces.”
In general, schematic subversion involves the presentation
by the work of third-order informativity, that is,
information which has high novelty and surprise given
audience schemas. Expectations which are forcibly or
15
opportunistically subverted include understandings based
on genre, tradition, author, cover, synopses, reviews, social
15 Schematic subversion can also be understood in the Murray Davis sense of
interestingness, where interesting information is that which updates the
assumption-ground (set of assumptions) of its audience (Davis 1971).
11
information, etc., and stem from a combination of previous
experiences and learned information, both within the space
of the work and outside it.
In other words, effective artists are keenly aware of our
predictive structures, and frequently leverage such
understandings to not just maintain attention or
communicate with clarity and economy, but also to create
suspense and upend conceptions. One mode, associated
with “high” or “avant” art, tends toward subverting our
generic, formal, and textual expectations of a work; in
certain subfields of visual arts, ideological subversion is a
primary aim (Reason 2018d). Within “low” or popular art,
these subversions are more often narrative-based than
structural, taking the form of plot twists and comedic
relief. Such works cause us to fall into predictive traps,
into understandings or anticipations of the artwork or
world of the artwork which are shown to be misguided.
Within the framework of cognitive poetics, we can say that
art is schema-disrupting, and much of its value to audience
members derives from its third-order informativity.
3.2 The artwork as a joke
The conception of artworks as schema-disruptive
presented above bears commonalities with the theory of
humor presented by Hurley, Adams, and Dennett in Inside
Jokes: Using Humor to Reverse-engineer the Mind (2011).
The essence of a joke, in the view of Hurley et al., is that
its teller surreptitiously introduces a certain epistemic
commitment, then reveals it to have been mistaken. When
we experience humor, we are led down a “garden path” of
a covertly introduced, mistaken assumption which is then
revealed to us via the punchline (Hurley, Dennett, and
Adams 2011, xi–6).
Hurley et al. use the computer science metaphor of
debugging, and make an evolutionary argument as to its
origins. “Mother Nature—natural selection—has
[stumbled upon humor as an incentive for] our brains to do
all the tedious debugging that they must do if they are to
live dangerously with the unruly pile of discoveries and
mistakes that we generate in our incessant heuristic
search” (Hurley, Dennett, and Adams 2011, xi).
The philosopher Henri Bergson presents a similar analysis
of the humor inherent in tripping and other physical or
social awkwardnesses. Bergson argues:
…when people are too trapped in the automaticity
of their mechanical movements and when these are
insufficient in dealing with the environment at
hand, a comical situation presents itself. Bound by
the habits of movement, people sometimes forget
to adjust for new terrain or unexpected obstacles,
or they get so accustomed to their standard
environment, they expect the body to do all the
work intuitively. (Wampole 2015)
When one's predictions fail in an art encounter, it is much
like Bergson's jolt: out of the automaticity of interpretation
(as well as the schema responsible for the misinterpretation
in the first place) and into a state of cognitive arousal [see
§4.1, Art as superstimulus, & §4.3 Surprisal as
situationally valenced].
3.3 Charged subjects and formal gestures
It is our belief that many artists, rather than making a work
which is “realistic” in itself, choose “representative”
scenes, characters, and images to be the building blocks of
their works. These elements are somehow archetypal or
demonstrative of an underlying reality, sometimes physical
but more often social, psychological, or cultural. In this,
we have the support of reader-response theorist Wolfgang
Iser: “No literary text relates to contingent reality, but to
models or concepts of reality, in which contingencies and
complexities are reduced to a meaningful structure”
(Duckworth 1979).
Readers and viewers therefore encounter “charged”
scenarios—data points which, because they represent, to
the artist, rich, complex phenomena in the real world, tend
to be highly meaningful, intended to be contemplated at
some length after the experiential fact (and which often
naturally are, given their difficulty, complexity, and
interestingness). From a humanistic perspective, the
subversion of these scenarios is extra meaningful; the
subversion is not “arbitrary” but tethered to a world and
acts on the array of related schemas which might
generalize to other works and domains. In §4.4, we’ll
further consider how these “charged” or representative
data points might be perceived as higher-importance,
highly weighted inputs to a mind’s probabilistic model of
reality.
12
In humanistic, cartographic art, schematic subversion is
typically done of a charged subject, one which relates to
the real world and has a significant bearing on its
audience’s schemas, e.g. their understandings of social
psychology. In more formally oriented works, the
subverted subject might relate to specifically linguistic or
spatial expectations, such as the probabilistic
sense-meanings of a word (§2.3) or else the unspoken rules
and boundaries of a discipline. Artists like Duchamp and
Cage epitomize the latter approach, while pure poetry is
chock-full with the former. Both constitute foundational
priorities of modernist and twentieth century art practices
more generally.
16
From James Schuyler’s “The Morning of the Poem”:
(1) Force, fate, will, and you being you: a
painter, you drink
Your Ovaltine and climb to the city roof, “to
find a view,” and
(2) Wings in fierce blue delphinium depths I think
About those two blue jays like me, too
chubby, and Baudelaire’s skull
In the first excerpt, “drink” leads to predictions of alcohol
that are revised (or doubled) by “Ovaltine.” In the second,
“too” is initially likely to be understood as meaning “as
well,” an assumption which is revised (or doubled) by
“chubby” to meaning “overly.”
From Bernadette Mayer’s “Synesthetes at the Writers
House”:
I'm pleased to announce
that staying at the Writers House
is like living under a multi-colored apple tree
in winter;
Here, the divergence of figurative connotations between a
Writers House “like living under a multi-colored apple
tree” (with its suggestions of bounty, fecundity, beauty,
16 Again following Bourdieu, we can speculate that this is due in part to the
difference in audiences between avant and non-avant work. In the 20th
century, the visual art field especially but also more largely the fields of
theater and “high” literature became increasingly “autonomous”:
self-sufficient and “insider,” with a large percentage of its consumers
themselves being invested players (artists, curators, collectors) in the field, as
opposed to a more disinterested observer. As fields gain autonomy from the
larger society and market, they increasingly devote attention to the specific
methods, techniques, and approaches of the field (since these topics will be
of greater interest to “insiders” as opposed to “outsiders.”)
and shade) is subverted in the next line by the suggestions
of that same tree in winter: barren, providing no shade,
ugly in its nakedness. Mayer’s writing, like so many poets
and artists, has been praised for its subversion and
manipulation of reader interpretation: “[She] postpones
interpretation, perhaps forever, in an attempt to chip away
at both her and her reader’s compulsion to know where the
writing is going [next],” writes Maggie Nelson in Women,
the New York School, and Other True Abstractions
(Nelson 2007).
Lastly, it is worth drawing attention to the concept of the
“turn” or volta, in theories of poetry. This is the moment in
the poem in which its tone or rhetoric pivots noticeably
into another register; it is often seen as the heart of the
turning work, the revelation which reveals its center. In
17
other words, what comes after, or with, the volta
retroactively modifies our understanding of what has come
before, and often goes so far as to directly subvert the
understanding which the poem has been building up
theretofore.
3.4 Hermeneutic revision and the effect idea
Schematic subversion is an effective way to not just adjust
but also bare or expose expectations (e.g. bias,
preconception, worldview). Effect ideas are artistic
mechanisms of action—by a work, onto an audience
member—in which the baring of the assumptive ideology
(i.e. schema) conveys with it valuable information, or else
poses valuable questions, about the world. A “form of
philosophy,” effect ideas exert themselves through the
reader watching himself watch the text, predicated on
cognitive-predictive self-awareness throughout the art
encounter. Through this self-watching, the reader comes to
understand more about not just art, reality, and the world
but about the schematic self—the set of probabilistic
expectations brought to the encounter—through said
schema’s enactment on the work (Reason 2017a).
Frequently, though not always, the effect idea hinges on a
reader’s confused, ambiguated, or otherwise subverted
reading of a charged subject.
In this way, the artist catches us in our prejudices and
assumptions, and suddenly bares them—which is to say he
flashes us with our usually invisible ideologies, suddenly
17 “We could say that for the sonnet, the volta is the seat of its soul.” (Levin
2001)
13
and with recognition. Like Bergson’s “jolt” describing a
tripping pedestrian (§3.2), the automaticity of
interpretation has been disrupted, forcing an evaluation of
what went wrong. The attentive audience member gains an
awareness of his own, revealed schema, thus we can
creatively classify the effect idea, following Stockwell’s
system of naming, as schema-baring.
18
Though the effect idea plays an important role in all
mediums, the visual fine arts and literary fiction especially
encourage this category of response. Their consumption
involves prolonged pondering and self-evaluation when
faced with the art object. Moreover, literary and gallery
audiences have been trained to treat these mediums this
way (that is, self-reflectively), and in response, the
mediums’ works are created with such a treatment in mind.
Grietzer 2014: “Art is more ‘artsy’ the more [the
hermeneutic] process bolsters its intended impact.”
Tellingly, conceptual art hinges on the effect idea, using
the subjected experience as a way to
communicate—allowing for an observer experience that is
less pre-determined or ‘railroaded’ compared to alternative
models.
Still, effect ideas crop up in unusual places, including the
oeuvres of artists perceived as “middlebrow” or lacking in
cultural prestige. Spielberg employs a predictive lure in the
openings of many of his films, in which the audience is
cued into misapprehending the opening shot. In Close
Encounters of the Third Kind, Spielberg leverages the
audience’s trailer-cued expectations (that the film’s subject
will be extraterrestrials) by showing two bright
beam-lights on a dark screen. What appears at first glance
to be a spacecraft is revealed seconds later to be a jeep in a
sandstorm. In Jurassic Park, Spielberg uses audio
production, rustling leaves, and ambiguous shot framing in
order to convince its audiences that they are looking at a
dinosaur—only to reveal that the object in question is a
large transport truck. In both cases, what is initially
presumed to be a mysterious or monstrous “other” is
revealed in fact to be an instrument of man, an effect
which carries with it an implicit set of ideas.
18 For clarity, Stockwell’s original carving consisted of schema-preserving,
schema-accreting, and schema-disrupting modes of discourse. To this trio
we have additionally proposed the “schema-subverting” (§3.1) and
“schema-baring” modes (§3.4).
§4 Characteristics of schema-subversive
art
In §3, we theorized the schema-subversive (both forced
and opportunistic) and schema-baring functions of art
objects within the predictive framework introduced in §1
and 2. Here, in §4 and 5, we speculate on other possible
dynamics between art and an audience’s schemas.
4.1 Art as superstimulus
Following Hurley et al. (2011) on humor , we propose that
art is, among other functions, a kind of higher-level
cognitive superstimulus culturally evolved to target
humans’ innate predictive structure. Agent arousal
correlates with the properties of high perceived relevance
or precision as a Bayesian input, derived, respectively,
from the work’s (perceived) topicality and the author’s
(perceived) credibility. Like a joke, which is tailored to
guide listeners to a specific interpretation of events only to
pull the rug out (Hurley, Dennett, and Adams 2011), art is
tailored to target existing compressions in a subject’s
schemas. Where classical art often reifies or activates
familiar patterns, e.g. patterns used for object recognition
(Evans 2019), contemporary works often exploit and
subvert regularity observable in the real word. In either
case, the artwork provides an intense encounter between a
subject and schema, on one side, and the art object with its
highly compression-prone or compression-breaking
information on the other.
19
4.2 Truthiness and adherence to model
In an art encounter, our mind updates its inferential models
about the world with respect to the work’s perceived
accuracy, an assessment made by the model itself. The
observer’s schema acts as a “check” or arbitrator on its
own incorporation of the artworks’ worldview (dynamics
and concepts); when the cartography of the work is too
20
implausible in the eyes of an apprehending schema, it may
be dismissed entirely. This is to say that concepts learned
directly from personal experience, and indirectly from
19 The aesthetic theories of John Dewey, especially 1934’s Art as Experience,
is a guiding force here, stipulating as it does that art exists not as an inherent
property of a material, but as a dynamic exchange or interaction between an
activating viewer and activated object.
20 More accurate within our PEM framework, we can say that the work’s
accordance with the viewer’s schema plays a key part in guiding the
inferences of, or updates to, the HGM, based on the art encounter.
14
outside sources, are used to assess the likelihood of a
work’s worldview as conveyed through its components.
21
The artwork’s “truthiness” as estimated by the viewer can
be understood as its precision, or reliability (Clark 2016).
In this way the schema can be understood as a gatekeeper
to its own revision: only stimulus surpassing some level of
intelligibility and precision for a viewer will be able to
interact with the viewer’s schema and incite revision. One
consequence is that information which fits closely with an
existing schema but poorly with a ground-truth reality is
perceived by that schema as more, rather than less, likely
to be the case.
Art that presents worldviews or models of reality that are
congenial with the observer’s can be termed resonant; art
that is presented by a source who we deem authoritative is
termed credible.
4.2.1 Resonance
Resonance is a term used primarily by non-scholarly
readers (Stockwell 2009), and the vagueness of the term as
used in non-scholarly discourse fields has led to its being
avoided in academic discourse.
As we use it here, to resonate is to oscillate in such a way
that (figuratively) “a sympathetic oscillation occurs in a
similar nearby structure,” (Reason 2018e) and for most
readers, engaged in a highly personal encounter with a
text, the original oscillating structure is the text, and the
sympathetic body, in whom resonant vibrations occur, is
the reader’s body of personal experience or
worldview—their schema of the world.
While resonance is one of our key checks in art-based
learning, it is highly susceptible to confirmation bias.
Radical updates moving toward truth may be dismissed for
their implausibility according to the observer’s standing
world model, while minor updates moving toward
objective falsehood might be deemed more plausible
within the standing schema. Moreover, the kinds of
transformative texts that we look to in literature often work
off mental models of the world so distant from our own
that they ring alarm bells, but it is precisely these works
which are so valuable to learning (Reason 2018e). This
21 Following memory research, we can think of concepts derived from
personal experience as episodic, and concepts learned second-hand as
semantic. (Tulving, Donaldson 1972).
situation is compatible with the understanding of psychosis
through “pathological priors,” beliefs which while
damaging or patently untrue, nonetheless are sought out by
a dissonance-minimizing system to reduce entropic
uncertainty (Carhart-Harris and Friston 2019).
4.2.2 Credibility
We appear to treat artistic, literary, and cinematic works
simultaneously as communications, where an
interlocutor’s intent constitutes valuable information, and
as maps or models of reality (the “cartographic function”),
which is perhaps why concepts like authenticity and
sincerity are among the most contested and consecrated in
literary-artistic production. Insofar as we see artworks as
reality-updating inputs to the brain, our trust in and
opinion of the source of information will be highly valued.
In rhetoric, “credibility” is important as a marker of the
accuracy of presented information, or the efficacy of an
action proposal.
One significant innovation in fiction was the pointed use
of an unreliable narrator. Narrators are considered
“unreliable” when their reports significantly diverge from
what we, the readers, are led to perceive as the “reality” of
the textual world. We can understand this, similar to how
pseudonyms disrupt readers’ hyperpriors of author
identity, as a means of further destabilizing the interpretive
process (see §5.3 Hyperpriors).
4.3 Surprisal as situationally valenced
The disruption of predictive response, i.e. surprisal, has
been theorized to cause arousal in the subject due to its
signaling “important changes in the environment” which
require acting on (Mandler 2003). This arousal may be
neutral in that the situational context of its occurrence
affects whether it manifests for the subject as a positive
emotion (e.g. interest) or a negative one (e.g. stress); in
other words, the arousal is situationally valenced (Van de
Cruys and Wagemans 2011).
In an art encounter, the situational context contains both
the generic set of expectations specific to the artwork as
well as an implicit understanding that no action need be
taken in response to the environment of the work. On some
level, even as much as one “mistakenly” views an artwork
(especially a fictional narrative or world) as real-world
15
data to populate one’s models, the artwork is still cordoned
off as a sense-datum of low-risk or minimal consequence
(and which, opportunely, might reflect properties or
phenomena of reality). We propose that in an art
encounter, the observer is immersed in events and
predictive work within a controlled, low-risk “sandbox,”
causing predictive updating (i.e. inference work) to be
low-anxiety, even pleasurable, rather than stressful
(Reason 2018d). This low-risk quality is reinforced by
22
the fact that predictive events and failures in art
experiences are typically either asocial (updates happening
in one’s private thoughts) or shared and automatic (the
collective gasp of an audience at a horror twist)—thus they
further lack the social cost of having one’s poor calibration
publicly revealed.
An alternative theory of art arousal, presented by Van de
Cruys and Wagemans (2011) and taken up by Kesner
(2014), argues that the prediction-defying nature of a work
creates an initial displeasurable dissonance that can be
resolved—pleasurably—when the work’s deeper structure
(or grammar, or logic) is eventually discovered (Van de
Cruys and Wagemans 2011); (Kesner 2014). Phrasing this
within our PEM framework, we might say that an observer
enters an art encounter with a set of activated interpretive
concepts, C, judged appropriate to the artwork. When
those concepts (and therefore the larger schema) are defied
(as they inevitably are, in non-superfluous, meaningful
works), the observer must propose different hypotheses by
which to understand the work, and the eventual discovery
of an appropriate schema produces subjective pleasure.
This alternative hypothesis finds support in Schmidhuber
and Friston, who both ascribe pleasurability to discovering
efficient, instrumentally accurate explanations (or
“compressions”) of previously unpredictable stimuli (K.
Friston, Kilner, and Harrison 2006; Schmidhuber 2010;
Schwarz 2013).
4.4 Art as “weighted” or high gain
Art encounters, while inevitably “noisy” at the cognitive
level (one is distracted, the mind drifts…), are a culturally
22 Speculatively speaking, it would be surprising, given apparent bodily
physiological changes in response to low-stakes vs. high-stakes situations
(anxious vs. relaxed states), if there were not a corresponding difference in
the feeling of arousal which results from prediction errors while in each of
these states. We can imagine that predictions about the relative danger of a
prediction error would even be baked into the prediction structure itself.
special practice characterized by low-distraction
environments (gallery white walls, dimmed theaters, a
spotlight) in which the subject's attention is focused
primarily on the stimulus at hand. Attention, as Clark
advances, increases the gain (i.e. weighting) on prediction
errors, increasing in turn the learning potential of an
encounter, that is, the significance of its effect on a
subject’s schema (Clark 2016). After the fact, art
encounters are frequently contemplated at length, often
socially or textually (e.g. in critical reception) in discursive
scenarios not dissimilar to Hohwy’s model of
introspection as inference found in 2013’s The Predictive
Mind (Hohwy 2013).
Further, we can see art as a culturally special practice in
which not only are large amounts of attention focused, and
focused specifically on the inferential work of figuring out
and sense-making, but there is additionally a credibility (in
the sense of §4.2.2) to both the specific artist and the fact
of an encounter itself (artworks being culturally special)
which might assign high precision to error signals
propagated during the art encounter.
§5 Further applications in aesthetics
The predictive model of cognition also casts light on a
number of related terms from literary theory and
aesthetics. Here we’ll frame some of these concepts in a
predictive framework in order to illustrate PEM’s
explanatory power within established humanities
frameworks.
5.1 Accessibility
When we say a work (or category of works, e.g.
conceptual art) is inaccessible to the general public, what
we mean is that the ratio of familiarity to foreignness, of
predictable to unpredictable, is so low as to make the work
unassailable by the observer. There is no puzzling to do
because there is a dearth of priors from which to puzzle,
and the reader is left, in the words of (Csikszentmihalyi
and Robinson 1990), “on the outside, unable to interact
with the work.” In an environment dominated by noise, the
observer is left without a starting place, without any priors
or constraints on whose basis he can expect certain
characteristics over others, or form any predictions at all
about the work. This phenomenon might reflect an
16
inability of the viewer’s hierarchical predictive models to
provide meaningful hypotheses about the artwork.
5.2 Ideal vs. implied readers
The “ideal reader” as a literary-theoretic concept is often
used (implicitly) to mean a reader with an identical
significative and interpretive schema to the author, such
that communication between them is, so the thinking goes,
losslessly transmissited. But this ideal reader (as Wolfgang
Iser astutely points out) would find the author’s text
entirely “superfluous,” having nothing to learn from it.
Instead, Iser points to an implied reader, who “embodies
all those dispositions necessary for a literary work to
exercise its effect” (Iser 1978). In a schematic frame, we
can understand this implied reader—ideal in his own
way—as possessing beforehand the very priors and
assumptions which the author assumes in crafting a
situated sequences of meanings and subversions in his
writing the text (assumes both conceptually and also with
respect to the probabilistic “sense-meanings” in the
language, see §2.3 & glossary).
5.3 Hyperpriors
Hyperpriors in PEM systems are high-level priors
employed in an interpretive situation, such as an art
encounter. Pervading hyperpriors in our cognition can
include not just time and space—e.g., following (Clark
2013), “that there [can] only [be] one object… in one
place, at a given scale, at a given moment”—but identity
—following (Carhart-Harris and Friston 2019), our
“narrative of self.” Hyperpriors, in our probabilistic
formulation of art encounters, are high-level priors that
constrain the space of possible subnetworks of concepts
and their relations (Clark 2013; Swanson 2016). As such,
hyperpriors invariably influence the selection of a
dominant hypothesis for given sense-data, and all
hermeneutic work is guided and biased by them.
Hyperpriors tend towards being “fixed” as opposed to
“fungible”—for instance, hyperpriors are less susceptible
to revision, both generally and in an art encounter, than
hypotheses regarding the work’s meaning or content.
However, many artists do, in fact, attempt to misguide
audiences with faked or absent hyperpriors, such as by
masking their identities through pseudonyms, or
presenting their works in non-art contexts.
5.4 Defamiliarization
A well-known phenomenon in psychology is the cessation
of full awareness of familiar stimuli. The phenomenon has
been called “compiling” by Herbert A. Simon, “tacit
dimensionality” by philosopher Michael Polanyi, the
“ready-to-hand” by Martin Heidegger, and “driving on
autopilot” in casual parlance (Ekman 2013). In the terms
of the Russian Formalist school, it is the difference
between recognition and seeing, where to recognize is to
perceive in a minimal, peripheral way. Under the PEM
framework, we can understand this as schemas “explaining
away” well-integrated stimuli. Where high-novelty
information, or significant clashes between our
expectations and sense data, earns more awareness in this
model, low novelty or schema-congruent sensory data is
allocated less space in the consciousness field.
To the Formalists, a crucial aim of art was
defamiliarization, where the everyday and banal (or
“ready-to-hand,” or “pre-compiled”) is presented in a way
which distorts it into newness. Audiences appreciate a
fresh sight of what was previously merely recognized.
Often, the defamiliarized subject is not immediately
recognizable for what it is; only when the mind connects
the defamiliarly presented with the familiarly known, an
analogic link is created between the two which upcycles
into new models, or interpretations, of the familiar.
23
An adjacent concept to defamiliarization, taken from
psychology, is cognitive disfluency, which describes
effortful attention to a stimulus (in contrast to automatic or
effortless attention). Disfluent experiences, such as the art
encounter, have been found to “improve syllogistic
reasoning and reduce reliance on heuristics” (Hurley,
Dennett, and Adams 2011). Tellingly, both awkward
situations (see §3.2) and avant artistic encounters are
characterised by the disruption of automaticity in favor of
more disfluent hermeneutic or affective states.
For discussion of the design of intentionally fluent
experiences, see §5.8, “Design as the antithesis of art.”
23 Jonathan Richman, ArtNews: “[Dalí's] paintings helped me find my way
with their tone of foreboding and mockery but also a sense of wonder at the
universe itself and what Aldous Huxley once wrote about something else:
‘the sinister otherness of familiar things.’”
17
5.5 Ambiguity
Ambiguous states are those which lack a dominant
explanatory hypothesis. Often, multiple alternative
meanings are visible in what is known as polysemy; other
times, the allusion of the referent is unclear and no
hypotheses generated by an observer’s schemas earn a
high enough probability to be considered likely.
5.5.1 Temporal decay in probability mapping
For further context, see §2.3, Interpretation of
intentionality. Shared probability distributions of meaning
and reference allows artists and audiences to communicate.
Over time and across cultures, this symmetry between
interlocutors decays, leading to an ambiguity of reference
wherein no dominant meaning—or a different meaning
than before—presents itself. As a result, the work becomes
less accessible. The hook of Nina Simone’s “Mississippi
Goddam”—“And everybody knows about Mississippi,
goddam!”—once unambiguously (“everybody knows”)
referenced the state’s racial regressiveness in the
midcentury. The structural racism sense of the allusion
“everybody knows about Mississippi” might still retain a
plurality of likelihood if used today, but this sense is no
longer clear-cut.
5.5.2 Ambiguity of priority and grounding
When an artist is ambiguous or counter-intuitive in the
presentation of an object or occurrence such that it is
unclear where attention ought to be paid—e.g. by placing
signals in the background, or among noise—our predictive
models of perceptual triage are questioned. If the artwork’s
ambiguities accurately map real-world ambiguities, then
we are updating our predictive models with valuable
nuance (that there is not a single, dominant interpretation
but many). While we may not be able to make stronger
predictions of reality, our confidence in probability
distributions (a prior on the distributions) has been altered,
which will change the structure of future predictions
(Reason 2017b). The effects of ambiguous artworks, in
this sense, are not dissimilar to the effects of psychedelics
as outlined by Carthart-Harris & Friston in “REBUS & the
Anarchic Brain,” where REBUS refers to the relaxation of
priors under the influence of psychedelic compounds
(Carhart-Harris and Friston 2019).
5.6 Predictive coding and music
Music is the arts discipline which has been most
interfaced, to-date, with predictive models of cognition.
Various studies have linked the anticipation of melodic
patterns to lower-level auditory coding. Researchers have
noted that unfamiliar music is less dopaminergically
rewarding to listeners, whereas more familiar music, in
which listeners can actively anticipate coming notes and
chord shifts, is more rewarding (Salimpoor et al. 2011,
2015; Dura 2007); similarly, chance compositions have
been found to be less rewarding or pleasurable than those
which work within more predictable structures. These
observations can be applied in thinking about why most
popular music operates off a small bank of chord
progressions, why records get “worn in,” and why atonal
and achromatic music is less appealing to casual listeners
(Reason 2016). It may further help explain rhyme
schemes, chorus/verse structure, and the concept of
“difficult” versus “easy” listening.
24
In musicology, a relative of predictive processing models
has emerged in Eugene Narmour's implication-realization
model of melodic expectation (Gjerdingen and Narmour
1992; Narmour 1990, 1992).
5.7 Mass culture and mutual error minimization
Clark (2013) frames the phenomenon of media-enabled
communication (e.g. references to fictional situations and
characters, or description through analogy to mutually
known media) in terms of collective prediction error
minimization:
Using a variety of tricks, tools, notations,
practices, and media, we structure our physical and
social worlds so as to make them friendlier for
brains like ours. We color-code consumer
products, we drive on the right (or left), paint
white lines on roads, and post prices in
supermarkets. At multiple time-scales, and using a
wide variety of means (including words,
equations, graphs, other agents, pictures, and all
24“Difficult” records are understood to be those which take time and
patience to “wear in”; the internal structure or organizing logic proves
difficult to grok, or else so far from familiar structures that they cannot be
laterally understood. “Easy” records, a term often used pejoratively in
referring to pop music (see “easy listening”), are experienced as having an
“instant” or first impression charm.
18
the tools of modern consumer electronics) we thus
stack the dice so that we can more easily minimize
costly prediction errors in an endlessly
empowering cascade of contexts from shopping
and socializing, to astronomy, philosophy, and
logic (Clark 2013).
This mechanism as it applies to mass-consumed art sees
support in media studies thought (e.g. vis-a-vis
globalization), where ink has been spilled over the ways a
mass medium like film creates more homogeneous (and
thus, implicitly, predictable) cultural rituals, for example in
romantic norms. Thus the literary-theoretic concept of
texts as “models” (Iser 1993) takes on further meaning (as
not merely descriptive but also normative).
5.8 Design as the antithesis of art
Philosophers of art such as Alva Nöe (2015) have
previously defined art by its distinction from design,
similar to the ever-present contrasting of craft and art in
19th and 20th-century discourse. Insofar as art’s
subversion, confounding, resisting, and problematizing of
interpretation can be understood as purposeful illegibility
and surprise, the world of design is characterized by its
adherence to the principle of supreme, low-surprisal
legibility. Successful design appears to require insight into
audience (or “user”) schemas and procedures equal to that
of successful art. However, this insight is used to cue and
prime users of designed technologies and interfaces to
ensure smooth experience and intuitive navigability. Much
of our designed world has, by design, become thoroughly
familiarized, backgrounded into our lives and processed
unconsciously (i.e. unproblematically; see §5.4 on
“explaining away” sense data) (Reason 2018c; Noë 2015;
Reason 2018b).
As Nöe writes:
A designer of doorknobs makes a simple artifact,
but does so with an eye to its mesh with this larger
cognitive and anthropological framework. When
you walk up to a door, you don’t stop to inspect
the doorknob; you just turn it and go right through.
Doorknobs don’t puzzle us. They do not puzzle us
just to the degree that we are able to take
everything that they presuppose—the whole
background practice—for granted (Noë 2015).
5.8.1 Temporal decay in considerations of “art”
Moreover, as any one artwork fades into the horizon line
of history, it can cease to register as art because the
schema-subverting and schema-confounding effects which
once characterized it are temporally and culturally (i.e.
chronotypically, following (Bakhtin 2014) hyper-specific
and thus no longer in play for contemporary audiences.
Works which we grow up with can rarely surprise us, are
difficult to anew. Nöe again:
Very often we find ourselves admiring old
masters, for example, more or less solely for their
decorative aspects, or because of their supposed
historical significance or monetary value, or
perhaps because they exhibit virtuosity in
craftsmanship. And so of course it seems
implausible that we admire works of this sort
because of the way they subvert or undercut or
abrogate the authority of what is normally taken
for granted. After all, that’s just not what these
works do for us, at least most of the time. They
have expired. Or stopped being artworks (Noë
2015).
5.9 Future directions?
This paper cannot provide a full interfacing of literary
theoretic, art historical, and aesthetic frameworks with
theories of predictive error minimization. We’ll close only
by laying out a handful of concepts which jump readily to
mind as potentially congenial with such theories : critical
reagents, free indirect discourse, gappiness, genre,
grammar, indirect speech acts, interestingness (as defined
by Murray Davis), irony, the logical consistency of fantasy
worlds, narrator reliability, punning, red herrings,
representation, rhetoric, rhyme, suspense, unreliable
narrators, values hierarchies, and visual patterns.
19
GLOSSARY
art encounter: following John Dewey’s Art as Experience,
a dynamic exchange between a viewer and an art object.
Bayesian inference: a statistical inference method in which
Bayes’ theorem is used to update the probability of a
hypothesis of a model using prior beliefs and new evidence
as it becomes available.
bidirectional: referring to the way, in multilevel predictive
systems, bottom-up sense data and top-down schemas
intertact.
cartographic: referring to the way an artwork “maps” or
models reality, e.g. as prominently found in the social
realism of 19th century fiction.
compressability: regularity of a data set, either inherent or
with respect to some pattern identifiable or predictable by
a subject’s knowledge structure.
hermeneutics: the practice of interpreting texts, used
loosely in this paper to mean the interpretative work
necessary in any art encounter.
hierarchical Bayesian model: a class of Bayesian models
which learn and reason probabilistically about data by
employing knowledge encoded by its structure, i.e.
probability distributions connected in a hierarchical graph.
HBMs are an instance of predictive schemas.
hierarchical generative model/s (HGM): Clark’s carving
of the predictive schema, which may or may not utilize
Bayesian learning.
hyperpriors: a systemic prior, tending toward the abstract
and fundamental, in subjects' predictive models (see
Hohwy 2013, Carhart-Harris and Friston 2019, Clark
2013, Swanson 2016).
interestingness: a subject-dependent property of
information occuring when said information allows an
observer to predict or compress the world either more
efficiently, or more accurately (see Schmidhuber 2009,
Davis 1971). As Davis argues, schematically redundant or
random information is typically understood as
uninteresting, while interesting theories subvert, compress
(“help cohere”), or complicate and help revise one’s
models.
knowledge structure: the set of concepts or priors and their
relationships constituting an HGM, which guide its
predictions and probabilistic evaluations of the
environment.
noise: incoming information left out of models as too high
entropy (low compressability) or irrelevant (and thus
excluded from conscious attention).
precision: the reliability or certainty of information in an
inferential system, as estimated by that system.
prediction error: a measure of distance between the
predicted value (by a model) and the actual value. In
practice, the distance between a predictive agent’s
expectation of certain evidence according to their priors,
and the actual perceived evidence.
predictive coding: a framework in sensory domains like
visual processing that the brain actively anticipates, rather
than passively receives, incoming sensory information
PEM: prediction error minimization, a cognitive science
framework which relates prediction in agents (see
predictive coding) to the ideas of hierarchical generative
models and probabilistic inference
observer: a generic term referring to both the “viewer” (of
film or visual art theory) and “reader” (of literary theory)
allowing for a liberal discussion of artistic media more
generally.
schema: An agent’s knowledge structure operating on
various timescales and domains of information to make
predictions about reality, including art events; analogous
to literary-theoretic terms like worldview, framework,
mental model, and interpretive set, and also to PEM terms
like top-down models and rules.
sense-meaning: the intended signification of a text segment
(word, phrase, string), which is situated within both the
text’s “coherence of senses” and the larger cultural context
of the utterance.
20
surprisal: the degree to which data under a certain a
occurrence is unlikely given the subject’s models of reality
(see Clark 2017).
21
WORKS CITED
Bakhtin, Mikhail. 2014. “1981[1934-1935]. ‘Discourse in
the Novel’. In The Dialogic Imagination. Four
Essays, 259–422, Selected 261–265, 268–275.
Austin: University of Texas Press.” In The Discourse
Studies Reader, 28–35.
Beaugrande, Robert de. 1982. “Text, Discourse, and
Process: Toward a Multidisciplinary Science of
Texts.” Poetics Today 3 (1): 195.
Beni, Majid D. 2017. “Commentary: The Predictive
Processing Paradigm Has Roots in Kant.” Frontiers
in Systems Neuroscience.
Bensman, Joseph, and Israel Gerver. 1958. “Art and the
Mass Society.” Social Problems 6 (1): 4–10.
Bourdieu, Pierre. 1985. “The Market of Symbolic Goods.”
Poetics 14 (1-2): 13–44.
———. 1998. On Journalism and Television. Pluto Press.
Bourdieu, Pierre, and Randal Johnson. 1993. The Field of
Cultural Production: Essays on Art and Literature.
Bruner, Jerome S., and A. Leigh Minturn. 1955.
“Perceptual Identification and Perceptual
Organization.” The Journal of General Psychology 53
(1): 21–28.
Carhart-Harris, R. L., and K. J. Friston. 2019. “REBUS
and the Anarchic Brain: Toward a Unified Model of
the Brain Action of Psychedelics.” Pharmacological
Reviews. https://doi.org/10.1124/pr.118.017160.
Clark, Andy. 2013. “Whatever next? Predictive Brains,
Situated Agents, and the Future of Cognitive
Science.” The Behavioral and Brain Sciences 36 (3):
181–204.
———. 2016. Surfing Uncertainty: Prediction, Action,
and the Embodied Mind. Oxford University Press,
USA.
———. 2017. “A Nice Surprise? Predictive Processing
and the Active Pursuit of Novelty.” Phenomenology
and the Cognitive Sciences 17 (3): 521–34.
Clignet, Remi, and Diana Crane. 1988. “The
Transformation of the Avant-Garde: The New York
Art World, 1940-1985.” Social Forces; a Scientific
Medium of Social Study and Interpretation 67 (1):
276.
Crane, Diana. 2015. “Avant-Garde Art and Artists.” In
International Encyclopedia of the Social &
Behavioral Sciences, 320–25.
Csikszentmihalyi, Mihaly, and Rick Emery Robinson.
1990. The Art of Seeing: An Interpretation of the
Aesthetic Encounter. Getty Publications.
Davis, Murray S. 1971. “That’s Interesting!” Philosophy of
the Social Sciences 1 (2): 309–44.
De Beaugrande, Robert, and Wolfgang U. Dressler. 2016.
Introduction to Text Linguistics. Longman Linguistics
Library.
Duchamp, Marcel. 1956. Marcel Duchamp interview on
Art and Dada Interview by James Johnson Sweeney.
Video.
https://www.youtube.com/watch?v=DzwADsrOEJk.
Duckworth, Alistair M. 1979. “The Act of Reading: A
Theory of Aesthetic Response. Wolfgang Iser.”
Nineteenth-Century Fiction 34 (3): 337–43.
Dura, Marian T. 2007. “Sweet Anticipation: Music and the
Psychology of Expectation. By David Huron. MIT
Press, 55 Hayward St., Cambridge, MA 02142-1315;
800-405-1619; Http://mitpress.mit.edu. 2006. 512 Pp.
Black-and-White Illustrations, Music Examples,
Notes, Glossary, References, Index. Hardcover,
$40.00.” Music Educators Journal 94 (2): 54–55.
Ekman, Ulrik. 2013. Throughout: Art and Culture
Emerging with Ubiquitous Computing. MIT Press.
Evans, Owain. 2019. “Sensory Optimization: Neural
Networks as a Model for Understanding and Creating
Art,” November. https://arxiv.org/abs/1911.07068v1.
Friston, Karl J. 2013. “The Fantastic Organ.” Brain: A
Journal of Neurology 136 (4): 1328–32.
Friston, Karl, James Kilner, and Lee Harrison. 2006. “A
Free Energy Principle for the Brain.” Journal of
Physiology, Paris 100 (1-3): 70–87.
Gjerdingen, Robert O., and Eugene Narmour. 1992. “The
Analysis and Cognition of Basic Melodic Structures:
The Implication-Realization Model.” Notes 49 (2):
588.
Godfrey-Smith, Peter. 2016. Other Minds: The Octopus,
the Sea, and the Deep Origins of Consciousness.
Farrar, Straus, Giroux.
Gopnik, Alison. 1996. “The Scientist as Child.”
Philosophy of Science.
https://doi.org/10.1086/289970.
Gordon, Jan B., and Renato Poggioli. 1968. “The Theory
of the Avant-Garde.” The Journal of Aesthetics and
Art Criticism 27 (2): 244.
Grietzer, Peli. 2014. Amerikkkka. Vol. 004. GPDF
Editions. Gauss PDF.
———. 2017. “A Theory of Vibe.” Glass-Bead 1 (1).
https://www.glass-bead.org/article/a-theory-of-vibe/.
Griffiths, Thomas L., Charles Kemp, and Joshua B.
Tenenbaum. n.d. “Bayesian Models of Cognition.” In
The Cambridge Handbook of Computational
Psychology, 59–100.
Hill, Grahame. 2001. A Level Psychology Through
Diagrams.
Hohwy, Jakob. 2013. The Predictive Mind.