ArticlePDF Available

Relevance Realization and the Emerging Framework in Cognitive Science



We argue that an explanation of relevance realization is a pervasive problem within cognitive science, and that it is becoming the criterion of the cognitive in terms of which a new framework for doing cognitive science is emerging. We articulate that framework and then make use of it to provide the beginnings of a theory of relevance realization that incorporates many existing insights implicit within the contributing disciplines of cognitive science. We also introduce some theoretical and potentially technical innovations motivated by the articulation of those insights. Finally, we show how the explication of the framework and development of the theory help to clear up some important incompleteness and confusions within both Montague's work and Sperber and Wilson's theory of relevance.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 1 1–21
Relevance Realization and the Emerging
Framework in Cognitive Science
JOHN VERVAEKE, Cognitive Science Program, and Psychology Department
University of Toronto, University College Toronto, ON, Canada, M5S 3H7.
TIMOTHY P. LILLICRAP, Centre for Neuroscience Studies, Queen’s University,
Kingston, ON, Canada K7L 3N6.
BLAKE A. RICHARDS, Department of Pharmacology, University of Oxford,
Mansfield Road Oxford, OX1 3QT UK.
We argue that an explanation of relevance realization is a pervasive problem within cognitive science, and that it is becoming
the criterion of the cognitive in terms of which a new framework for doing cognitive science is emerging. We articulate that
framework and then make use of it to provide the beginnings of a theory of relevance realization that incorporates many
existing insights implicit within the contributing disciplines of cognitive science. We also introduce some theoretical and
potentially technical innovations motivated by the articulation of those insights. Finally, we show how the explication of
the framework and development of the theory help to clear up some important incompleteness and confusions within both
Montague’s work and Sperber and Wilson’s theory of relevance.
Keywords: relevance, constraints, self-organization, opponent processing, framework.
1 Introduction
There is a family of seemingly intractable problems in cognitive science. In each individual case,
it is unclear how it might be resolved, but the problems are central and so cannot be ignored or
marginalized. A well-known example is the problem of combinatorial explosion which faced the
general problem solving (GPS) framework of Newell and Simon [33]. We demonstrate here that this
problem, and others, have remained intractable because of a theoretical circularity caused by the
centrality of relevance to cognitive function. Attempts to deal with this circularity appear to have
been hampered by what we think is a confusion about what it is we can scientifically explain. We will
argue that one cannot have a theory of relevance itself because there is no stable, homogeneous class
of entities which correspond to the term ‘relevance’. However, we believe a theory of the mechanisms
of how relevance can be realized is tractable. We will call this a theory of relevance realization. Our
argument is analogous to the idea that one cannot have a theory of biological fitness, but one can
have a theory of the mechanisms of natural selection that realize it. To draw this all together, we
argue that perhaps the only way cognitive science can hope to circumvent this family of problems is
to develop a non-circular theory of relevance realization.
In this essay, we begin by describing how these intractable problems in cognitive science lead
to circular theories due to the issue of relevance. We then discuss an important methodological
© The Author, 2009. Published by Oxford University Press. All rights reserved.
For Permissions, please email:
Journal of Logic and Computation Advance Access published October 26, 2009
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 2 1–21
2Relevance Realization
move: the attempt to circumvent these problems through what Dennett calls ‘reverse engineering’.
But following an argument from Green, reverse engineering will require a criterion of the cognitive
which does not rely upon folk-psychological intuitions about the nature of cognition; they themselves
presuppose relevance realization and so will simply return us back to the intractable problems. We
argue that the best way to avoid this threat is to make relevance realization the criterion of the
If this is the case, we need some plausible account of what the mechanisms of relevance realization
would look like. Of course, a complete specification of these mechanisms will require significant
empirical work. The best that we can hope to provide here is a plausible account of the required
structural principles for these mechanisms. To this end, we describe three important lower order
constraints and a fourth higher order constraint used by cognitive agents which, when considered
as dynamic, opponent processes, could help to produce a non-circular structural theory of relevance
realization. Finally, we use our theory of relevance realization to critique previous theories of
relevance (e.g. Sperber and Wilson’s [47]) which we believe fall prey to a recursive regress.
Much of the machinery on which our theory runs has been borrowed directly from modern practice
in the sub-disciplines of cognitive science (e.g. linguistics, machine learning, neuroscience and
psychology). In this article, we attempt to draw connections between these various pieces of work,
and in so doing, our aim is to contribute to what we view as the already emerging framework in
cognitive science.
1.1 The centrality of relevance realization to cognitive science
The origins of cognitive science as a discipline are rooted in the research programs of the 20th century
that sought to explain cognition in terms of computation and language, thus bringing together the
disparate disciplines of cognitive psychology, artificial intelligence and linguistics. Historically, some
of the most central areas of research within this interdisciplinary framework have been problem
solving, causal interaction with the world, categorization, induction and communication [15,44].
In each of these areas significant problems arose that prevented any single theory from gaining
widespread acceptance. We will argue that these problems were related via relevance realization.
1.2 Relevance realization and problem solving
The foundational framework for understanding problem solving for both cognitive psychology and
AI is the GPS framework of Newell and Simon [33]. One of the successes of this framework was
to reveal an aspect of problem solving which initially was counter-intuitive, namely, combinatorial
explosion [26,32]. In order to understand combinatorial explosion, one needs to understand how
problems were represented in the GPS framework. In this framework, a problem is represented by
four elements: a representation of the initial state, a representation of the goal state, a representation
of all of the operators an agent can use to turn one state into another and finally path constraints
which disallow certain types of solutions [26,32]. Taken together, these elements generate a problem
space or search space which consists of all the possible sequences of states that the agent could take.
A solution consists in finding the sequence of operations which will take the agent from the initial
state to the goal state while obeying path constraints.
The GPS model was useful because it made apparent that for most problems which humans solve,
the associated search spaces are vast and complex. For example, consider a typical chess game. On
average, for each turn there are 30 legal operations you can perform, and there are typically 60
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 3 1–21
Relevance Realization 3
turns in a game. So the number of alternative sequences you would have to search in order to find a
path from the initial state to the goal is FD, where Fis the number of operators and Dis the number
of turns. So in our chess example the number of pathways you would have to search would be 3060
which is a very large number. This number of paths is far too large for any conceivable computer to
search exhaustively (consider for comparison that the number of electrons in the entire universe is
estimated at 1079).
Nevertheless, humans successfully wend their way from initial states to goal states all the time
(while respecting path constraints). How do they do it? How do they do a search through that space
in a way that is non-exhaustive, but still intelligent? In practice, people make use of heuristic search
in which large regions of the search space are not considered [33, p. 96]. But typically models
of such heuristics are hand crafted by AI practitioners for the problem of interest, and there are
no general theoretical accounts of how to mechanistically generate heuristics powerful enough to
produce human-level competence.
A major failure of the GPS framework was that it relied on the assumption that problems form a
well-defined class and that most problems could easily be turned into well defined or formal problems
[21]. It became clear that for many real-world problems (e.g. taking good notes during a lecture),
the initial state, goal states, path constraints and operators are either incomplete, vague or missing.
Such problems are known as ‘ill-defined problems’. Human being’s ability to avoid combinatorial
explosion stems from the way they can convert ill-defined problems into well-defined problems [26].
Certain actions are immediately ruled out via non-inclusion in the problem space during problem
formulation, thereby making the search space far more tractable (e.g. no one even considers including
the ambient temperature of the room while taking notes in a lecture). The key is our ability to zero
in on the relevant information and the relevant structure of the information to perform the actions
needed for good problem formulation.
However, to determine what is relevant to a problem also involves determining what is irrelevant!
As such, the smaller search space which problem formulation affords us is only achievable if we
initially consider the larger search space and segregate the relevant from the irrelevant. This leads us
back into the problem of combinatorial explosion. We are caught in a nasty circle here in which we
need good problem formulation in order to deal with combinatorial explosion, and yet good problem
formulation seems to be a combinatorially explosive problem. Only an account of how people realize
what is relevant while avoiding combinatorial explosion can break through this theoretical circle.
1.3 Relevance realization and interaction with the environment
The necessity of breaking this circle is especially apparent when we consider how agents take action
in the environment while intelligently dealing with unintended side effects. This is best illustrated
with an example from Dennett [8] about a robot trying to acquire its food in a very basic manner.
Suppose we have a robot designed to retrieve batteries as its food source and then transport those
batteries to a location where they can be consumed. Also, suppose our robot comes upon a wagon
upon which there is the battery, but unfortunately there is also a bomb on the wagon. The robot
correctly deduces that if it pulls the wagon then the battery will come along as an intended effect.
However, an unintended side effect is that the bomb comes along and destroys the robot. As its
designers, we attempt to remedy this situation by having the robot deduce not only the intended
effects of its actions but also potential side effects. When we test our improved robot in a repeat
of the original situation we find it stopped and endlessly calculating for the simple reason that the
number of potential side effects it can consider is indefinitely large. We seek to remedy this by having
the robot form a list of potentially relevant side effects. This part of our thought experiment requires
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 4 1–21
4Relevance Realization
that we have some theory of the properties of information that renders it relevant. However difficult
this assumption is let us grant it for the sake of continuing the thought experiment in order to further
see what it reveals. Once again we find our robot stopped endlessly calculating. This is because it is
creating two lists. One of potentially relevant side effects and one of irrelevant potential side effects,
and because each list is again indefinitely large, the calculation cannot be completed.
As Dennett’s example illustrates, for an agent to take action in even relatively simple circumstances
it must somehow intelligently ignore a great deal of information, but this seems paradoxical. This
requires zeroing in on the relevant information while not even considering most of the irrelevant
information. It requires putting a frame around one’s cognition. This is therefore a generalized
version of the frame problem [36], indicating that we have on our hands an essential problem that
extends well beyond this case.
1.4 Relevance realization and categorization
The same dilemmas plague another central component of research in cognitive science, namely,
categorization. The importance researchers have placed on categorization has been in part motivated
by the acceptance that there are deep connections between what is required for problem formulation
and what is required for categorization of novel information [21]. Categorization is the process
by which we create classes which support powerful inductive generalizations that are relative to the
features held in common by the members. In other words, the members of a category are only identical
in so far as they contribute to the inductive generalizations. Yet, as Goodman [16] famously noted, any
two objects can be infinitely similar or dissimilar, which presents any formal theory of categorization
with a seemingly insurmountable computational task similar to the combinatorial explosion faced
by problem solvers. Thus, the immediate issue is how we zero in on the relevantly shared properties
which will be useful for inductive generalizations. We think that this issue is encountered by schema
[38], script [39] and stereotype or prototype theories [21], as well as more recent theories which
deal with context sensitivity in categorization [14]. Each one of these families of theories involves
an implicit theory of relevance realization that is presumed to solve this computational nightmare,
wherein relevance is usually specified in terms of one static property of information such as frequency,
or invariance, or prototypicality, etc. However, it quickly became apparent that the attempt to capture
relevance in this manner fails. For example, Medin, in an influential review article noted the following
about prototype theory [29]:
Prototype theories imply constraints that are not observed in human categorization, predict
insensitivity to information that people readily use and fail to reflect the context sensitivity that
is evident in human categorization. Rather than getting at the character of human conceptual
representation, prototypes appear to be more of a caricature of it. (p. 1472)
Medin’s point is that prototypicality does not capture the relevance realization needed for
categorization, and it is therefore failing in a deep way. A much more complex and dynamic
account of relevance realization is needed. (See [30] for a more recent review that argues that the
dynamic complexity of the information integration within concept formation and use may require a
‘psychometaphysics’ in which concepts are embedded in theoretical and explanatory projects. These
are projects that would clearly require sophisticated relevance realization.)
Despite the empirical merit possessed by these implicit theories of relevance realization, they have
been inadequate in generating good models of categorization [29,30]. We suggest that this is because
relevant information cannot simply be always identical to frequent, or invariant, or prototypical
information, because relevance is context sensitive. Dissenting voices such as Barsalou [2] have
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 5 1–21
Relevance Realization 5
presented similar arguments, and argued for making the context sensitive application of information
central to what it is to be a concept. Indeed, he claimed that ‘a concept can be viewed as an agent-
dependent instruction manual that delivers specialized packages of inferences to guide an agent’s
interactions with particular category members in specific situations’ [2, p. 626].The circle that looms
here, of course, is that at any given time a context could contain an infinite number of variables or
predicates regarding the environment. A much more sophisticated account of relevance realization
is needed, which breaks through this threatening circle by dynamically integrating features such as
frequency, invariance, etc., rather than following a strict, context insensitive rule.
It is interesting to note that as it became apparent that script, schema and stereotype theories
suffered from ‘tunnel vision’ [21,51,56], researchers tried to understand intelligence in terms of
how we are embodied and embedded in the world, i.e. how we operate in the environment [59], in
order to get at the missing contextual sensitivity. On this view, if we can understand how agents
utilize the world around them that would provide foundations for how categorization and problem
solving take place. Before an agent can generate facts or formulate problems, information needs to be
processed in terms of how it is relevant for successful action in the environment. Thus, at the core of
any embodiment thesis is the idea that we are acting in the world. Of course, intelligent agents need
to be able to successfully couple their actions to future effects in the world. However, once you try
to have a reliable tracking between cause and effect you face a deep problem, namely, the problem
of dealing with side effects as discussed above. We are unaware of any research program within the
embodied cognition movement that can solve this problem.
1.5 Relevance realization and rationality
There is a tempting philosophical strategy in the face of such problems within psychology and artificial
intelligence. Most theories within these disciplines presuppose a background normativity, specifically,
that cognitive processes should be rational in nature. Perhaps one could alleviate the problems of
circularity that we have highlighted if we explicate the normativity of rationality presupposed in
much of cognitive science. Many philosophers have undertaken this project [4,20,43,49]. We
will examine two whose work we consider representative in nature and whose work highlights the
centrality of relevance realization to rational induction.
Cherniak [4] has influentially argued that you cannot say that to be rational is to simply be logical
because many algorithms couched in logic lead to combinatorial explosion. He argues instead that
to say that we are rational means that we have zeroed in on some of the relevant subset of logical
inferences for the task at hand. He therefore calls attention to the fact that relevance realization is
central to the issue of rationality in general. Cherniak attempts to explain relevance realization in
terms of memory compartmentalization. Cherniak’s idea is that a system can avoid a combinatorially
explosive search through memory for relevant information by dividing memory into compartments.
The system only searches one compartment at a time. If the compartments are labelled in some
fashion, i.e. if they are content addressable, then the search goes right to the relevant compartment
and the search is thereby constrained to just that compartment. However, Chiappe and Vervaeke
[5] have argued that this account presupposes in a vicious way the very thing it is an attempt to
explain because the formation and use of such compartments requires the ability to determine how
things are relevant to each other. We will go a step further and suggest that any explanation of
relevance realization in terms of pre-existing memory organization will presuppose the very thing it
is attempting to explain. This is because any content addressable information organization scheme
requires successful categorization of the information, and as we have outlined above, categorization
itself requires relevance realization! Naturally, memory organization can play a role within an account
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 6 1–21
6Relevance Realization
of relevance realization, but it cannot be the foundational role assigned to it by Cherniak, or to theories
like his.
In a fashion similar to Cherniak, Putnam [34] argues that inductive inferences cannot rely on an
invariant syntactic formalization. This is because the success of any particular inductive logic is
relative to the environment in which it is operational. For example, one could have an environment
which is noisy, and so requires cautious induction, but if the environment contains little noise, it is
beneficial to act less cautiously. Putnam’s idea is that inductive logics can vary in a context-dependent
fashion, e.g. by changing a caution parameter or the mechanism for generating inferences, but still
arrive at similarly rational results. To extend the example: one person can use a specific inductive
logic in some environment and inductively conclude some particular belief, while another person
could be using a different inductive logic in a different epistemic environment, and produce the
same inductively justified belief, though the underlying computational processing/states would be
different. Of course, to implement this sort of environmentally dependent inference, a system must
be able to determine the context of the inductive inference, and as we have already noted, such an
ability presupposes relevance realization. Thus, Putnam’s refinement of the normativity of rationality
returns to the problem of relevance realization, just as Cherniak’s did. We would also like to note that
Putnam’s argument highlights a point we will develop later, namely, that the attempt to solve such
problems through an invariant syntax of inference seems to be seriously misplaced.
1.6 Relevance realization and communication
Similar problems are faced by research into language and communication. Although a great deal of the
modern generative linguistics program is focused on purely syntactic questions [7], it is undeniable
that an important aspect of our linguistic cognition is the pragmatic aspect. Any attempt to explain
the nature of language must account for the relationship between language and communication, and
any theory which fails to do this is a failed theory of language. Indeed, a central insight from seminal
philosophers like Austin and Grice about language is that we are often doing more than just making
statements [1,19]. For Austin, we are often performing actions instead of stating things, and for
Grice we are often conveying more information than we are stating.
Grice developed this theory of conveyance in his theory of conversational implicature. What
comes out of this work is that you cannot fold conversational implicature into semantics because
you would overload the lexicon. Instead of folding implicature into a lexicon, it must be something
that is continually worked out between individuals. Grice’s insight is that ‘working out’ is rational
cooperation, and is thus constrained by four maxims: (i) quantity, (ii) quality, (iii) manner and (iv)
relevance. Thus, Grice’s work gives relevance a central role in communication.
It is possible to take this even further, though. Sperber and Wilson [47] argue that the first three of
these maxims all collapse into the fourth maxim of being relevant. They argue along the following
lines, the maxim of quantity is just ‘provide the relevant amount of information’, the maxim of
manner just ends up being ‘use the relevant format’. The maxim of quality is a little less clear
because it requires truthfulness, and truthfulness is not as easily seen to boil down to relevance. But,
the maxim of quality cannot be the rule ‘convey all that is possibly true about what you are thinking’,
as we have already seen in our discussion of Cherniak. So complete truth cannot be the normative
standard that people imply. Therefore, what people must be sharing are the relevant truths. We are
now presented with the same computational dilemma we encountered before: there are an infinite
number of truths, and prima facie it would be impossible to segregate the relevant from the irrelevant
in a computationally tractable manner.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 7 1–21
Relevance Realization 7
Thus, theories of pragmatics suffer the same problems with relevance realization that theories
of problem solving, categorization and action do. This time though, Sperber and Wilson explicitly
identify this problem and attempt to generate a theory of relevance. We think that is not a sufficient
account, but their account will play an important role in the framework we lay out later in this article.
To summarize, we have demonstrated that fundamental research streams in cognitive science
have encountered essential problems related to computational intractability that have frustrated any
attempts at formal theories of problem solving, categorization, action, induction and communication.
We have also shown that these problems form a family of interdependence, with the problems from
one area leading to difficulties in others. Thus, we have not only shown the pervasiveness of relevance
realization but also that its theoretical appearance is systematic.
1.7 Reverse engineering and the emerging criterion of the cognitive
In the face of these daunting problems, one may attempt to get around these thorny concep-
tual/theoretical issues concerning relevance realization via ‘reverse engineering’ [8]. Under this
methodology, attempts to directly understand cognition are abandoned, and research is instead
focused onto designing an intelligent machine. The hope is that if this research program is successful
we will understand various cognitive phenomena in virtue of having designed a machine that
exhibits the phenomena. Such principles therefore become one’s theoretical account of the cognitive
processes that generate the intelligent behaviour. Hence, artificial intelligence would allow us to
work backwards into theories of things like categorization, communication, etc., while avoiding the
theoretical circularities we have encountered in the other direction.
However, we are assuming here that we will know when a machine is in fact intelligent. The
obvious methodological question is how to establish this identity. Turing [53] famously proposed
just such a test. As Fodor [11] points out, the Turing test has an important methodological principle at
work; namely, that the test screens off certain factors of comparison. However, this principle tacitly
makes use of our assumptions about what constitutes intelligence. These assumptions will of course
be shaped by our current folk psychology and/or our explicit scientific psychology.
In 1994, Green [17], pointed out that unless we have some agreement about how to pick out
cognitive phenomena (i.e. a criterion of the cognitive), we will never be able to determine the correct
factors for comparison. Green reviews some of the common criteria. None of these is widely agreed
upon, i.e. they cannot be used to powerfully pick out examples of cognitive processes. For example,
the view that seemed to be achieving success in the mid-1980s was that cognitive processes are
inferential processes operating on syntactic representations [35]. But this fell under heavy criticism
from the connectionists [18,37,45,46]. The original formulation of the Turing test biases one towards
paying attention to inferential and language-like features of cognition. Aconnectionist would be very
unhappy with any version of a test for the cognitive which only pays attention to these factors. Without
a powerfully applicable criterion of the cognitive, the interpretation of simulations, and hence the
whole AI project, will remain seriously in question.
We propose that the systematic importance of relevance realization to cognitive processes makes it
the obvious choice for a criterion of the cognitive. Put succinctly, any attempt to engineer an intelligent
system must ultimately focus on the development of a system that can realize relevance. In fact, we
believe that a great deal of current work in machine learning and theoretical neuroscience should be
viewed in these terms. Examples include current work in categorization [24,25,55], optimal control
[28,40,52] and reinforcement learning [9,10,50]. In each of these cases a significant focus of the
research program is the development of systems that can cope with the computational intractability
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 8 1–21
8Relevance Realization
rooted in the need for relevance realization. Indeed, the goal of such research could be said to be
the development of systems that can determine relevant features, controls or actions for problems
encountered in the real world. Thus, we put forward that such a criterion of the cognitive is already
emerging [59]. Unlike in the past where generally the criterion was intuitively generated from our
folk psychology, current work focuses on methods to solve the problems that were encountered in
the past when applying such intuitive criteria. Therefore, the methodological move is to base one’s
criterion of the cognitive on whatever facilitates solving these difficult technical problems. As such,
an explicitly developed theory of relevance realization will help the development and application of
new techniques and theories, and ultimately, our understanding of cognition.1
1.8 Towards a theory of relevance realization
Before we begin to articulate even a cursory theory of relevance realization, we feel it is important
to identify a framework that will avoid circularities. This is because theoretical work on relevance
will lead to regress both due to its privileged position within cognition, as we have shown above, and
due to its position within the practice of science itself. In this section, we describe how this leads to
circularities and we identify three guidelines for theorists that will help avoid them. The first guideline
is that a theory of relevance is impossible, and instead a theory of self-organizing mechanisms for
relevance realization is what is required. The second is that a theory of the mechanisms of relevance
realization must not be representational or syntactic, but economic. The third is that a theory of
relevance realization cannot rely on a completely general purpose learning algorithm, but must
involve competition between multiple competing learning strategies.
2 The importance of a theory of a self-organizing relevance realization
If relevance realization is to be the criterion of the cognitive, a naive assumption might be that you
have to come up with an account of relevance with which you can pick out all relevant things in order
to find generalizations over the class. In this sense, you may try to come up with a theory of what
relevance is. This is the tactic that has been employed by other researchers to date [48]. However,
we believe that this is a fundamental mistake.
1It may be helpful to contrast what we believe to be the central framing metaphor for the criterion which held prominence
until the mid-1980s with the new emerging metaphor. The classical metaphor is that cognition is essentially computation and
that the brain is essentially a computer. The idea is that information is organized around inferential, and syntactic relations
which are isomorphic with a linear causal order. This program is implemented on a static hardware, and because the hardware
is static it is almost completely irrelevant to the software. The development of the hardware is also irrelevant except insofar
as it is a process for producing the mature hardware that can run the software. The computer is a stable logic machine.
In contrast, we might call the new metaphor the Logos (capturing both the sense of logistics and the Greek sense of the
term as: making information belong together) Multi-Machine (LMM). Here, information is organized in terms of economic
properties (which we will discuss in greater detail later) and relevance relation, namely, how information can be economically
integrated together to support successful interaction with the world. This is isomorphic not with a linear causal order, but with
a circular causal order of a self-organizing dynamic system. This system is instantiated in a plastic neural network. The brain
is a Multi-Machine: a machine which can make itself into new kinds of machine such that it not only learns but also increases
its capacity for learning. In this new model there is no clear line between the ‘hardware’ and ‘software’, since both influence
each other as they run. Thus, the developmental history of the hardware is always relevant to the explanation of cognition.
Note also that one of the things that an LMM can develop is a computer, i.e. one of the machines within the multi-machine
of the brain can be a computer.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 9 1–21
Relevance Realization 9
To begin with, consider by analogy the role played by ‘fitness’ in the theory of evolution. A
common confusion regarding fitness is that what makes a creature fit is the possession of one or more
of the defining features of fitness. Thus, people sometimes misconceive evolution as the designing
of particular features like speed, intelligence, acute vision, etc. Of course, in reality the class of all
possible organisms which are fit is completely heterogeneous, unstable and dependent on context.
As such, we can make no systematic inductive generalizations about the class of fit organisms.
What evolutionary theory provides is not an account of the biological features that define fitness,
but a mechanism by which fitness is realized in a contextually sensitive manner. Therefore, by
strong analogy with the centrality of natural selection in biology, we do not really want a theory
of relevance. We want instead a theory that articulates a mechanism for how relevance is realized
in a contextually sensitive manner. We assert that the need for approaching relevance in this way
is a direct result of circularities inherent in attempting to build a theory of a phenomenon which
underpins all the other phenomena within the discipline. This is especially the case for relevance due
to the fact that the practice of science itself relies on these cognitive phenomena which are dependent
on it. This has previously been highlighted by Chiappe and Vervaeke [5]. We briefly review the
issue here.
Any scientific statement has to be protected by provisos which keep it from being trivially falsified.
For example, we say that ‘sugar is soluble’, even though someone might for instance freeze the water
just as the sugar was added, etc. The list of such ‘falsifications’ is obviously long and heterogeneous,
but they are considered irrelevant to the scientific statement. It looks like we have run into a circularity
here: the very articulation of a theory of relevance would require an implicit identification of what
is and is not relevant regarding the application of the statement. This runs us into a kind of chicken
and egg problem for cognition.
In a similar fashion, any scientific statement is going to rely upon pragmatic conveyance for
its interpretation and understanding. But attempts to render pragmatic conveyances into semantic
statements meets the problem that a stipulation of these conveyances relies upon implicatures.
Normally, we rely upon people’s ability to realize relevant implications and implicatures in order to
make communication practicable (see Grice/Sperber and Wilson above). But, if our goal is to design
a theory of relevance itself it should not include an assumption of the relevant implications for a
cognitive agent. Whatever else a scientific statement needs to be, it needs to be communicable since
science is essentially a community enterprise, so this difficulty with conveyance cannot be ignored.
Finally, any scientific explanatory statement(s) involves inference to the best explanation which
requires reference to the contrast class (the set of explanations it is better than). Of course, given
the underdetermination of a theory by its data, the number of alternative yet acceptable explanations
is infinite. Typically the abductive selection that allows scientists to ignore these alternatives is
usually explained in terms of concepts such as simplicity and similarity. Yet, this relies heavily upon
relevance since it will depend on the relevant features of the explanations for determining simplicity
and similarity. So a theoretical explanatory statement about the nature of relevance would rely again
upon the very thing we are attempting to explain.
For all of these examples we end up with a seemingly vicious infinite regress. Importantly though,
there is an implicit foundationalism in this discussion; i.e. that the only way to stop this infinite
regress is to find some ‘proto-relevance’ around which the chain of explanation is designed. However,
there is a ready alternative to such foundationalism—an evolutionary coherentism in which relevance
realization is a dynamic, self-designing, self-organizing process that is not equated with an immutable
identity. In the same fashion that evolutionary theory dissolves the chicken and egg problem,
we believe that a non-definitional theory that proposes self-organizing mechanisms for relevance
realization will be able to dissolve the threat of vacuous or cyclic explanations in cognitive science.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 10 1–21
10 Relevance Realization
We need to radically give up an ‘intelligent design’ framework for understanding the generation of
relevance realization by cognitive systems.
3 The requirement for an economic model of relevance realization
Even if we resolve the theoretical regress by moving to a self-organizing, non-definitional model,
there is a further difficulty remaining which was described by Vervaeke [57]. It seems prima facie
that relevance realization is going to be relative to the interests and goals of an organism, and interests
and goals are about future states of affairs. To direct behaviour towards some future state of affairs
requires some representation of that state of affairs. However, representations are aspectual by nature
since one does not represent all of the aspects or features of a thing. This means that one needs
relevance realization to generate good representations because one has to pick which aspects are
relevant to represent. This is a patently vicious explanatory circle. Note that this is a problem in our
attempts to explain the brain’s behaviour; it is obviously not a problem with the brain’s functioning.
So, the somewhat unintuitive move is to consider that theories of relevance realization should
be pitched at a sub-representational level to avoid this theoretical conundrum. What this means is
that the theory can only initially make use of completely immanent properties and relations of the
information available to the brain. The prototypical response to this issue in cognitive science has
been to drop to the logical/syntactic level of explanation since the formal properties are completely
self-contained and only become representation when they have been assigned content. This is the
theoretical framework behind the many variants of computation functionalism.
One of the founding figures of computational functionalism, Jerry Fodor [12], has recently made an
argument that one cannot capture relevance in the syntactic structure of tokens within a formal system.
Fodor’s argument is that notions like relevance or centrality or importance are all aspects of cognitive
commitment, i.e. how much a system cares about something and devotes its resources to it. Cognitive
commitment is an economic issue, and as such it is both globally defined and contextually sensitive.
This means that it cannot be captured in the syntax of a token since the syntax by its very logical nature
does not consider economic issues, must be locally defined, and operates in a contextually invariant
manner. As Fodor puts it, syntax is locally defined but relevance is globally defined and therefore
cannot run off of syntax alone. Fodor rightly regards this as a devastating problem for computational
functionalism because very many important processes cannot be captured by syntax. This means
that any viable theory of relevance realization is going to have to make use of sub-syntactic (e.g.
using vectoral representations) properties of information whose operations are locally defined but
have global effects. This tact is becoming relatively widespread in cognitive science, with examples
surfacing in the work of Hinton [24] and Gärdenfors [14].
One possible way to meet this demand is to turn to the economic properties of information and
action. We notice that in economies, decisions are made locally (e.g. you buy milk at the grocery
store), but these local decisions contribute to the global organization, and of course, the global
organization constrains future local processing. This is all done without centralized control, i.e. it is
self-governing. This is exactly the sort of mechanism which Fodor was looking for.
The economic approach uses internal measures of cost (e.g. metabolic), and reward (e.g.
dopaminergic). The reader may worry that we are invoking a pre-established harmony between
the internal running of the cognitive economy and successful behaviour in the world. The worry of
course is that pre-established harmonies often presuppose a god-like designer. However, we have no
such worry since we can confidently presume that, to the extent that it exists, evolution has worked
out this pre-established harmony.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 11 1–21
Relevance Realization 11
Obviously, the fact that the cognitive economy is internal, does not mean it is causally isolated.
Like all economies the cognitive economy has imports and exports. It imports sensory data, and
exports motor commands to the world. The exports have effects in the world which to some degree
constrain the future imports. Just as we can intelligently talk about the American economy as a
distinct entity that is nevertheless causally interwoven with other economies, we can talk about the
internal cognitive economy even though it is causally interwoven with the world. The important
relation to discover then, the pre-established harmony, is the balance of internal economic variables
whose processing results in successful interaction with the world.
4 The impossibility of a general learning algorithm solution
Relevance realization must be a set of pervasive constraints on processing rather than a specific
machine for realizing relevance. The reason for this is that if relevance realization took place within
a specialized mechanism, the mechanism itself would merely confront the frame problem. The
problems of how to decide what to ship there, and how it might avoid the combinatorial explosion of
doing so, would immediately and continually arise. This sort of a move only shifts the problem. It does
nothing to solve it. Thus, there cannot be a relevance realization mechanism in any straightforward
sense. As we have said we need to give up any intelligent design framework for understanding
relevance realization.
One area where this point can be brought to bear is in the field of meta-learning. This is an
exciting new field of research that exemplifies many of the features of the emerging framework
we are explicating. In meta-learning systems, one has basic learners that apply specific learning
algorithms/strategies to current problems. In addition to this one has a higher order meta-learner
whose job it is to assess the applicability (an important concept as we shall see) of various different
learning strategies across time so as to improve the selection and operation of the base learning
strategies. This makes the system a self-adaptive learner. However, one conceptual problem which
seems to confront this approach is explained well in a recent review by Vilata and Drissi [58].
They point out that the advantage of meta-learning over standard learning is that the bias (the set of
assumptions in the algorithm that restricts and structures the problem space) in the base-level learning
algorithm is no longer fixed a priori but is dynamically modified by the meta-learner. This addresses
one of the central issues of relevance realization in problem solving. However, the meta-learner itself
must have a learning algorithm with such an a priori fixed bias. One can address this by having a
meta-meta-learner, but an obvious infinite regress now ensues. This conceptual problem seems to
involve an imposition of an intelligent design framework in which there must be a specific machine
that ultimately decides how relevance is assigned. We argue that the solution to such problems is
to make the meta-learning self-organizing and immanent to the learning. We outline below how we
think this might be done.
It is important to note though that a meta-learner cannot simply be a general purpose, higher order
learning algorithm. There were hopes in the early machine learning literature that a sort of general
learning algorithm could be found that would do well on all problems [3,54,60]. But, just as the
GPS framework met insurmountable problems, the hopes of a general purpose learning algorithm
were eventually dashed as well:
Although the human brain is sometimes cited as an existence proof of a general-purpose learning
algorithm, appearances can be deceiving: the so-called no-free-lunch theorems [Wolpert, 1996],
as well as Vapnik’s necessary and sufficient conditions for consistency [Vapnik, 1998, see],
clearly show that there is no such thing as a completely general learning algorithm. All practical
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 12 1–21
12 Relevance Realization
Table 1. Economic constraints and interactional problems from which relevance realization
Internal Economic Properties External Interactional Properties
Cognitive scope Applicability
Compression Particularization General Purpose Special Purpose
wij =−ηJ(·)
wij αwij
Cognitive tempering Projectability
TD Learning Inhibition of Return Exploiting Exploring
Cognitive prioritization Flexible gambling
Cost function #1 Cost function #2 Focusing Diversifying
J1,2(β,α )=1
learning algorithms are associated with some sort of explicit or implicit prior or bias that favours
some functions over others. Since a quest for a completely general learning method is doomed
to failure, one is reduced to searching for learning models that are well suited for a particular
type of task.
Given the no-free-lunch theorems—which demonstrates that all learning algorithms are inherently
tuned to some subset of problems—one important tact a cognitive system can use is to adopt strategies
that are complementary in that they have goals that are in a trade-off relationship. The system can then
use opponent processing in order to continually redesign the learning strategy it is using. Opponent
processing is a powerful way to have self-organization implement a heuristic solution to the no-free-
lunch restriction by having a continual competitive trade-off between two or more complementary
strategies. In biology there is evidence at many levels of analysis that evolution has centred upon this
as a solution (e.g. the control and maintenance of homeostasis in the body by the autonomic nervous
system which plays the sympathetic and parasympathetic systems against one another). In this way
nature creates mechanisms that can strategy shift in a completely self-organizing manner and also
immanent to the processing. This means that we need an account of how processing is constrained
to operate in this manner. The meta-learners speciality is thus the problem of balancing a series of
constraints. In the remainder of this article, we speculate about the nature of some of these constraints.
In the next three sections, we identify three important interactional problems that cognitive agents
face and attempt to specify three corresponding internal economic processes that involve competition
between opposing goals (Table 1). We believe that the mechanism behind relevance realization is
ultimately the process that enables the brain to balance these competing goals. Thus, we argue that
relevance is never explicitly calculated by the brain at all, but the high-level phenomena of relevance
realization emerges from the brain’s attempt to dynamically balance its economic requirements. The
following characterization of these economic processes is still in its infancy, but we hope that this
initial sketch will inspire other researchers, and possibly lay the foundation for a much more detailed
account of the mechanisms underlying this core component of cognition.
4.1 The applicability problem: cognitive scope
Let us say that you are trying to engineer processes to control your cognitive economy in order to
maximize some sort of future-discounted reward. One of the engineering problems which confront
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 13 1–21
Relevance Realization 13
you at the interactional level is whether to go for general purpose or special purpose machinery. The
question is what internal economic processing can you use to optimize your applicability as a machine
in the world. The point—stemming from optimal control and the no-free-lunch theorems—seems
to be that hard commitment to either of these strategies (general purpose or specific) is undesirable.
The engineering trick is to make use of opponent processing so that one is continuously dynamically
designing and redesigning the right kind of tool for the task at hand (i.e. the sort of tool which will
allow you to maximize reward signals for the task). How though does one get internal economic
processing to trade-off between these interactional styles of processing (general purpose versus
special purpose)?
One formal example is typified by an engineering solution for training neural networks. When
updating the weights in a neural network, one typically follows the negative of the gradient (or
estimate of the gradient) of some performance metric, J(·) with respect to the weights, e.g.
wij =−ηJ(·)
where, ηis a learning rate and wij is a synaptic weight in the network which connects neuron iwith
neuron j. If the weight update term consists solely of this negative gradient term, networks will tend
to over-fit the data that it trains on. A common solution to this problem is to introduce a ‘weight
decay’, or regularization term into the weight update so that,
wij =−ηJ(·)
αwij (2)
where, αis the magnitude of weight decay applied. Such a term penalizes large weights and pushes
them towards zero [22]. This has the effect of simplifying the network which in turn makes it better
at generalizing from its training experience. The ratio of the gradient and weight decay terms will
determine the extent to which a network focuses on generalization at the expense of performance on
data it has trained with.
A network which is making use of weight decay is opponent processing compression against
particularization (i.e. explaining the data well). This opponent processing we refer to as cognitive
scope (CS)—trying to capture both the spatial analogy of general versus specific and the perceptual
analogy of effectiveness of perception as in ‘microscope’ and ‘telescope’. Our claim is that a brain
constrained to internally processing CS tracks the opponent processing between general purpose and
special purpose machines and thereby optimizes its applicability of information for action within the
Here we are broadening Turing’s insight that you can track rationality in the world by having
a device that just internally pays attention to the logical syntax of the information. However,
given what we have said about rationality, we think that Turing’s conception of rationality as
just the logical management of inference is insufficient. A lot of rationality has to do with the
realization of relevance [13]. Our point is that we can track this aspect of behavioural relevance
realization by internally processing CS. This extension makes use of Hinton’s insight, which we
term internalization (in contrast with representation), and which is typified in such methods as the
wake/sleep learning algorithm [23]. In such learning algorithms, the way one gets neural networks
to engage in unsupervised learning, i.e. learning where the target value in the world is not known,
is by having the network relate to itself in a completely internal fashion. It relates to itself in an
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 14 1–21
14 Relevance Realization
internal fashion such that it treats itself as a micro-environment2in which the procedural abilities to
track environmental variables can be trained. These procedural abilities are then turned on the world
in an attempt obtain the actual information from the world which can then be used to improve the
micro-environment so that the whole system bootstraps itself up into effective interaction with the
world. This internalized training of CS is what we mean by saying that the internal processing of
economic variables, such as CS, track the interactional properties such as determining applicability.
So, part of what constitutes relevance realization is when CS tracks applicability so that the
cognitive organism is continually redesigning itself as some trade-off between a general purpose and
a special purpose machine. Note also, how we have specified a general constraint on information
processing that is completely immanent to such processing.
4.2 The projectability problem: cognitive tempering
Another engineering problem that faces individuals in the world is whether to go for exploitative
machinery or exploratory machinery (Table 1). An exploitative machine is one that tends to stick
with its currently known actions and to simply select from those actions the one that it thinks will
have the highest payoff now. In contrast, the exploratory machine will forgo immediately available
payoffs to look for the possibility of contexts/actions which will produce higher payoffs later on.
Neither one of these strategies pursued exclusively is a good overall problem solving strategy. The
problem with being exclusively exploitative is that there is a good chance that there are much higher
payoffs elsewhere and this therefore constitutes a significant opportunity cost for the explorer. You
do not want to have a machine that only looks for information which facilitates exploitation in the
immediate future since other information may help to build towards larger future payoffs.
On the other hand, a machine which spends almost all of its time exploring for opportunity runs
the risk of losing available payoff as it searches for better payoff. In a similar fashion to the opponent
processing mentioned for general purpose versus special purpose, we would want a machine that is
dynamically moving between these strategies. We want a system that is optimizing for the interactional
property that we call projectability, which is the dynamic balance between exploiting the here-and-
now and exploring the there-and-then. The system has to setup a projective relationship between the
actual here-and-now with the possible there-and-thens. How then do we construct a machine which
is able to optimize for projectability solely in terms of internal economic properties?
One promising way of doing this is to couple powerful reinforcement algorithms such eligibility
trace temporal difference (TD) learning [50] with an temporally decaying inhibition of return trace. In
TD learning (Table 1shows the value function update equation for the most basic form ofTD learning
as discussed by Sutton and Barto (1998)), a memory trace (e.g. an eligibility trace) of recent actions
performed by a machine may be kept. When reward (or punishment) is encountered, the actions in
the trace are credited with having brought about the reward (or punishment), and so are reinforced (or
weakened). Those actions which were performed most recently are reinforced most strongly, while
2The original internal environment is nothing more than statistical patterns in the data stored within a neural network.
However, such networks are very good at picking up very complex statistical patterns that pick up correlational and causal
patterns. These complex structures within the data can serve as a virtual world upon which to train procedural abilities. It
is important to remember that at first the structures of this internal ‘world’ need not be an accurate representation of the
world. Internalization only needs a demanding informational environment in which to train the skills for picking up complex
information from complex environments. Of course, as an internal environment the target value is known to the system and can
be used in correction. The rate of practice can be adjusted, and variation can also be introduced, all to improve the procedural
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 15 1–21
Relevance Realization 15
those which happened long ago are reinforced only a little. This technique of apportioning reward to
actions backward in time allows an agent to learn to perform actions for payoffs in the distant future.
TD learning reinforces action patterns (i.e. reinforcement of return), which tend to have you return
to states/actions temporally associated with high reward (this is the element which pushes you to be
exploitative). On the other hand, to explore in an intelligent fashion, the machine can lay down an
inhibition of return trace a temporally decaying memory trace which indicates to the machine not to
return (in a soft way) to states/actions it has recently seen and thereby promotes exploration [27].
This trace would be traded off against the reinforced action patterns to find an intelligent way to
trade-off exploitation with exploration.
We call this internal processing of the opposition between inhibition of return and reinforcement
of return cognitive tempering (CT). We call it CT because we are trying to capture the metallurgic
sense of neither being too flexible nor too inflexible, and also the root word ‘temp’—having to do
with time. What we are proposing is that CT can be trained to track projectability. A system which
is constrained for working out CT will find good projectability in the world.
4.3 The problem of flexibly gambling: cognitive prioritization
A third interactional problem faced by cognitive agents is how to gamble flexibly in the face of
ambiguous information (Table 1). Ambiguity, whether caused by the introduction of noise (e.g.
perceptual) or in overlap of the entities in the environment which can generate the same information,
means that one is always gambling with the commitment of one’s cognitive resources. Then the
interactional issue is how one is to wager one’s cognitive resources in the world. Betting should be
flexible because, for instance, the scarcity of one’s internal reserves ought to cause a trade between
focusing and diversifying as betting strategies. For example, if an agent is very thirsty it will tend
to gamble all of its efforts on getting water, i.e. it will focus its wagers on this project. As soon as
thirst is satiated though, the agent will begin to pursue a diversity of problems. As we will discuss
in the following section, Montague points out that agents ought to ‘care’ differentially about the
environment because they run on batteries [31].
What are the internal economic constraints that can track these interactional trade-offs? In contrast
with the other two economic constraints, this one is much more conjectural in nature. What we intend
to do here is to provide an argument for the plausibility of being able to produce a mathematical
formalization of this constraint. Whereas the internal economic constraints we call CS and CT had to
do with how cost functions might be heuristically optimized, cognitive prioritization (CP) has to do
with the structure and prioritization of cost functions. In short, one of the things which allows agents
to be truly successful in the world is adapt not only their behaviour to suite a given task, but also to
adapt the sorts of tasks they are interested in optimizing.
In order to make clear what we mean, consider the following example: an animal has two basic
operational goals in its life; one is to find food to sustain its energy stores, and the second is to
avoid being food for another larger animal. All other projects are asymmetrically dependent on these
ongoing projects.3Both of these external goals are tracked internally by two cost functions, J1(·) and
J2(·), respectively. One of these cost functions, J1, is an internal metric which tracks how well you
doing at maintaining your energy reserves. The second, J2, is an internal metric which tracks how
3In a personal communication, Zachary Irving pointed out that didactically it may help to think that, instead of just having
two cost functions—food and predator avoidance—you had 10 (e.g. mate seeking, water seeking, sleep, young rearing, etc.).
If this was the case, then when beta gets very low, a learning system looks very specifically at food seeking, at the cost of all
those other activities that are potentially relevant.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 16 1–21
16 Relevance Realization
Table 2. Higher order oppositional processes on economic constraints underlying relevance
Efficiency (Selection)
Compression-Generalization TD Learning-Exploiting -Focusing
Particularization-Specialization Inhibition of Return-Exploring -Diversifying
Resiliency (Variation)
well you are doing at predator avoidance. Suppose that J1is multiplied by a leaky integrator function
which tracks the level of satiation for the animal (it integrates acquisition of resources and leaks as
they are depleted).
Now, suppose that the animal is therefore interested in what we will call a joint cost function, J1,2:
βJ1(α,·)+J2(α,·) (3)
where, βis the leaky integrator which indicates the level of satiation, and αis a vector containing
the animal’s adjustable parameters governing action, J2is a relatively constant cost function which
resolves ambiguity in this manner: it emphasizes misses over mistakes; i.e. it is much more important
to not miss the predators approach than it is to mistake something for a predator that is not one. On
the other hand, J1will tend to resolve ambiguity by emphasizing mistakes over misses: i.e. it is
initially more important not to mistake poisonous or inedible material for food than to miss food.
But, as energy resources deplete, the system gives more and more emphasis to not missing food
opportunities than to mistakenly eating poison or inedibles. When satiation, β, is low (e.g. 1is
large), J1becomes dominant and can put pressure on J1,2to focus resource investment into the food
acquisition project.
Of course, we are not at all committed to there being only two cost functions—there are likely many
which are flexibly trade-off against each other. The example presented above is merely illustrative of
how CP may operate as a constraint within relevance realization. Also, the approach we have taken
here easily permits the addition and balance of other cost functions.
4.4 Interaction between the three constrains
We think that the three internal economic constraints: CS, CT and CP, are all mutually constraining
within an internal economic arena. In addition, we think that there are higher order constraints on this
process of running the internal cognitive economy. There is a selective constraint to be as efficient
as possible in this economy, but there is also an opposing constraint to be resilient, which creates a
higher order oppositional process on top of the constraints we have identified (Table 2). The main
problem is if you just push for efficiency, you can lose a lot of latent preadaptive functions which
may turn out to have long term value to you. So, you do not want to downsize too much and become
brittle in your ability to handle environmental perturbations—you want some overlap, redundancy
and variation in your processing.
Resiliency introduces variation into the economy and efficiency introduces selection—they are
opponent to one another and so the whole system will tend to evolve. This is very similar to Siegler’s
idea that cognitive development shows significant parallels to the process of evolution [42]. Thus,
relevance realization is continually evolving. It is continuously self-adaptively self-designing. This
is cashed out neurologically in the developmental complexification [41] (a dialectic of integration for
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 17 1–21
Relevance Realization 17
efficiency and diversification for resilience) of the brain’s functionality. Putting these ideas together
implies that cognition is inherently developmental in nature, rather than development just being the
peripheral issue of how cognition emerges. Thus, we may speculate that developmental psychology
ought be seen as central to cognitive psychology.
The constraint of efficiency is specifically discussed by Montague [31] as being important to getting
cognitive systems to ‘care’ about information, i.e. find information relevant.According to Montague,
such caring will make it possible for cognitive systems to choose what information to pay attention to
and which actions to perform. His basic argument is that because organisms run on energy reserves,
what he calls ‘batteries’, all of the cognitive processing of real-world organisms is constrained to be as
efficient as possible. Although he does discuss some interesting ideas about internal communication
and modelling, he does not explicate how in general the constraint for efficiency is implemented in
all processing.
In contrast, Sperber and Wilson [47] much more explicitly develop such an account. According
to Sperber and Wilson information is relevant to the degree to which it trades off between the
maximization of cognitive effect and the minimization of cognitive effort. Relevance is a kind of
cognitive profit, and information is more relevant if it is more efficiently obtained, i.e. more effect
for less effort. We think that there are very important insights in this approach. There is an emphasis
on economic properties that are internally specified, and there is use (at least implicitly) of self-
organization through opponent processing.
However, we do think that there are important problems with the attempt to equate relevance
with efficiency. First, is that since relevance is defined as efficiency it is not possible according to
Sperber and Wilson to be inefficient in processing and realize relevance. Since it is plausible that
the brain also pursues resiliency, it may often process information in a manner that is currently
inefficient so that it does not lose the ability to repair, relearn or redesign itself in the future.
We suggest rather than efficiency defining relevance, it should be thought of as a higher order
constraint operating in an opponent fashion with the higher order constraint of resiliency. This
opponent processing instantiates the goal of making cognition continually evolve. Naturally saying
that relevance realization is cognitive evolution is irredeemably vague. In our proposal, the evolution
is specified in terms of the interaction of higher order constraints which are further specified in
terms of the interaction of lower order constraints which are further specified in terms of opponent
processing in and between cost functions operating in an economic manner. Though of course, there
is still an enormous amount of work to do, both theoretically and experimentally, to build on the
sketch we have given here.
The second problem for Sperber and Wilson’s theory is the problem of how the efficiency is
specified in terms of economic properties. Although there is great insight on Sperber and Wilson’s
part about the role of economic properties, there is nevertheless confusion in the formalization
of these properties. Sperber and Wilson confuse the economic level with both the syntactic and
semantic levels of processing. For example, they discuss relevance processing in terms of logical
inferential relations which are clearly at the syntactic level, and they discuss cognitive effect in terms
of belief revision which is clearly pitched at the semantic/representational level. This intrusion of
the syntactic and semantic levels subjects their theory to the kinds of circularities we noted earlier.
This conclusion was clearly pointed out by Chiappe and Kukla [6]. They note that for Sperber and
Wilson the calculation of cognitive profit is constrained by the information that is active in the current
context. We would argue that this is a dim recognition of the importance of interactional properties
like the applicability of processing that need to be further specified and incorporated significantly
into the theory. In any case, Chiappe and Kukla point out that Sperber and Wilson need to explicate
how the initial context is updated and developed. Sperber and Wilson do this in terms other potential
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 18 1–21
18 Relevance Realization
contexts from memory which the organism can select. According to the theory the organism needs to
select a context that maximizes relevance (understood as cognitive profit). This is, of course, circular
in that determining which context from all potential contexts will do this is the kind of problem
for which relevance realization is needed. Sperber and Wilson attempt to address this by arguing
that contexts (in memory) have accessibility relations to each other by means of which the benefits
and costs of moving between contexts can be calculated according to cognitive profit. Note that
there is an insight here into relevance realization constraints operating at different levels of analysis.
However, please also note that the explore versus exploit problem has been transferred inside in that
as the cognitive system moves between candidate contexts in memory it must decide if it should stop
at any currently valuable context and exploit it or keep exploring for a potentially more valuable
context. While this is problematic for Sperber and Wilson’s explanation, it does suggest a way in
which memory organization could serve as an internal environment for the further internalization
and training, i.e. bootstrapping, of CT.
Yet, such memory organization cannot be foundational as Chiappe and Kukla point out. As we
noted earlier, memory organization itself presupposes relevance realization and so presupposes the
very process it is being used to explain. Chiappe and Kukla point out that Sperber and Wilson have no
account of memory organization and so the whole account is circular in nature. Sperber and Wilson
[48] inadequately address this by arguing that these foundational problems are probably addressed
by some combination of modularity (special purpose machinery) and something like a blind hill-
climbing algorithm (general purpose machinery). There is a lot of confusion in this answer between
interactional properties and economic properties, and about how these interactional properties are
organized in processing, and how this processing is realized in a completely internal economic
fashion. There is no clear discussion of how relevance realization could develop and bootstrap itself
up into sophisticated relevance realization. All of these things need to be carefully teased apart and
the relations between them worked out if the charge of circularity is going to be dissolved. We believe
our current theory begins to do just this, although much more work needs to be done.
5 Conclusion
We have argued that relevance realization is a pervasive problem within cognitive science and a new
framework for doing cognitive science is emerging in which relevance realization is the criterion of
the cognitive. As such, we believe that the explication and explanation of cognition will ultimately
be in terms of processes of relevance realization. Further, we have argued that this framework
is beginning to discover and grapple with the theoretical and technical tools required to address
questions concerning the mechanisms of relevance realization in the brain. We have sketched what
we believe are the crucial points in such an explanation of relevance realization. Although we are no
doubt wrong in detail, we believe that we have shown that it is very plausible that the correct answer
will be turn out to be relevantly similar to the one we have presented here.
We would like to acknowledge the help of Zachary C. Irving, Najam Tirmizi, Leo Ferraro, Vladislav
Sekulic and Alex Lamey.
[1] J. L. Austin. How to do Things with Words. Oxford University Press, 1962.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 19 1–21
Relevance Realization 19
[2] L. Barsalou. Situated Conceptualization. In Handbook of Categorization in Cognitive Science,
H. Cohen, and C. Lefebvre, eds, Chapter 28, pp. 619–650. Elsevier, 2005.
[3] Y. Bengio and Y. LeCun. Scaling learning algorithms towards AI. In Large-Scale Kernel
Machines, pp. 321–358. MIT Press, 2007.
[4] C. Cherniak. Minimal Rationality. MIT Press, 1986.
[5] D. Chiappe and J. Vervaeke. Fodor, cherniak, and the naturalization of rationality. Theory and
Psychology,7, 799–821, 1997.
[6] D. L. Chiappe and A. Kukla. Context selection and the frame problem. Behavioral and Brain
Sciences,19, 529–530, 1996.
[7] N. Chomsky. The Minimalist Program, 3rd edn. The MIT Press, 1997.
[8] D. Dennett. The robot’s dilema: the frame problem in artificial intelligence. In Cognitive Wheels:
The Frame Problem of AI, Greenwood Publishing Group Inc., 1987.
[9] K. Doya, K. Samejima, K. Katagiri, and M. Kawato. Multiple model-based reinforcement
learning. Neural Computation,14, 1347–1369, 2002.
[10] K. Doya. Reinforcement learning in continuous time and space. Neural Computation,12,
219–245, 2000.
[11] J. Fodor. Psychological Explanation: An Introduction to the Philosophy of Psychology. Random
House, 1968.
[12] J. Fodor. How the mind works: what we still don’t know. Daedalus,135, 86–94, 2006.
[13] R. M. French. Subcognition and the limits of the Turing test. Mind,99, 53–65, 1990.
[14] P. Gärdenfors. Conceptual Spaces: The geometry of thought. MIT Press, 2000.
[15] L. R. Gleitman and M. Liberman. An Invitation to Cognitive Science:Volume 1: Language, 2nd
edn. MIT Press, 1995.
[16] N. Goodman. Seven strictures on similarity. Bobbs-Merrill, 1972.
[17] C. Green. Fodor, functions, physics and fantasyland: is AI a Mickey Mouse discipline? Journal
of Experimental Theoretical Artificial Intelligence,8, 95–106, 1996.
[18] C. D. Green and J. Vervaeke. What Kind of an Explanation, if Any, is a Connectionist Net?
pp. 201–208. Captus, 1996.
[19] H. P. Grice. Studies in the Way of Words. Harvard University Press, 1989.
[20] G. Harman. Change in View: Principles of Reasoning. MIT Press, 1988.
[21] J. Haugeland. Artificial Intelligence: The Very Idea. MIT Press, 1985.
[22] G. Hinton and D. van Camp. Keeping neural networks simple by minimizing the description
length of the weights. In Proceedings of the Sixth Annual ACM Conference on Computational
Learning Theory, pp. 5–13. ACM Press, 1993.
[23] G. E. Hinton, P. Dayan, B. J. Frey, and R. M. Neal. The “wake-sleep” algorithm for unsupervised
neural networks. Science,268, 1158–1161, 1995.
[24] G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief nets. Neural
Computation,18, 1527–1554, 2006.
[25] G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural networks.
Science,313, 504–507, 2006.
[26] K. J. Holyoak. Problem Solving. An Invitation to Cognitive Science: Thinking, E. S. Edward
and N. O. Daniel, eds, Chapter 8, 2nd edn, pp. 267–295. MIT Press, 1995.
[27] R. M Klein and W. MacInnes. Inhibition of return is a foraging facilitator in visual search.
Psychological Science,10, 346–362, 1999.
[28] W. Li and E. Todorov. Iterative linear quadratic regulator design for nonlinear biological
movement systems. 1st International Conference on Informatics in Control, Automation and
Robotics,1, 222–229, 2004.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 20 1–21
20 Relevance Realization
[29] D. Medin. Concepts and conceptual structures. American Psychologist,44, 1469–1481,
[30] D. Medin and L. Rips. Concepts and Categories: Memory, Meaning and Metaphysics. The
Cambridge Handbook of Thinking and Reasoning, K. J. Holyoak and R. G. Morrison, eds,
Chapter 3. Cambridge University Press, 2005.
[31] R. Montague. Your Brain is (Almost) Perfect. Plume, 2006.
[32] A. Newell and H. Simon. Human Problem Solving. Princeton University, 1972.
[33] A. Newell and H. A. Simon. Computer Science as Empirical Inquiry: Symbols and Search.
Mind Design II, J. Haugeland, ed., Chapter 4. MIT Press, 1997.
[34] H. Putnam. Representation and Reality. MIT Press, 1991.
[35] Z. Pylyshyn. Computation and Cognition: Toward a Foundation for Cognitive Science. MIT
Press, 1984.
[36] Z. W. Pylyshyn. The Robot’S Dilemma: The Frame Problem in Artificial Intelligence. Ablex
Publishing Corporation, 1988.
[37] W. Ramsey, S. P. Stich, and J. Garon. Connectionism, Eliminativism, and the Future of Folk
Psychology, pp. 199–228. Lawrence Erlbaum, 1991.
[38] D. E. Rumelhart and A. Ortony. The Representation of Knowledge in Memory. Lawrence
Erlbaum Associates Inc., 1977.
[39] R. C. Schank and R. P. Abelson. Scripts, Plans, Goals and Understanding. Lawrence Erlbaum
Associates Inc., 1977.
[40] Y. P. Shimansky, T. Kang, and J. He. A novel model of motor learning capable of developing
an optimal movement control law online from scratch. Biological Cybernetics,90, 133–145,
[41] D. J. Siegel. Toward and interpersonal neurobiology of the developing mind:attachment
relationships, “mindsight” and neural integration. Infant Mental Health Journal,22, 67–94,
[42] R. S. Siegler and C. Shipley. Variation, Selection, and Cognitive Change. Developing cognitive
competence: New approaches to process modeling, T. Simon and G. Halford, eds, Chapter 2,
pp. 31–76. Erlbaum, 1995.
[43] H. A. Simon. Reasoning in Human Affairs. Standford University Press, 1983.
[44] E. Smith and D. Osherson. An Invitation to Cognitive Science: Volume 3: Thinking, 2nd edn.
MIT Press, 1995.
[45] P. Smolensky. On the proper treatment of connectionism. Behavioural and Brain Sciences,11,
1–73, 1988.
[46] P. Smolensky. Connectionism, Constituency, and the Language of Thought, pp. 201–227.
Blackwell, 1991.
[47] D. Sperber and D. Wilson. Relevance: Communication and Cogniton, 2nd edn. Blackwell,
[48] D. Sperber and D. Wilson. Fodor’s frame problem and relevance theory. Behavioral and Brain
Sciences,19, 530–532, 1996.
[49] S. Stich. The Fragmentation of Reason: Preface to a Pragmatic Theory of Cognitive Evaluation.
MIT Press, 1993.
[50] R. S. Sutton and A. G. Barto. Reinforcement Learning. The MIT Press, 1998.
[51] E. Thompson. Mind in Life: Biology, Phenomenology, and the Sciences of the Mind. Belknap
Publishing, 2007.
[52] E. Todorov and M. I. Jordan. Optimal feedback control as a theory of motor coordination.
Nature Neuroscience,16, 1226–1235, 2002.
[09:37 27/10/2009 exp067.tex] LogCom: Journal of Logic and Computation Page: 21 1–21
Relevance Realization 21
[53] A. M. Turing. Computing machinery and intelligence. Mind,59, 433–460, 1950.
[54] V. N. Vapnik. Statistical Learning Theory. John Wiley & Sons, Inc., 1998.
[55] V. N. Vapnik. An overview of statistical learning theory. IEEE Transactions on Neural Networks,
10, 988–999, 1999.
[56] F. Varela, E. T. Thompson, and E. Rosch. The Embodied Mind: Cognitive Science and Human
Experience. The MIT Press, 1991.
[57] J. Vervaeke. The Naturalistic Imperative in Cognitive Science. PhD Thesis edition, University
of Toronto, Toronto, Canada, 1997.
[58] R. Vilata and Y. Drissi. A perspective view and survey of meta-learning. Artificial Intelligence
Review,18, 77–95, 2002.
[59] M. Wheeler. Reconstructing the Cognitive World. MIT Press, 2005.
[60] D. H. Wolpert and W. G. Macready. No free lunch theorems for optimization. IEEE Transaction
on Evolutionary Computation,1, 67–82, 1997.
Received 20 June 2008
... In order to meet both of these requirements, we first position ourselves within a recent literature that promises to circumvent the frame problem in a non-representational and non-computational manner-a necessary condition for affording theoretical compatibility with the enactivist paradigm. The theoretical framework that is articulated in this literature explains that cognition realizes relevance (i.e., circumvents the frame problem) through processes of optimization involving trade-off relationships between opponent yet complementary learning strategies with respect to a mutual goal (Hovhannisyan & Dewey, 2017;Vervaeke & Ferraro, 2013;Vervaeke et al., 2012). For example, cognition can realize relevance with respect to the goal of "evading threats" by dynamically optimizing the trade-off between "fight" and "flight" in an opponent-processing fashion. ...
... However, we must note that as a cognitive theory of personality function, it must also be ensured that CB5T does not succumb to the notorious frame problem. The frame problem is the problem of explaining how cognition realizes what is relevant and is arguably a basic theoretical constraint that applies to all explanatory accounts of cognitive function (Dennett, 1987;Vervaeke et al., 2012). The immanent problem here for CB5T, therefore, is that representational and computational accounts of cognitive function have been shown to necessarily succumb to the frame problem in virtue of their very commitments (Dennett, 1987;Vervaeke et al., 2012), meaning that slipping representational or computational meanings into a cognitive scientific account of personality is to be avoided not only to ensure compatibility with enactivism but also, indeed, to evade the frame problem. ...
... The frame problem is the problem of explaining how cognition realizes what is relevant and is arguably a basic theoretical constraint that applies to all explanatory accounts of cognitive function (Dennett, 1987;Vervaeke et al., 2012). The immanent problem here for CB5T, therefore, is that representational and computational accounts of cognitive function have been shown to necessarily succumb to the frame problem in virtue of their very commitments (Dennett, 1987;Vervaeke et al., 2012), meaning that slipping representational or computational meanings into a cognitive scientific account of personality is to be avoided not only to ensure compatibility with enactivism but also, indeed, to evade the frame problem. In the following section, we discuss the importance of the frame problem for cognitive science, in general, and the (enactive) cognitive science of personality, in particular, as we review an account of cognition that promises to circumvent the frame problem in a manner that fits with the theoretical aims of our article. ...
Full-text available
The distinguishing feature of enactivist cognitive science is arguably its commitment to non-reductionism and its philosophical allegiance to first-person approaches, like phenomenology. The guiding theme of this article is that a theoretically mature enactivism is bound to be humanistic in its articulation, and only by becoming more humanistic can enactivism more fully embody the non-reductionist spirit that lay at its foundation. Our explanatory task is thus to bring forth such an articulation by advancing an enactivist theory of human personality. To this end, we synthesize core concepts from cognitive science, personality theory, and phenomenological philosophy in order to develop an Enactivist Big-5 Theory (EB5T) of personality. According to EB5T, personality traits are dispositional tendencies for how we come to optimally grip our distinctly human worlds. Individual differences in personality are therefore reflective of stylistic differences in optimal gripping tendencies between human beings. EB5T affords a non-reductionist understanding of the immanent teleology of the autopoietically embodied human mind as a kind of full-scale optimal gripping process that is achieved along five major dimensions of personality. To the degree that these dimensions are universal, therefore, we argue that our theory offers a viable path forward in advancing enactivist cognitive science beyond the life of a cell and into the mind of a person, a longstanding hope and ambition held by proponents of the enactive approach.
... referred to as the "frame problem" and cognitive scientists have long grappled with some version of this question (Dennett, 1987;McCarthy & Hayes, 1969;Vervaeke et al., 2012). In this paper we aim to show that two important frameworks for thinking about how we separate signal from noise have converged on a very similar answer, though they have used a different vocabulary and set of concepts to describe it. ...
... This particular version of the frame problem turned out to be only one manifestation of a more general problem in cognitive science. How can a cognitive agent zero in on relevant aspects of the world (i.e., signal) and intelligently ignore irrelevant aspects of the world (i.e., noise; Vervaeke et al., 2012)? Indeed, this problem is so pervasive that Vervaeke and colleagues (2012) label relevance realization as the criterion of the cognitive, suggesting that "any attempt to engineer an intelligent system must ultimately focus on the development of a system that can realize relevance" (p. 7). ...
... Relevance realization (RR) is another emerging framework in cognitive science which takes a dynamical systems perspective to explain how the mind separates signal from noise (Vervaeke et al., 2012;Vervaeke & Ferraro, 2013). RR refers to a cognitive agent's ability to intelligently ignore irrelevant information and zero in on those aspects of the world that are relevant to their goals. ...
Full-text available
The frame problem refers to the fact that organisms must be able to zero in on relevant aspects of the world and intelligently ignore the vast majority of the world that is irrelevant to their goals. In this paper we aim to point out the connection between two leading frameworks for thinking about how organisms achieve this. Predictive processing is a rapidly growing framework within cognitive science which suggests that organisms assign a high 'weight' to relevant aspects of the world, effectively treating relevant aspects of the world as if they are more precise (in a Bayesian sense). This assignment of weight is called precision-weighting and is the predictive processing account of how organisms allocate their attention. Relevance Realization is a framework that conceptualizes an organism's ability to realize relevance as resulting from a dynamical system in which a cognitive agent makes use of opponent processing relationships to zero in on relevant aspects of the world. In this paper we use recent work on the diametric model of autism and psychosis to demonstrate that the tradeoffs inherent to precision-weighting are also inherent to relevance realization. This connection will demonstrate that although these frameworks have a different intellectual background and use a different set of concepts and vocabulary, they are both pointing to the same underlying process. The fact that these different frameworks have converged on such a similar solution to the problem of how organisms realize relevance serves to demonstrate the plausibility of that solution.
... The task of intelligent systems is to identify the phenomena or concepts that are the most appropriate to describe the observed data [1,2]. While these concepts are hard to approach in tasks like image recognition or text analysis, in the case of description of observed motions of dynamical systems, we have some clues of the general form of these concepts. ...
... We determine not only the equation of motion but also other 'laws' of the systems, using different levels of embedding. A first-order embedding leads to a law of the form C (1) Δt (x n ) = constant, this includes the holonomic constraints of the system. Second-order constraints describe anholonomic constraints as well as other conserved quantities like energy. ...
Full-text available
The goal of this paper is to determine the laws of observed trajectories assuming that there is a mechanical system in the background and using these laws to continue the observed motion in a plausible way. The laws are represented by neural networks with a limited number of parameters. The training of the networks follows the Extreme Learning Machine idea. We determine laws for different levels of embedding, thus we can represent not only the equation of motion but also the symmetries of different kinds. In the recursive numerical evolution of the system, we require the fulfillment of all the observed laws, within the determined numerical precision. In this way, we can successfully reconstruct both integrable and chaotic motions, as we demonstrate in the example of the gravity pendulum and the double pendulum.
... Everything the system knows is of possible relevance. How then does the system determine what is of actual relevance without engaging in an exhaustive search of everything it knows (Dennett, 1984;Dreyfus, 1992;Fodor, 2000;Wheeler, 2005;Samuels, 2010;Vervaeke et al., 2012;Vervaeke and Ferraro, 2013;Danks, 2014;Shanahan, 2016)? The problem could perhaps be solved if artificial intelligence researchers could find a way to make an internal model that represents all possible contexts in terms of their determinate properties. ...
... For example, if the probability of a coin landing heads is 50/50 then over the course of the time spent flipping a coin, the coin will spend 50% of this time landing heads, and 50% of this time landing tails. We can think of the average 11 Our argument that agents with sensorimotor autonomy will circumvent the problem of meaning shares much in common with the account of relevance realization developed by John Vervaeke et al. in a number of publications (e.g., Vervaeke et al., 2012;Vervaeke and Ferraro, 2013). Vervaeke et al. understand relevance realization in terms of the self-organizing optimisation of trade-offs between opponent yet complementary learning strategies. ...
Full-text available
Biological agents can act in ways that express a sensitivity to context-dependent relevance. So far it has proven difficult to engineer this capacity for context-dependent sensitivity to relevance in artificial agents. We give this problem the label the “problem of meaning”. The problem of meaning could be circumvented if artificial intelligence researchers were to design agents based on the assumption of the continuity of life and mind. In this paper, we focus on the proposal made by enactive cognitive scientists to design artificial agents that possess sensorimotor autonomy—stable, self-sustaining patterns of sensorimotor interaction that can ground values, norms and goals necessary for encountering a meaningful environment. More specifically, we consider whether the Free Energy Principle (FEP) can provide formal tools for modeling sensorimotor autonomy. There is currently no consensus on how to understand the relationship between enactive cognitive science and the FEP. However, a number of recent papers have argued that the two frameworks are fundamentally incompatible. Some argue that biological systems exhibit historical path-dependent learning that is absent from systems that minimize free energy. Others have argued that a free energy minimizing system would fail to satisfy a key condition for sensorimotor agency referred to as “interactional asymmetry”. These critics question the claim we defend in this paper that the FEP can be used to formally model autonomy and adaptivity. We will argue it is too soon to conclude that the two frameworks are incompatible. There are undeniable conceptual differences between the two frameworks but in our view each has something important and necessary to offer. The FEP needs enactive cognitive science for the solution it provides to the problem of meaning. Enactive cognitive science needs the FEP to formally model the properties it argues to be constitutive of agency. Our conclusion will be that active inference models based on the FEP provides a way by which scientists can think about how to address the problems of engineering autonomy and adaptivity in artificial agents in formal terms. In the end engaging more closely with this formalism and its further developments will benefit those working within the enactive framework.
... The task of intelligent systems is to identify the phenomena or concepts that are the most appropriate to describe the observed data [1,2]. While these concepts are hard to approach in tasks like image recognition or text analysis, in the case of description of observed motions of dynamical systems, we have some clues of the general form of these concepts. ...
... We determine not only the equation of motion but also other "laws" of the systems, using different levels of embedding. A first-order embedding leads to a law of the form C (1) ∆t (x n ) = constant, this includes the holonomic constraints of the system. Second-order constraints describe anholonomic constraints as well as other conserved quantities like energy. ...
Full-text available
The goal of this paper is to determine the laws of observed trajectories assuming that there is a mechanical system in the background and using these laws to continue the observed motion in a plausible way. The laws are represented by neural networks with a limited number of parameters. The training of the networks follows the Extreme Learning Machine idea. We determine laws for different levels of embedding, thus we can represent not only the equation of motion but also the symmetries of different kinds. In the recursive numerical evolution of the system, we require the fulfillment of all the observed laws, within the determined numerical precision. In this way, we can successfully reconstruct both integrable and chaotic motions, as we demonstrate in the example of the gravity pendulum and the double pendulum.
... This is particularly important for psycho-oncology, since different cancer phases and stages mean varying affordances and salience landscapes. What information discloses itself to us as relevant is at the core of our agency [77]. This translates into an altered access to and to varied sets of ER skills and strategies depending on the cancer phase. ...
Full-text available
Emotion dysregulation is regarded as a driving mechanism for the development of mental health problems and psychopathology. The role of emotion regulation (ER) in the management of cancer distress and quality of life (QoL) has recently been recognized in psycho-oncology. The latest technological advances afford ways to assess ER, affective experiences and QoL in child, adolescent and young adult (CAYA) cancer patients through electronic patient-reported outcomes (ePRO) in their daily environment in real-time. Such tools facilitate ways to study the dynamics of affect and the flexibility of ER. However, technological advancement is not risk-free. We critically review the literature on ePRO in cancer existing models of ER in pediatric psycho-oncology and analyze strength, weaknesses, opportunities and threats of ePRO with a focus on CAYA cancer research and care. Supported by personal study-based experiences, this narrative review serves as a foundation to propose a novel methodological and metatheoretical framework based on: (a) an extended notion of ER, which includes its dynamic, adaptive and flexible nature and focuses on processes and conditions rather than fixed categorical strategies; (b) ePRO as a means to measure emotion regulation flexibility and affect dynamics; (c) identifying early warning signals for symptom change via ePRO and building forecasting models using dynamical systems theory.
... This is particularly important for psycho-oncology, since different cancer phases and stages mean varying affordances and salience landscapes. What information discloses itself to us as relevant is at the core of our agency [77]. This translates into an altered access to and to varied sets of ER skills and strategies depending on the cancer phase. ...
... Sprenger and Hartmann 2019), predictive processing is a transparently evolutionary process in which priors are 'selected' in terms of their capacity to accurately predict incoming sense data. Part of this process must also include something akin to what John Vervaeke calls 'relevance realisation' (Vervaeke, Lillicrap, and Blake 2009). There is vastly more incoming sense data than we could possibly attend to, thus our priors must also function to narrow this 'blooming, buzzing, confusion' (James 1890) down into that subset of the data that is relevant to us as agents with specific goals. ...
For an ‘evolutionary thinker’, storytelling may be considered a shared derived trait (synapomorphy) of the human lineage. Once we began to tell stories, they became a key trait shaping our subsequent evolution. Furthermore, whether the ‘centre of narrative gravity’ is cosmological (as in creation myths) or focused on an individual character, the form of narrative is itself evolutionary, describing a process of transformation from an initial situation to subsequent states (with or without ‘resolution’). This article articulates the basic logic of an ‘evolutionary stance’ and applies this heuristic to a consideration of narrative, epistemology, religion, and the status of ‘self’. In the process, insights from evolutionary biology (including the concept of ‘major evolutionary transitions’), cognitive science (including predictive processing and relevance realisation as well as computational definitions of ‘self’) and analytical psychology are drawn upon. The article ends with a consideration of practices – from meditation and active imagination to psychedelic-assisted psychotherapies – aimed at suspending the activity of the personal ‘self’ and shifting the ‘centre of narrative gravity’ to reveal transpersonal elements of the psyche. Although we inevitably resume narrativizing our existence, the experience of temporary breaks in our personal narratives may enable us to tell more inclusive stories.
... Sentence similarity is not a completely well-defined downstream task (e.g. are the sentences, 'Jack loves Jill' and 'Mary loves chocolates', similar?). 1,2 For example, Goodman (1972) argue that two objects can be infinitely similar or dissimilar (Vervaeke et al., 2012). A possible explanation for why our models perform better than prior work on search and classification but not on these tasks is that our models might not be optimized for the specific definition used by these sentence similarity benchmarks. ...
Full-text available
Text embeddings are useful features in many applications such as semantic search and computing text similarity. Previous work typically trains models customized for different use cases, varying in dataset choice, training objective and model architecture. In this work, we show that contrastive pre-training on unsupervised data at scale leads to high quality vector representations of text and code. The same unsupervised text embeddings that achieve new state-of-the-art results in linear-probe classification also display impressive semantic search capabilities and sometimes even perform competitively with fine-tuned models. On linear-probe classification accuracy averaging over 7 tasks, our best unsupervised model achieves a relative improvement of 4% and 1.8% over previous best unsupervised and supervised text embedding models respectively. The same text embeddings when evaluated on large-scale semantic search attains a relative improvement of 23.4%, 14.7%, and 10.6% over previous best unsupervised methods on MSMARCO, Natural Questions and TriviaQA benchmarks, respectively. Similarly to text embeddings, we train code embedding models on (text, code) pairs, obtaining a 20.8% relative improvement over prior best work on code search.
... 4 . 1 . s u m m a r y o f r e s u l t s Determining relevance is an essential capability of human cognition; considering all stimuli relevant all the time would be impossible (Vervaeke, Lillicrap, & Richards, 2012). Context clearly affects what is relevant and what is not, as we saw in the opening example with the sand on the beach, but in out-of-theblue contexts, we argue that people default to a more general notion of relevance that they have learned by generalising over past experiences in context. ...
Full-text available
Scalar inferences occur when a weaker statement like It’s warm is used when a stronger one like It’s hot could have been used instead, resulting in the inference that whoever produced the weaker statement believes that the stronger statement does not hold. The rate at which this inference is drawn varies across scalar words, a result termed ‘scalar diversity’. Here, we study scalar diversity in adjectival scalar words from a usage-based perspective. We introduce novel operationalisations of several previously observed predictors of scalar diversity using computational tools based on usage data, allowing us to move away from existing judgment-based methods. In addition, we show in two experiments that, above and beyond these previously observed predictors, scalar diversity is predicted in part by the relevance of the scalar inference at hand. We introduce a corpus-based measure of relevance based on the idea that scalar inferences that are more relevant are more likely to occur in scalar constructions that draw an explicit contrast between scalar words (e.g., It’s warm but not hot ). We conclude that usage has an important role to play in the establishment of common ground, a requirement for pragmatic inferencing.
Full-text available
Chiappe & Kukla argue that relevance theory fails to solve the frame problem as defined by Fodor. They are right. They are wrong, however, to take Fodor's frame problem too seriously. Fodor's concerns, on the other hand, even though they are wrongly framed, are worth addressing. We argue that relevance theory helps address them.
Two themes about the conceptual system are developed: (1) modal simulations underlie conceptual processing; (2) conceptual representations are situated. The construct of situated conceptualization—a multimodal simulation that supports one specific course of situated action with a particular category instance—integrates these themes. A given concept produces many different situated conceptualizations, each tailored to different instances in different settings. A situated conceptualization creates the experience of “being there” with a category instance in a setting via integrated simulations of objects, settings, actions, and introspections. On recognizing a familiar type of instance, an entrenched situated conceptualization associated with it becomes active, which provides relevant inferences via pattern completion. Supporting empirical evidence from cognitive psychology, social psychology, and cognitive neuroscience is reviewed.
Conference Paper
This article reviews findings from a wide range of scientific disciplines in exploring the idea that the mind develops at the interface between human relationships and the unfolding structure and function of the brain. Recent discoveries from a number of independent fields, including those of developmental psychology and cognitive neuroscience, can be synthesized into an integrated framework for understanding how the brain gives rise to mental processes and is directly shaped by interpersonal experiences. This "interpersonal neurobiology" (Siegel, 1999) presents an integrated view of how human development occurs within a social world in transaction with the functions of the brain that give rise to the mind. This framework suggests some basic principles for conceptualizing the essential experiential ingredients that may facilitate the development of the mind, emotional well-being, and psychological resilience during early childhood and perhaps throughout the lifespan. At the core of these processes is a fundamental mechanism of integration which can be seen at a variety of levels, from the interpersonal to the neurological. Integration may be conceptualized as the basic process that secure attachments facilitate in promoting psychological well-being. This article will summarize these concepts and offer some ideas about their implications for practice and future investigations.
The question, "What is Cognitive Science?" is often asked but seldom answered to anyone's satisfaction. Until now, most of the answers have come from the new breed of philosophers of mind. This book, however, is written by a distinguished psychologist and computer scientist who is well-known for his work on the conceptual foundations of cognitive science, and especially for his research on mental imagery, representation, and perception. In Computation and Cognition, Pylyshyn argues that computation must not be viewed as just a convenient metaphor for mental activity, but as a literal empirical hypothesis. Such a view must face a number of serious challenges. For example, it must address the question of "strong equivalents" of processes, and must empirically distinguish between phenomena which reveal what knowledge the organism has, phenomena which reveal properties of the biologically determined "functional architecture" of the mind. The principles and ideas Pylyshyn develops are applied to a number of contentious areas of cognitive science, including theories of vision and mental imagery. In illuminating such timely theoretical problems, he draws on insights from psychology, theoretical computer science, artificial intelligence, and psychology of mind. A Bradford Book