Opinion TRENDS in Cognitive Sciences Vol.11 No.

The proactive brain: using analogies

and associations to generate
predictions
Moshe Bar
Martinos Center for Biomedical Imaging at MGH, Harvard Medical School, 149 Thirteenth Street, Charlestown, MA 02129, USA

Rather than passively ‘waiting’ to be activated by we seek correspondence between a novel input and existing
sensations, it is proposed that the human brain is con- representations in memory (e.g. ‘what does this look like?’).
tinuously busy generating predictions that approximate Finally, these analogies activate associated representa-
the relevant future. Building on previous work, this pro- tions that translate into predictions (Figure 1).
posal posits that rudimentary information is extracted Each of these key components – associations, analogies
rapidly from the input to derive analogies linking that and predictions – has been the focus of rich and active
input with representations in memory. The linked stored research for a long time. By connecting these concepts in
representations then activate the associations that are one unifying principle of memory-based predictions, the
relevant in the specific context, which provides focused framework proposed here builds on this valuable back-
predictions. These predictions facilitate perception and ground to emphasize the functional coherence between the
cognition by pre-sensitizing relevant representations. three processes.
Predictions regarding complex information, such as those To make the underlying mechanism more explicit, I will
required in social interactions, integrate multiple analo- elaborate on each of the elements that mediate the gener-
gies. This cognitive neuroscience framework can help ation of predictions. I will start with the proposal that the
explain a variety of phenomena, ranging from recognition foundation of predictions is provided by the associative
to first impressions, and from the brain’s ‘default mode’ to nature of memory organization.
a host of mental disorders.
Associations as the building blocks of predictions
How does our experience translate into focused, testable
General framework
predictions? The answer proposed is that memory is used to
When we are immersed in the world of neuroscience
generate predictions via associative activation. In memory,
findings, the brain might seem like a collection of many
our experiences are represented in structures that cluster
little modules, each expert in a specific task. Is it possible
together related information. For example, objects that
that, instead, one can account for much of the brain’s oper-
tend to appear together are linked on some level, and these
ation using a small set of unifying principles? One such
representations include properties that are inherent to and
principle could be that the brain is proactive in that it
typical of that same experience. Such structures have been
regularly anticipates the future, a proposal that has been
termed ‘context frames’ [1,2], which are reminiscent of ear-
promoted in the past in different forms and contexts. Specifi-
lier concepts such as schemata [3], scripts [4] and frames [5],
cally, I propose that the cognitive brain relies on mem-
which all imply a unified, global representation of perceptual
ory-based predictions, and these predictions are generated
and semantic associated attributes. The structure of these
continually either based on gist information gleaned from
context frames enables co-activations that prime our sub-
the senses or driven by thought. The emphasis in this
sequent perception, cognition and action by remaining ‘on-
proposal is on the analogical link to memory and the role
line’ and making available predictions of what to expect in
of associations in predictions, as well as on the idea that we
the immediate environment. For example, placing a picture
use rudimentary information to generate these predictions
of a certain recognizable object next to an ambiguous object
efficiently. Furthermore, by developing this framework
can make that object recognizable if it looks like something
using a cognitive neuroscience approach and a minimalistic
familiar that is contextually congruent with the clear object
terminology, key concepts can directly be tested and used in
(i.e. an analogy) [1]. This principle operates similarly in other
empirical and theoretical future research.
domains. For example, contextual framing has a direct
The proposed account integrates three primary
influence on our judgments of the emotions of others [6].
components. The first is associations, which are formed
Taken together, the associative nature of memory makes
by a lifetime of extracting repeating patterns and statisti-
it possible to take advantage of frequent trends in the
cal regularities from our environment, and storing them in
environment to help interpret and anticipate immediate
memory. The second is the concept of analogies, whereby
and future events. One basis for this proposal is provided
Corresponding author: Bar, M. ([email protected]). by the literature on priming, with its various types (e.g.
Available online 4 June 2007. perceptual, semantic and contextual). These studies support
www.sciencedirect.com 1364-6613/$ – see front matter ß 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.tics.2007.05.005
 Opinion TRENDS in Cognitive Sciences Vol.11 No.7 281

Box 1. The brain’s ‘default’ mode and associative

processing
A collection of cortical regions is consistently active when human
subjects are not engaged in a goal-directed behavior. This network
has been termed the ‘default network’ [9], and is currently drawing a
large amount of attention.
The primary method of considering neuroimaging data is to
subtract the signal elicited by one condition from the signal elicited
by another condition. To look at the main effect of a single
condition – how did this condition affect activity in the brain – one
typically uses a ‘baseline’ condition in which a fixation cross is
presented. The implicit assumption in the many studies that have
used this method was that the brain uses these fixation intervals for
resting. However, increasingly more imaging studies are reporting
negative activations (‘deactivations’) when one condition is
compared with the fixation ‘baseline’. Regions that demonstrate
Figure 1. A minimalistic depiction of the proposed principle. An input (A0 ) is such deactivation are taken to have been more active during the
connected with an analogous representation in memory (A). This link results in the fixation baseline then during the compared experimental condi-
co-activation of related representations (B, C, D), associations that remain ‘online’ tion, although what are the cognitive processes that are carried out
to provide testable predictions. The input can be either an external, sensory input, by this network during rest is still unclear. This default network is
or internally generated during thought processes. Furthermore, the input can be of remarkably similar, in its medial view, to the network activated by
varying degrees of complexity, which would result in predictions that are contextual associations (Figure I) [8], supporting the proposal put
correspondingly of different levels of elaboration, encompassing the range from
forward here that the brain is continually engaged in the generation
perceptual to executive predictions.
of associations-based predictions.
the idea that the perception of a certain stimulus
co-activates the representations of related items [7], altho-
ugh these items have not been experienced as part of the
present physical environment. Indeed, recent neuroimaging
studies demonstrate the involvement of associative predic-
tions in cognitive facilitation [2].
The proposal promoted here is that the brain is
continually engaged in generating predictions, and that
these predictions rely on associative activation. Taken
together, it is important to demonstrate that the brain is
frequently busy with associative processing, an idea dating
back to Plato, Aristotle, Hobbes and the Empiricists. Sup-
port for this idea comes from a recent link we have made [8]
between the neural underpinnings of associative proces-
sing and reports regarding the cortical activation that is
considered to reflect the brain’s ‘default’ mode [9]. Specifi-
cally, we showed that significant parts of the default net-
work, which refers to the collection of brain regions that are Figure I. Medial view of the typical default network and the typical contextual
associations network. The default network regions are those that tend to show
consistently activated when subjects are not engaged in a deactivation during the experimental task. In other words, these regions are
task-specific cognitive effort, overlap with the regions acti- more active during fixation rest than during task performance. The context
vated by tasks that recruit associative processing [8] (Box network activations were obtained from the contrast between strongly
contextually associative objects (e.g. a tennis racket) and weakly contextual
1). This remarkable overlap between the default network objects (e.g. cherries). Importantly, the superimposed labels demonstrate the
and activity attributable to contextual associative proces- striking overlap between associative processes and resting, task-independent
sing demonstrates that what people do when their mental processes [8].

capacity is not completely consumed by a specific task is to

generate associations. In other words, associative acti-
vation is an integral process of the brain’s mental ‘default’ situation and to use it to derive an analogy, mapping the
mode. Given the proposal that predictions are derived from novel input to similar representations in memory. Figure 2
associations, this overlap is in agreement with the idea of a depicts a simple example, where a new exemplar of
continuous generation of predictions. a certain object class is analogically mapped to the
Associations, therefore, provide the representational corresponding prototype, and in Box 2 I describe a
tool used for predictions. In the next section I will consider model of how such an analogy can be accomplished rapidly
the mechanism that activates these associations, and how using coarse information.
it does so most effectively such that only the associations Traditionally this process has been considered as
and predictions that are most relevant and most helpful for recognition, classification, or even a type of memory retrie-
a given situation are activated. val, but in the present context I treat this process as analogy-
making instead. In the process of recognition or retrieval,
Analogies as the trigger of predictions the task is to answer the ‘what is it?’ question, whereas, in
I propose that our brains are equipped with the ability analogy, the emphasis is on ‘what is it like?’ In other words,
to extract gist, minimally analyzed information, from a although it might seem like an issue of terminology, in
www.sciencedirect.com
282 Opinion TRENDS in Cognitive Sciences Vol.11 No.7

Box 2. Top-down facilitation based on rudimentary Box 3. Questions for future research
information
 What are the computational operations and the underlying
In the framework outlined here, the activation of a memory cortical mechanisms mediating the transformation of a past
representation based on a sensory or internally generated input is memory into a future thought?
a process of analogical mapping. A central question is how gist  How does the brain handle completely novel situations where no
information, how ever defined, can be sufficient for mapping the reliable predictions can be generated?
input onto an analogous memory. One model (Figure I), from object  To what extent are we aware of our predictions and their origin? In
recognition, postulates that rudimentary information in the image some cases, such as in stereotypical thinking, being aware of
(i.e. low spatial frequencies), which is extracted rapidly, is suffi- these predictions can eliminate unwanted influences.
ciently powerful to activate expectations about what the observed  What does it mean for predictions to provide a perception of
object might be [14,75]. A similar mechanism is proposed to be stable environment? In most typical situations, we know what to
operating on multiple levels, although the representation of gist expect and what not to expect. How is finding something
information on higher levels of analysis is yet to be defined (see alarmingly incongruent with our expectations (e.g. an elephant
Concluding remarks section). Note that the gist-based initial guess in the living-room) different from finding something unexpected
could elicit more than a single alternative. This ambiguity is yet insignificant (e.g. a shoe in the living-room)?
resolved gradually as high-spatial frequencies arrive with the  How do we become aware of a mismatch between predictions
bottom-up streams. But it can also be resolved more quickly by and perception? And how do we incorporate lessons from
incorporating other rapidly extracted sources of information, such prediction errors into future behavior?
as context [2], which would fine-tune this analogical mapping to
have fewer alternatives and, thus, less ambiguity.
This principle is not limited to the realm of visual
recognition, but rather encompasses a wide variety of
domains where input can be linked to memory to generate
predictions. For example, imagine meeting a new person.
Our first impressions are rapid [15,16] and are based on
rapidly extracted coarse information [15]. According to the
present proposal, this process is mediated by linking the
features of the new person to the most similar representa-
tion in memory; someone we know and that looks to some
extent like this new person. We automatically project
information such as personality attributes to the new
person based simply on this analogy. Although this ana-
logy is an approximated set of traits, it might be beneficial,
at least under some circumstances, to not start inter-
Figure I. A top-down facilitation model. A partially processed, low spatial
frequency (LSF) image of the visual input is rapidly projected to OFC from the
actions without any assumptions on that new person.
early visual cortex and/or from subcortical structures such as the amygdala, Analogies can be based on similarity on various levels,
while detailed, slower analysis of the visual input is being performed along the including perceptual similarity (e.g. in shape or smell),
ventral visual stream. This ‘gist’ image activates predictions about candidate
objects that are similar to the image in their LSF appearance, which are fed
abstract conceptual dimensions, and goals [12]. Analogy-
back to the ventral object recognition regions to facilitate bottom-up based mappings of properties manifest themselves in
processing. Reproduced with permission from Ref. [14]. processes ranging from perception and memory [17] to
stereotypic judgments and prejudice [18].
It is important to note that the input is rarely mapped
recognition per se we recognize by linking to memory to with a single analogy directly to memory. Instead, the
interpret the input, whereas in analogy the input is linked to function of analogies can be based on the integration of
memory not only for the sake of interpretation, but also for multiple analogies that accumulate to complex mapping.
the purpose of projecting attributes and generating predic- For example, if you are trying to understand a conversation
tions. Therefore, by using the term analogy, the emphasis is that is taking place on a screen when watching a new
placed on the associations-based predictions that analogies movie, you will have to map novel sounds to similar and
elicit beyond mere recognition, and it is this extra step that is familiar sounds in memory (which will then be connected
the focus of the proposed framework. with their associated linguistic meaning), to map the novel
Nevertheless, analogical mapping still serves to interpret face appearances to similar and familiar face expressions
the input: inferring what physical input caused a certain (which will then be connected with the intentions associ-
percept, an issue that has received a lot of attention [10–12]. ated with them), the context in which the conversation is
Therefore, the analogy itself also provides an important top- taking place will be mapped to other similar contexts in
down prediction regarding the identity of the input using memory and, when combined, these analogies can help
initial bottom-up information [13,14] (Box 2). However, the map the complete, new situation to a collection of frag-
focus here is on the considerably less explored type of pre- ments in memory that together can allow you to under-
dictions: forecasting that pertains to what is about to hap- stand the scene, and to forecast what is likely to be next.
pen, what is likely to appear in the same context, and what is While our existing memories are used to derive analo-
the most beneficial action that needs to be taken given the gies and activate predictions, they are constantly being
specific input. In other words, the analogy in Figure 2 med- updated. The analogical process, in addition to affording
iates interpretation, by linking input to memory, whereas the interpretation of our environment, subsequently aug-
forecasting predictions stem from the subsequent activation ments previous representations in a way that fosters
of information associated with that analogy (e.g. Figure 3). increasingly flexible future analogies.
www.sciencedirect.com
 Opinion TRENDS in Cognitive Sciences Vol.11 No.7 283

Figure 2. A simplified mechanism for analogy whereby the analogous representation is selected based on the maximal amount of feature (F) overlap with the input. This is
a basic model, which has been used often in the past, and is introduced here to demonstrate the primary element of analogical mapping between input and memory. Such
analogies can rely on similarity on various levels (e.g. physical, conceptual and function). I treat this process here as analogy, rather than classification, to emphasize that
the main purpose of this process goes beyond mere interpretation of the input, to connect the input with known and predicted attributes that are manifested as predictions.

Finally, the activation of associations for prediction will Possible neural underpinnings
not be as useful if it simply activates automatically all the Neural substrates
information associated with the linked representation(s) in Many cortical projections that connect separate regions are
memory. Instead, it needs to take into account the context known to be reciprocal [19–22], which suggests bi-direc-
in which this input is encountered, and selectively activate tional cortical communication. According to some esti-
the most relevant associations [2]. For example, an object mates, the number of feedback (top-down) projections
such as a hairdryer can be naturally encountered in several might even exceed the number of feedforward (bottom-
possible contexts: hair-salon, appliance store, bathroom, as up) connections [23]. Although this aspect of the anatomy
well as associated with abstract contexts (Figure 3). If the is known, and the implication of omnipresent bi-directional
hairdryer is encountered in the hair-salon context, there is flow consequently seems highly reasonable, this finding
no need to activate objects that are typically found in has not yet been sufficiently incorporated into contempor-
bathrooms and appliance stores, which would not only ary thinking regarding cognitive processing. However, this
be wasteful but also generate incorrect predictions. By ‘provocative’ anatomy implies something profound about
taking context cues into account, this switchboard-like how the brain works. Specifically, the reciprocal connec-
mechanism will only activate the most relevant associ- tions might provide the infrastructure that supports the
ations, which will result in the generation of the most continuous top-down involvement of internal representa-
accurate predictions. tions with the interpretation of the world around us.
To summarize, analogies map novel inputs to There are three main components in this proposal:
representations in memory that most resemble this input. associations, analogies and predictions, and they interact
Subsequently, information associated with these repres- with each other regularly. Associations have largely been
entations is activated to provide predictions about what found in the medial temporal lobe (MTL), in the hippo-
else might be expected in the same situation. By taking campus [24,25], and in the parahippocampal cortex (PHC)
context into account in this associative activation, only the [26]. As reviewed above with regard to the overlap seen
most relevant predictions are generated. In the next sec- between associative processing and the default network,
tion I will describe possible neural underpinnings for other medial regions, such as the medial prefrontal cortex
associations, analogies and predictions. (MPFC) and medial parietal cortex (MPC) are involved as
www.sciencedirect.com
284 Opinion TRENDS in Cognitive Sciences Vol.11 No.7

Figure 3. Translating analogies to predictive associative frames. The linked stored representation (i.e. an analogy mapped as depicted in Figure 2) activates only the
associations that are relevant in the specific context, which provides focused predictions. In the example depicted here, although a hairdryer is associated with many
objects, and multiple contexts (e.g. bath, hair-salon and appliance store), only the appropriate subset of predictions will be activated based on the associations relevant in
the specific context (e.g. hair-salon). These connections result in predictions that are then tested against the input (sensory, or internally generated).

well. Given the diverse types of possible associations, it is and the PHC in particular. The retrosplenial complex (RSC
indeed expected that they will be mediated by a large in the MPC), which is consistently found to be recruited in
collection of regions, depending on complexity, modality associative tasks (Figure I in Box 1), has been suggested to
and purpose. For example, other types of associations, such represent prototypical, rather than episodic, information
as visuomotor associations, seem to be represented in other about associations [26]. In other words, the PHC represents
regions, such the basal ganglia. stimulus-specific context and associations, which are sensi-
The brain regions that mediate analogical thinking are tive to specific appearance (e.g. my kitchen), whereas the
much less explored. Nevertheless, some types of analogical RSC/MPC represents knowledge about associations related
thinking have been found to activate the lateral and medial to the prototypical context (e.g. a kitchen). Their combined
PFC [27,28]. contribution presumably elicits prediction-related repres-
Regarding the neural regions that mediate predictions, entations in the PFC (orbitofrontal cortex, OFC, in particu-
there are multiple sub-processes that need to be considered: lar), as well as in a domain-specific cortex such as the
the generation of predictions, their verification, and their fusiform gyrus in the case of object recognition.
updating. Expectations-based preparatory activation has
been observed in numerous domains. For example, anticip- Neural mechanisms
ating a somatosensory stimulus activates the somatosen- The neural mechanism that generates predictions is
sory cortex [29]; pictures of food activate gustatory cortices largely unknown, but it is thought to be mediated, or at
[30]; visual imagery, even if not in a directly predictive task, least balanced, by neural oscillations and synchrony [32],
activates the visual cortex, and has even been shown to and some evidence for such a mechanism in the PFC exists
activate early visual cortex in a retinotopically organized [33]. However, in addition to synchrony, there are other
manner [31]. promising proposals [10,34,35] that are appealing and could
We have previously proposed a neural mechanism in the benefit from further physiological and cognitive testing.
domain of context-based predictions and visual recognition Predictions are proposed to be activated rapidly, using
[2,14,36]. Briefly, the associations relevant to the present information that is available relatively early (i.e. low spatial
discussion seem to be represented by the MTL in general, frequencies; LSF). Such LSF-based predictions can be
www.sciencedirect.com
 Opinion TRENDS in Cognitive Sciences Vol.11 No.7 285

triggered in the PFC for the purpose of object recognition same prefrontal regions that are proposed to be involved in
while interacting with the visual occipito-temporal cortex predictions, OFC and the ACC in MPFC (e.g. Ref. [8]), have
[14], and for context-related predictions while interacting been theorized to monitor exploration and exploitation
with MTL and MPC [2]. The generation of object-related behavior (via the locus coeruleus) [50]. The MPFC, in
predictions in the PFC by rudimentary information is ela- particular, has been proposed to be crucial for cognitive
borated in Ref. [14], and how LSF rudimentary information control and specifically for evaluation of potential out-
might activate context-related predictions is elaborated in comes (see Ref. [51] for review). How is this related to
Refs [2,37,38]. the present attribution of a predictive function to regions of
Bayesian analysis is a natural computational the MPFC? Processes such as evaluation of performance
formulation that can be used for capturing the idea that and outcomes, coping with response errors, response con-
experiences accumulated in our memory guide our predic- flicts, and decision uncertainty all have in common an
tions and behavior. Every moment of our lives involves extensive reliance on predictions. This can further help
some level of uncertainty, and associative memories can explain why the same MPFC area is also activated by a
help cope with this uncertainty. The statistical history of diverse range of additional processes, including error
events and stimuli in our surroundings guides the acti- detection [51], contextual activations [2], and ‘reading’
vation of what is the most appropriate information to somebody else’s mind (i.e. theory-of-mind) [52].
expect. Such statistical regularities can be extracted from Additional factors related to the neural substrates that
our environment [39,40]. We frequently use the Bayesian subserve the proposed framework will be interspersed in
principle that past experiences shape the priors for pre- upcoming sections, but in general, given the amount of
dicting aspects of the environment in everyday life, and available reports, it is hard to say substantially more on
there are sound reasons to suggest that the brain operates this theme without being overly speculative. Among the
according to these principles [10,41,42]. relevant issues that deserve extensive future research (Box
3), of particular interest are the spatiotemporal inter-
Detecting errors in predictions actions among the various components of this network.
Humans can detect and learn from regularities in the
environment, even without awareness [40]. When the pre- Applied predictions
dictions derived from these learned associations are vio- Predictions, as described here, span a wide spectrum of
lated, people can nevertheless detect them [43]. The complexity and function. At one extreme, there is the
neuronal and computational mechanisms that subserve simple prediction that stems from a simple association
the detection of prediction violation have been considered (e.g. knowing to expect pain after bumping your toe on
in the past [44,45]. The fMRI signal that corresponds to a the leg of the sofa). At the other extreme, one can anticipate
detection of a violation of expectations seems to concentrate complicated experiences, plan far ahead, or mentally
on the lateral ventral prefrontal cortex [46]. This suggests ‘travel’ in time to the future, based on simulations and
that the mechanisms that detect and potentially correct memory. In spite of this extreme variety, there is no
mismatches are mediated by at least partially different evidence to indicate that predictions of various levels of
regions than those that participate in the medial network complexity, abstraction, timescale and purpose use mech-
that is proposed here to generate predictions. Various other anisms that are qualitatively different. As suggested ear-
cortical regions have been implicated in processes related to lier, simple perceptual predictions can be a result of direct
prediction errors. For example, activity in the anterior analogical mapping followed by the automatic activation of
cingulate cortex (ACC) has been reported to correlate with a Hebbian-like association. By contrast, complex executive
the likelihood of prediction error [47]. Prediction errors are predictions result from the integration of multiple smaller
detected for the purpose of adjusting the specific prediction analogies and ‘atomic’ associations.
for the immediate context, as well as for learning and Our accumulated experience creates an ever-evolving
updating internal generalizations that will help improve platform for predictions, and the influence of these predic-
predictions in the future. Such errors can be detected intern- tions can be observed on many levels. In this section I will
ally, via mismatches between predictions and perceptions, briefly review findings from such various domains that all
and they can be detected using external cues, such as reward demonstrate the ubiquitous role of predictions in the brain
and punishment. Our brains seem wired to detect change and in behavior, as well as the ways in which predictions
efficiently, at various levels. It is possible that the compari- rely on associations and analogies.
son between predictions and perception is processed outside The first is a phenomenon termed ‘representational
of awareness and we become aware of it only when a mean- momentum’ [53]. This term pertains to the finding that
ingful mismatch is encountered. a sequence of static photographs that implies a certain type
In learning and developing the ever-evolving of motion (e.g. imagine a rectangle that in each consequent
knowledge-base from which predictions are later gener- image is rotated a few degrees clockwise) significantly
ated, we frequently face a choice between relying on what affects the subjects’ performance on the last, target image,
we know (i.e. exploitation) or attending novel aspects that as a function of its rotation difference from the previous
can augment our existing knowledge (i.e. exploration). This image in the sequence. This robust demonstration suggests
is particularly pronounced in conditions of uncertainty that subjects generate the representation of the anticip-
[48,49], which in the present framework might reflect ated next image before it is presented. A related fMRI
reduced ability to rely on analogies when such analogies study [54] showed that when human subjects view static
cannot be mapped with high confidence. Interestingly, the images with implied motion (e.g. an athlete about to jump),
www.sciencedirect.com
286 Opinion TRENDS in Cognitive Sciences Vol.11 No.7

the motion-processing areas in the middle temporal and Predictions from within: mind-wandering and
the middle posterior superior temporal sulcus are acti- mental time travel
vated although the images are static. The focus here has been on predictions triggered by and
Demonstrations of contextual priming [55,56] provide geared towards the sensory world around us. But thinking
another example of the generation of visual predictions, about the future is often oriented internally, as we do when
where seeing a certain object activates the representa- we plan, imagine, reason, fantasize, and so on. There has
tions of other objects that are likely to appear in the same been a recent surge of interest in the subject, which
environment. Furthermore, predictions have a clear and includes stimulating findings from neuroimaging and
crucial role in language comprehension and sentence patient work [62–65]. This ability to project oneself into
integration [57,76], and being able to predict the inten- the future and imagine upcoming or imaginary situations
tions and actions of others is beneficial, and has recently can be seen as a prediction tool, which also relies on
been shown to confer direct facilitation on sensory proces- analogies and associative processing, and whose primary
sing [58]. Finally, studies have shown that activity in goal is preparatory in nature. The areas involved seem to
task-specific regions during anticipatory preparation is be highly overlapping with the associative and the default
predictive of subsequent performance success [59]. These networks (Box 1).
findings indicate that by generating a prediction, the brain An interesting aspect common to those findings is that
activates the specific regions that are responsible for systems that are typically associated with memory, prim-
processing the type of information that is about to be arily in MTL, are consistently active when people think
encountered. about the future. This does not change the function that
Louis Pasteur said that ‘chance favors the prepared these memory regions have been implicated with, as some
mind’. According to the proposal described here, the ‘pre- have suggested, but it does underscore the fact that to
paredness’ of the mind is derived from the generation of consider the future we need to retrieve information from
specific predictions, which shapes the pre-sensitization of our memory of the past (i.e. analogies and associations in
the representations of some information and not another. the terminology used here). This makes sense, given the
Therefore, what we see can be driven by our motivations, supreme importance of our acquired memory for our future
goals and desires, which all might bias attention and survival and interaction with the environment. In that
sometimes determine our perception [60]. Therefore, by regard, it might be overly broad to look at the involvement
enabling privileged processing of anticipated information, of memory structures in internal mental processes as link-
regardless of the source of this anticipation, predictions ing memory and imagination [66], because the specific part
might be seen as constraining the selection of what aspects of imagination that recruits memory is, according to the
of the environment will be primed and perceived, and how proposal described here, the generation of predictions. Con-
they will be interpreted. sequently, the cardinal purpose of memory starts to seem
Predictions have ubiquitous clinical implications. On less for leisured reminiscing, as in the famous example of
the debilitating end, predictions and associations can pro- Proust’s madeleine, and more as a knowledge-base that
vide the basis for anxiety disorders and phobias, as well as guides our lives in an increasingly more informed manner.
post-traumatic stress disorders. But predictions in other It is important to consider the potential functional
clinical circumstances can result in a positive influence, role of mental time travel. My proposal is that its primary
such as in placebo effects. Indeed, placebo analgesia that is role is to create new ‘memories’. We simulate, plan and
enhanced by expectations has been shown to activate combine past and future in our thoughts, and the result
regions in the PFC, such as the OFC [61], which are might be ‘written’ in memory for future use. These simu-
proposed here to be part of the network that generates lated memories are different from real memories in that
predictions. they have not happened in reality, but both real and
Beyond helping us anticipate, prepare and simulate, the simulated memories could be helpful later in the future
continual generation of predictions has another purpose, by providing approximated scripts for thought and action.
which is to help us to perceive stability and coherence in This could be why athletes often report that they find
our environment. Our senses deliver a profoundly impo- imagining a future race helpful in the race itself. Therefore,
verished version of the environment to us, due to factors simulating the future guides us by providing a rudimen-
such as the massive reduction in the amount of information tary script for the optimal course of action when the
reaching the visual cortex from the retina, the presence of anticipated future arrives.
clutter and noise, and the fact that even familiar items vary
in their appearance from one occasion to another. So pre- Concluding remarks and predictions about
dictions, derived from memory, fill-in the gaps based on predictions
experience. I propose that the knowledge that is stored in our memory
This reconstructive contribution of predictions can be exerts its contribution to behavior by way of predictions, and
helpful not only when dealing with impoverished sensory that our perception of the environment relies on existing
input, but could eliminate the need to attend every aspect knowledge as much as it does on incoming sensory infor-
of our environment exhaustively, and could also be the mation. The framework proposed here is composed of three
reason why our brain can often ignore details and encode in main ingredients: associations, analogies and predictions. It
memory only a reduced, gist-based version of actual mem- can be summarized as follows: the input is linked to memory
ories because these details can later be reconstructed with via analogies, and once an analogy has been found, it elicits
sufficient resolution. the activation of the representations associated with the
www.sciencedirect.com
 Opinion TRENDS in Cognitive Sciences Vol.11 No.7 287

input. This co-activation of related items provides on-line, ‘wrong path’. Examples range from a baseball illusion
focused predictions on what else is highly likely to be called the rising fastball [71], to smell judgments that vary
relevant in that specific situation. Regardless of whether significantly based on how that smell was labeled [72], to
oriented internally and externally, the primary role of these demonstrations of false-memory in social cognition [73].
predictions is to guide our actions, plans and thoughts. In addition, there are other phenomena that might be
As discussed in the introduction, several of these explained as interference from predictions, although they
concepts have been studied individually in the past. How- have not been portrayed as such. For example, that cog-
ever, the present proposal integrates these concepts in a nitive performance can be improved when subjects are
single unifying framework, with a cognitive neuroscience instructed to be passive and not to ‘try hard’ [74]; when
perspective. The synthesis proposed here might neverthe- we are unable to retrieve a memory and the same incorrect
less be seen as related to notions such as ‘embodied cogni- memory keeps coming up; or when a break from a persist-
tion’ [67,68]. Embodied cognition is a movement involving ing intellectual problem seems to be all that was needed to
primarily philosophy and artificial intelligence, and it is then find a quick solution upon returning to the problem.
related to situated cognition [69]. The sense of similarity The common component in these cognitive phenomena
between the frameworks stems from the following elements might be top-down predictions that are persistent but
of embodied cognition: that the brain ‘has an agenda’ and, unhelpful and we have to find ways to ignore them
thus, that it is proactive, as proposed here too; that it takes altogether in such cases.
context (i.e. the situation) into account; and that memory is The independent role of predictions, analogies and
embodied (e.g. not amodal) to promote interactions with the associations in our mental lives hardly requires further
world. However, other claims diverge significantly from demonstrations, given the rich and fruitful research of
what is promoted here. For example, the idea that the these concepts. However, their integration and orchestra-
function of cognitive processes is directed solely towards tion here under a single framework could nevertheless
the body’s interactions with the world, the suggestion that benefit from direct testing. One hypothesis that stems from
the mind cannot be studied in isolation, and that the outside the current proposal is that because predictions are
world is part of the cognitive system and, therefore, that derived from analogies, increased similarity between novel
representations in memory have a lesser role than the one input and representations in memory should result in
endorsed here. In other words, embodied cognition links increasingly efficient predictions. This could be tested on
even the most abstract processes to sensorimotor low-level multiple levels of predictions, ranging from perceptual to
processes, whereas the framework proposed here relies on more executive predictions. Another interesting test will
mental simulations, foresight and various types of anticip- be of the notion that analogies and predictions are gener-
ation that cannot always be directly mapped to low-level ated based on gist. Although this has been specified and
actions. tested in fields such as recognition (e.g. LSF-based; Box 2),
Humans seem to minimize processing of incoming it becomes more challenging when one considers higher
information when this information is predictable (e.g. habi- levels of abstraction. It seems reasonable that gist will be
tuation, repetition blindness, change blindness, inhibition of helpful in abstract analogies as well, and characterizing
return), and in parallel encourage the allocation of mental how abstract gist (such as that often also used in the
resources to unexpected and/or novel information (e.g. memory literature) is extracted, represented and used will
orienting response toward novel and unexpected stimuli, be necessary.
perceptual pop-out). How does this somewhat circular mech- Given the omnipresence of predictions, their influence
anism work? After all, we should process the input (i.e. find pervades far more processes than can be treated here.
an existing analogy) to understand whether there is some- These include predictions in emotion, their effects on
thing unexpected there or not. The bias to not invest in the awareness and their interactions with attentional allo-
predictable is proposed to be primarily a top-down, intern- cation. These are major issues, and addressing them will
ally driven process, whereas the alertness for detecting the be crucial for any complete theory of the human brain.
novel and unexpected is primarily a bottom-up, sensory- In regard to clinical implications, aging seems to
driven process [70]. One possibility is that the decision is compromise many of the regions that are proposed here
made based on the evaluation of a match between perception to constitute the network of associations-based predictions:
and prediction using only gist information that is available MTL, MPC and MPFC. Therefore, older adults should have
early (Box 3). a diminishing ability to predict future events. However, at
An interesting direction for future research concerns the the same time, the observation that older adults can use
roles of the different types of memory in predictions. I have their increased experience to derive improved expectations
concentrated here on episodic memory, the central role of is common. Gaining a better understanding of which pre-
which is supported by the repeated activation of the MTL dictions are affected by atrophy, and of what regions, will
in prediction-related processing, but it is reasonable to help chart the global network of predictions as well as
hypothesize that semantic memory plays a role as well. characterizing any qualitative differences between differ-
Studying how these memory types interact, as might be ent types of predictions, both of which will ultimately
reflected by interactions between the medial network dis- improve future cognitive approaches to aging-related
cussed here and, for example, the dorso-lateral PFC will be decline. Furthermore, several clinical disorders, such as
crucial for such understanding. major depression, are associated with a reduced ability to
In spite of the advantage that predictions provide for take context into account during everyday judgments
behavior, they can sometimes interfere and put us on the and decisions. This could be seen as intuitive when
www.sciencedirect.com
288 Opinion TRENDS in Cognitive Sciences Vol.11 No.7

considering the debilitating tendency of depressed patients 19 Pandya, D.N. (1995) Anatomy of the auditory cortex. Rev. Neurol.
(Paris) 151, 486–494
to ruminate on a (usually negative) thought. As elaborated
20 Ghashghaei, H.T. et al. (2007) Sequence of information processing for
here, context consists of associations, and associations emotions based on the anatomic dialogue between prefrontal cortex
provide the basis of predictions. It is possible that engaging and amygdala. Neuroimage 34, 905–923
such patients in cognitive programs that train them to 21 Rockland, K.S. and Drash, G.W. (1996) Collateralized divergent
activate predictions that rely on associations from a wide feedback connections that target multiple cortical areas. J. Comp.
Neurol. 373, 529–548
variety of contexts would help alleviate their symptoms.
22 Ergenzinger, E.R. et al. (1998) Cortically induced thalamic plasticity in
Such optimistic prospects will have to be tested explicitly the primate somatosensory system. Nat. Neurosci. 1, 226–229
in the future. 23 Salin, P.A. and Bullier, J. (1995) Corticocortical connections in the
Finally, according to the proposal outlined here, the visual system: structure and function. Physiol. Rev. 75, 107–154
brain constantly anticipates the future, and to do this it 24 Eichenbaum, H. (2000) A cortical-hippocampal system for declarative
memory. Nat. Rev. Neurosci. 1, 41–50
constantly accesses information in memory. This implies 25 Ranganath, C. et al. (2004) Inferior temporal, prefrontal, and
that our subjective feeling of present is actually an integ- hippocampal contributions to visual working memory maintenance
ration that also includes past and future information: we and associative memory retrieval. J. Neurosci. 24, 3917–3925
need to know where we were (both in time and space) and 26 Aminoff, E. et al. (2006) The parahippocampal cortex mediates spatial
what is next, in addition to our present state. Indeed, and nonspatial associations. Cereb. Cortex, DOI: 10.1093/cercor/bhl078
27 Bunge, S.A. et al. (2005) Analogical reasoning and prefrontal cortex:
people who study meditation initially find it extremely evidence for separable retrieval and integration mechanisms. Cereb.
difficult to eliminate any future and past thoughts. Un- Cortex 15, 239–249
derstanding how the brain integrates and holds simul- 28 Waltz, J.A. et al. (2000) The role of working memory in analogical
taneously information from multiple points in time is an mapping. Mem. Cognit. 28, 1205–1212
29 Carlsson, K. et al. (2000) Tickling expectations: neural processing in
important question for the future.
anticipation of a sensory stimulus. J. Cogn. Neurosci. 12, 691–703
30 Simmons, W.K. et al. (2005) Pictures of appetizing foods activate
Acknowledgements gustatory cortices for taste and reward. Cereb. Cortex 15, 1602–1608
I thank E. Aminoff, J. Boshyan, M. Fenske, N. Gronau, K. Kveraga, M. 31 Slotnick, S.D. et al. (2005) Visual mental imagery induces retinotopically
Mason and three anonymous reviewers for their help. Supported by organized activation of early visual areas. Cereb. Cortex 15, 1570–1583
NINDS R01-NS044319 and NS050615, and the MIND Institute. 32 Engel, A.K. et al. (2001) Dynamic predictions: oscillations and
synchrony in top-down processing. Nat. Rev. Neurosci. 2, 704–716
References 33 Liang, H. et al. (2002) Synchronized activity in prefrontal cortex during
1 Bar, M. and Ullman, S. (1996) Spatial context in recognition. Perception anticipation of visuomotor processing. Neuroreport 13, 2011–2015
25, 343–352 34 Ullman, S. (1995) Sequence seeking and counter streams: a
2 Bar, M. (2004) Visual objects in context. Nat. Rev. Neurosci. 5, 617– computational model for bidirectional information flow in the visual
629 cortex. Cereb. Cortex 1, 1–11
3 Mandler, J.M. and Johnson, N.S. (1976) Some of the thousand words a 35 Mumford, D. (1992) On the computational architecture of the neocortex.
picture is worth. J. Exp. Psychol. [Hum Learn] 2, 529–540 II. The role of cortico-cortical loops. Biol. Cybern. 66, 241–251
4 Schank, R.C. (1975) Using knowledge to understand. In Theoretical 36 Bar, M. and Aminoff, E. (2003) Cortical analysis of visual context.
Issues in Natural Language Processing (Schank, R.C. and Nash-Weber, Neuron 38, 347–358
B., eds), pp. 117–121, Tinlap Press 37 Oliva, A. and Torralba, A. (2001) Modeling the shape of a scene: a
5 Minsky, M. (1975) A framework for representing knowledge. In The holistic representation of the spatial envelope. Vision. Res. 42, 145–175
Psychology of Computer Vision (Winston, P.H., ed.), pp. 211–277, 38 Sanocki, T. (1993) Time course of object identification: evidence for a
McGraw-Hill global-to-local contingency. J. Exp. Psychol. Hum. Percept. Perform. 19,
6 Mobbs, D. et al. (2006) The Kuleshov Effect: the influence of contextual 878–898
framing on emotional attributions. Soc. Cogn. Affect. Neurosci 1, 95– 39 Torralba, A. and Oliva, A. (2003) Statistics of natural image categories.
106 Network 14, 391–412
7 Anderson, J.R. (1983) Architecture of Cognition, Harvard University 40 Chun, M.M. and Jiang, Y. (1998) Contextual cueing: implicit learning
Press and memory of visual context guides spatial attention. Cognit. Psychol.
8 Bar, M. et al. (2007) The units of thought. Hippocampus 17, 420–428 36, 28–71
9 Raichle, M.E. et al. (2001) A default mode of brain function. Proc. Natl. 41 Kersten, D. et al. (2004) Object perception as Bayesian inference. Annu.
Acad. Sci. U. S. A. 98, 676–682 Rev. Psychol. 55, 271–304
10 Friston, K. (2005) A theory of cortical responses. Philos. Trans. R. Soc. 42 Knill, D.C. and Pouget, A. (2004) The Bayesian brain: the role of
Lond. B Biol. Sci. 360, 815–836 uncertainty in neural coding and computation. Trends Neurosci. 27,
11 French, R.M. (2002) The computational modeling of analogy-making. 712–719
Trends Cogn. Sci. 6, 200–205 43 Biederman, I. et al. (1982) Scene perception: detecting and judging
12 Hummel, J.E. and Holyoak, K.J. (2003) A symbolic-connectionist objects undergoing relational violations. Cognit. Psychol. 14, 143–177
theory of relational inference and generalization. Psychol. Rev. 110, 44 Schultz, W. and Dickinson, A. (2000) Neuronal coding of prediction
220–264 errors. Annu. Rev. Neurosci. 23, 473–500
13 Rao, R.P. and Ballard, D.H. (1999) Predictive coding in the visual 45 Friston, K. (2003) Learning and inference in the brain. Neural Netw.
cortex: a functional interpretation of some extra-classical receptive- 16, 1325–1352
field effects. Nat. Neurosci. 2, 79–87 46 Nobre, A.C. et al. (1999) Orbitofrontal cortex is activated during breaches
14 Bar, M. et al. (2006) Top-down facilitation of visual recognition. of expectation in tasks of visual attention. Nat. Neurosci. 2, 11–12
Proc. Natl. Acad. Sci. U. S. A. 103, 449–454 47 Brown, J.W. and Braver, T.S. (2005) Learned predictions of error
15 Bar, M. et al. (2006) Very first impressions. Emotion 6, 269–278 likelihood in the anterior cingulate cortex. Science 307, 1118–1121
16 Willis, J. and Todorov, A. (2006) First impressions: making up your 48 Cohen, J.D. and Aston-Jones, G. (2005) Cognitive neuroscience:
mind after a 100-ms exposure to a face. Psychol. Sci. 17, 592–598 decision amid uncertainty. Nature 436, 471–472
17 Song, J.H. and Jiang, Y. (2005) Connecting the past with the present: 49 Daw, N.D. et al. (2006) Cortical substrates for exploratory decisions in
how do humans match an incoming visual display with visual memory? humans. Nature 441, 876–879
J. Vis. 5, 322–330 50 Aston-Jones, G. and Cohen, J.D. (2005) An integrative theory of locus
18 Devine, P.G. (1989) Stereotypes and prejudices: their automatic and coeruleus-norepinephrine function: adaptive gain and optimal
controlled components. J. Pers. Soc. Psychol. 56, 5–18 performance. Annu. Rev. Neurosci. 28, 403–450

www.sciencedirect.com
 Opinion TRENDS in Cognitive Sciences Vol.11 No.7 289

51 Ridderinkhof, K.R. et al. (2004) Neurocognitive mechanisms of 63 Addis, D.R. et al. (2007) Remembering the past and imagining the future:
cognitive control: the role of prefrontal cortex in action selection, common and distinct neural substrates during event construction and
response inhibition, performance monitoring, and reward-based elaboration. Neuropsychologia 45, 1363–1377
learning. Brain Cogn. 56, 129–140 64 Szpunar, K.K. et al. (2007) Neural substrates of envisioning the future.
52 Frith, C.D. and Frith, U. (2006) The neural basis of mentalizing. Proc. Natl. Acad. Sci. U. S. A. 104, 642–647
Neuron 50, 531–534 65 Buckner, R.L. and Carroll, D.C. (2007) Self-projection and the brain.
53 Freyd, J.J. (1983) The mental representation of movement when static Trends Cogn. Sci. 11, 49–57
stimuli are viewed. Percept. Psychophys. 33, 575–581 66 Miller, G. (2007) Neurobiology. A surprising connection between
54 Kourtzi, Z. and Kanwisher, N. (2000) Activation in human MT/ memory and imagination. Science 315, 312
MST by static images with implied motion. J. Cogn. Neurosci. 12, 67 Lakoff, G. and Johnson, M. (1980) Metaphores We Live By, University of
48–55 Chicago Press
55 Palmer, S.E. (1975) The effects of contextual scenes on the 68 Noe, A. (2005) Action in Perception (Representation and Mind), The
identification of objects. Mem. Cognit. 3, 519–526 MIT Press
56 Biederman, I. (1972) Perceiving real-world scenes. Science 177, 77–80 69 Barsalou, L.W. (2003) Situated simulation in the human conceptual
57 Duffy, S.A. (1986) Role of expectations in sentence integration. J. Exp. system. Lang. Cogn. Process. 18, 513–562
Psychol. Learn. Mem. Cogn. 12, 208–219 70 Grossberg, S. (1980) How does a brain build a cognitive code? Psychol.
58 Neri, P. et al. (2006) Meaningful interactions can enhance visual Rev. 87, 1–51
discrimination of human agents. Nat. Neurosci. 9, 1186–1192 71 McBeath, M.K. (1990) The rising fastball: baseball’s impossible pitch.
59 Wylie, G.R. et al. (2006) Jumping the gun: is effective preparation Perception 19, 545–552
contingent upon anticipatory activation in task-relevant neural 72 de Araujo, I.E. et al. (2005) Cognitive modulation of olfactory
circuitry? Cereb. Cortex 16, 394–404 processing. Neuron 46, 671–679
60 Balcetis, E. and Dunning, D. (2006) See what you want to see: 73 Macrae, C.N. et al. (2002) Creating memory illusions: expectancy-based
motivational influences on visual perception. J. Pers. Soc. Psychol. processing and the generation of false memories. Memory 10, 63–80
91, 612–625 74 Smilek, D. et al. (2006) Relax! Cognitive strategy influences visual
61 Kong, J. et al. (2006) Brain activity associated with expectancy- search. Vis. Cogn. 14, 543–564
enhanced placebo analgesia as measured by functional magnetic 75 Bar, M. (2003) A cortical mechanism for triggering top-down
resonance imaging. J. Neurosci. 26, 381–388 facilitation in visual object recognition. J. Cogn. Neurosci. 15, 600–609
62 Hassabis, D. et al. (2007) Patients with hippocampal amnesia cannot 76 DeLong, K.A. et al. (2005) Probabilistic word pre-activation during
imagine new experiences. Proc. Natl. Acad. Sci. U. S. A. 104, 1726– language comprehension inferred from electrical brain activity. Nat.
1731 Neurosci. 8, 1117–1121

Have you contributed to an Elsevier publication?

Did you know that you are entitled to a 30% discount on
books?
A 30% discount is available to all Elsevier book and journal contributors when ordering books or
stand-alone CD-ROMs directly from us.

To take advantage of your discount:

1. Choose your book(s) from www.elsevier.com or www.books.elsevier.com

2. Place your order

Americas:
Phone: +1 800 782 4927 for US customers
Phone: +1 800 460 3110 for Canada, South and Central America customers
Fax: +1 314 453 4898
[email protected]

All other countries:

Phone: +44 (0)1865 474 010
Fax: +44 (0)1865 474 011
[email protected]

You’ll need to provide the name of the Elsevier book or journal to which you have
contributed. Shipping is free on prepaid orders within the US.

If you are faxing your order, please enclose a copy of this page.

3. Make your payment

This discount is only available on prepaid orders. Please note that this offer does not
apply to multi-volume reference works or Elsevier Health Sciences products.

For more information, visit www.books.elsevier.com

www.sciencedirect.com