Balancing Perception and Cognition in an Attention

Balancing Perception and Cognition in an Attention-Centric Cognitive System
Anon Y. Mous
Abstract
That flexible, intelligent, goal-directed behavior in a dynamic world requires heuristic computation has been
widely appreciated since the early days of artificial intelligence and cognitive science. Ultimately, all heuristics share the property of being “selective” with respect
to a space of options. In (biological) cognitive systems,
attention is often billed as the mechanism that manifests this selectivity. Attention is thought to be a part
of a larger cluster of mechanisms that serve to orient a
cognitive system, to filter contents with respect to their
task relevance, and to devote more computation to certain options than to others. All these activities proceed
under the plausible assumption that not all information
can be or ought to be processed for a system to satisfice
in an ever changing world. In this paper, we describe
an attention-centric cognitive system called ARCADIA
that demonstrates the orienting, filtering, and resourceskewing functions mentioned above. The demonstration involves maintaining focus on a simple cognitive
task (i.e., counting certain perceptual events) in a dynamic environment. While the system successfully carries out the task, limits on attentional capacity result
in“inattentional blindness” under circumstances analogous to those where people show similar behavior.
Introduction
Flexible, intelligent behavior in a dynamic world requires
cognitive systems that strike a balance between reactivity
and high-level cognition. In the biological realm, selective
attention is often invoked as the means by which this balance may be struck. Attention is under our willful control
when we concentrate on a task at hand and can be captured
by highly salient stimuli like the sound of our name being
shouted by a friend. The interplay between focused attention and attentional capture is complex. When people are
involved in a task, the presence of too many highly salient,
task-irrelevant stimuli leads to degraded performance. Conversely, too much concentration on a particular thought or
task may lead to failures in noticing what would otherwise
be a highly salient (and perhaps important) stimulus.
To illustrate these failures, consider the case where a
driver is texting and fails to see the police car parked a few
meters ahead, recall trying to understand your friend over the
phone while standing in a noisy room, and reflect on being
“in the zone” while programming or writing and forgetting
to eat. As these examples show, success in the world involves
being able to strike an effective balance between being reactive and being task focused. Perhaps striking that balance involves reminding oneself to ignore the mobile phone while
driving or to set an alarm while coding as a way to enforce
breaks. Both of these strategies help control attention in situations where, possibly, one has experienced a failure in the
past and seeks to avoid them in the future. This paper describes a cognitive system that balances perception and cognition using strategies of this sort. The system is predicated
on the idea that attentional strategies, both learned and innate, serve to coordinate perception and cognition.
Given attention’s central role throughout human cognition, the limited treatment in the cognitive systems literature is disconcerting. To be sure, researchers within this field
have addressed certain aspects of attention, perception, and
cognition, but rarely has their intersection been explored.
One one side of a coin, the vast majority of computational
work on vision and especially on object recognition is massively implausible from a psychological perspective.1 The
copious amount of labeled training data and brute-force image processing strategies employed in those efforts are at
odds with both common sense and the deliverances of vision science. On the other side of the coin, psychologically
plausible approaches to perception in some of the betterknown cognitive architectures such as ACT-R (Nyamsuren
and Taatgen 2013) and EPIC (Kieras 2010) rely on symbolic
encodings of objects, relations, and events. There seems to
be an empty space between these two families of research
approaches: a space for systems that operate over real sensor data, but do so in psychologically plausible ways. An
attention-centric cognitive system may both exhibit perceptual selectivity in image processing and provide a theorygrounded platform for exploring the perception–cognition
interface.
Minding the Gap
The plan for the rest of the paper is as follows. First, we introduce an attention-centric cognitive system called ARCADIA, which stands for Adaptive, Reflective Cognition in an
Attention-Driven Integrated Architecture. To demonstrate
how the architecture can combine top-down (i.e., cognitive)
and bottom-up (i.e., perceptual) factors, we describe work to
date on a visual pipeline in ARCADIA. Bridewell and Bello
(2015) previously applied this pipeline to a change detection task and provided details on the how these factors determine where the system focuses its attention. Importantly,
1
We have no particular opinion about the utility of high fidelity,
psychologically plausible models in creating artificial, cognitive
systems. In our view, psychologically implausible approaches may
well be on the right track and the comments we make here should
not be taken as impugning those approaches.
Internal Identifier
ID
Name
Arguments
World
Source
Type
The element's variadic relation
Map from a relation's argument names to objects
Named collection of elements that belong
to a distinct perspective
Component that produced the element
Class of the element's relation
Figure 2: A schema for encoding an element in ARCADIA’s
interlingua.
Figure 1: ARCADIA’s tripartite informational structure
(top), and its cognitive cycle (bottom).
under certain, psychologically plausible conditions the architecture exhibited change blindness which is a surprising,
human insensitivity to dynamism in the world (Simons and
Rensink 2005).
After providing a walkthrough of ARCADIA’s visual
pipeline in the change-detection task, we report on ARCADIA’s ability to carry out the cognitive task of counting the
number of times that a moving ball changes color. This time
we find that under certain conditions ARCADIA exhibits
inattentional blindness, a psychological phenomenon where
people fail to recognize otherwise salient stimuli when their
attentional capacities are taxed. Here, a stimulus engineered
to provide a strong bottom-up saliency signal is ignored as
the system pursues its counting task. Importantly, when ARCADIA is not tasked with counting, its attention moves between the ball on the screen and the other salient stimulus
that was previously ignored due to limits on attention.
Finally, we explore trade-offs at the perception–cognition
interface; especially the idea that one could increase attentional capacity in artificial cognitive systems and hope to
mitigate the unwanted side-effects of these limitations.
The ARCADIA Cognitive System
ARCADIA embodies a theoretical commitment to attention
as a central, integrating mechanism for perception, cognition, and action. The system operates in distinct cognitive
cycles that are guided by a focus of attention. In terms
of intellectual roots, ARCADIA shares much of the structure found in the Global Workspace Theory of consciousness (Baars 1997). This connection should come as no surprise: the relationship between attention, perception, and
consciousness has generated an enormous amount of literature over the years and continues to be a fruitful area of research (Koch and Tsuchiya 2012; Dehaene et al. 2006). As
shown in the top of figure 1, ARCADIA’s general structure is
tripartite, consisting of components, accessible content, and
the aforementioned focus of attention, each of which will be
described in detail as our discussion proceeds.
Note that ARCADIA makes a principled distinction between aspects of cognition that are and are not able to be
introspected or verbally reported.2 Figure 1 illustrates this
distinction with uninspectable processing occurring at the
bottom in the activity of the system’s components, and potentially inspectable, reportable representations and activity
(more accurately, representations of activity) contained in
the middle and upper parts of the pyramid.
Representation in ARCADIA
Representation in ARCADIA is multifarious and differs between its inspectable and uninspectable layers. Components,
which are computational modules, encapsulate low-level
processes. ARCADIA places no theoretical restrictions on
the representational format or processing characteristics for
components in general, although one may develop components that reflect theoretical positions. This freedom at the
system level enables flexibility in design. Components can
employ any sort of data structures and algorithms, including
but not limited to neural networks, production rules, graphical models, frames, or constraint graphs, letting modelers
use whatever solution best fits the task at hand and trade it
for a better one when available. Similarly, the data processed
within components is equally unconstrained. Multidimensional matrices for images, waveforms for sounds, symbolic
structures for propositional and relationally structured data
are all admissible.
Interlingua ARCADIA is an integrated architecture, so
components with disparate representational formats must
communicate with each other. To this end, each component
interfaces with accessible content and the current focus of
attention through an interlingua. Figure 2 provides a schema
for an element expressed in the interlingua. Each element
consists of a unique identifier, a variadic argument-list, and
a symbolic name for the collection of arguments. The arguments contain labeled data produced by the components
and stored in formats that they can process. As a result,
the interlingua can bind visual, auditory, and other sensory
data to abstract, symbolically represented content retrieved
from long-term memory structures. Moreover, each element
tracks which component produced it by way of a source tag
2
The conscious–unconscious distinction does not quite work
here. Most memories are unconscious because they are inactive and
are not poised to be used in inference. Inactivity does not entail an
inability to become conscious or otherwise usable in inference.
and has type (e.g., action-description or object-instance). Finally, interlingua elements are indexed to worlds, which describe the situation that each element refers to. For instance,
elements describing aspects of ongoing perception are assigned the world “reality” whereas interlingua elements corresponding to mental simulations of fictional, hypothetical,
or counterfactual worlds would have a world argument corresponding to the name for the respective simulation.
Components and Inference
Although not shown explicitly in figure 1, there are three distinct classes of components in ARCADIA. The first kind of
component polls sensor data. When a sensory input channel is established for ARCADIA (e.g., a video, a soundscape), this class of components samples the input channel and passes the sample to the second class of components: those that operate directly on sensory data. For the
case of video processing, ARCADIA’s video-polling component samples a frame in the form of a matrix of image data
and hands it to one or more components that operate directly
on sensory data (examples to follow). After these low-level
(i.e., close to sensory input) components complete their processing, they encode their results in interlingua elements and
deliver them to ARCADIA’s middle layer where they may
be used by other components. Finally, all other components
implement routines that scan ARCADIA’s middle layer for
the particular type of contents they are specialized to process, extract data from the arguments in the elements, carry
out relevant processing over the extracted data, and encode
results into new interlingua elements that are pushed up into
the middle layer.
Accessible Content and the Focus of Attention
The top of figure 1 indicates that ARCADIA maintains a
separate space called accessible content. This area stores
ephemeral representations produced each cycle by low-level
components. In here, the system makes available the contents of working memory, the results of perceptual processing, and other potentially reportable information. Similar to
Baars’ concept of the global workspace, accessible content
is substantially larger than working memory, and we take as
an assumption, subject to revision, that it corresponds to the
informational contents of consciousness.3 Elements in this
space result from the attentional process that drives ARCADIA’s cognitive cycle and are produced by components as
they respond to the focus of attention or sensory input. Notably, the theoretical relationship between accessible content
and consciousness implies that verbal report is limited to the
items that accessible content contains.
As we have stated, ARCADIA maintains a focus of attention. Operationally, the focus of attention is a privileged
element of accessible content that the system broadcasts to
1. Change Event
2. Action Request
3. New Objects
4. Fixation Request
5. Candidate Fixation
6. Random Element
Task Specific Components
Task General Components
Figure 3: Attentional strategy for change detection.
the components,4 represented by the downward arrow in
figure 1. As a result, the focus has the power to alter the
direction of processing within the system, enabling shifts
from top-down, deliberative control to bottom-up, sensory
responsiveness and back. Note that the focus itself does no
inferential or perceptual work, rather it influences behavior
in the system indirectly. Only those (non-sensory) components responsive to the information in the focus will be active during a cycle, and they will therefore produce the focus
of attention and accessible content for the next cycle.
The Cognitive Cycle and Focus Selection
The bottom of figure 1 illustrates ARCADIA’s cognitive cycle. On each iteration, low-level components receive the current focus of attention and may query accessible content produced during the previous iteration. Designated pre-attentive
components connect directly to sensory systems such as
cameras, while others operate over only the focus and accessible content. Components automatically engage in processing and run to completion if they find input that they can respond to. The architecture deposits interlingua elements that
are produced by components into what will become the next
cycle’s accessible content and that may become the next focus of attention for the system. The structure, function, and
properties of the elements vary. Some components produce
bound object or event representations, whereas others produce abstract output, such as an expectation, whose arguments refer to other interlingua elements.
The only place where processing occurs outside of components is during focus selection. The focus selector is an
architectural mechanism that uses an attentional strategy to
choose the next focus of attention (see figure 3). Such strategies serve as control knowledge for how to pursue a task. In
principle, attentional strategies can be a function of arbitrary
complexity, but we currently use priority lists that prefer one
set of elements over others. For instance, one may specify
that the system should consider focusing on an available action request before focusing on a representation of an object.
In the next section, we describe an instantiation of a visual pipeline in ARCADIA and trace through its computations as the system constructs object representations. This
process occurs over multiple cycles, begins with raw image
data, and typically produces an element that is stored in visual short-term memory. In parallel, the system continues to
3
If one accepts the distinction between access consciousness
and phenomenal consciousness as championed by Ned Block
(1995), ARCADIA’s middle layer roughly corresponds to the former. Accessible content comprises elements poised for use in inference or verbal report.
4
Even though the focus is limited to a single element, the interlingua allows substantial internal structure in the argument list and
the focus is supported by the contextual information in accessible
content.
vSTM
Identity
Round, Red,
Circle
Focus of Attention
(Partial Illustration)
Currently Active
Attentional Strategy
!!!
Round, Red,
Circle
NEW
!!!
•
In region X:
Round, Red,
Circle...
Bottom-Up
Feature
Computation
Iconic
Memory
Bottom-Up
Fixation
Generation
Top-Down
Fixation
Generation
vSTM
Early Binder
Round,
Red, Circle
Early Binder
Property
Detectors
Figure 4: Processing during a model of vision in ARCADIA.
process sensor data, and the resulting elements may capture
attention, thereby disrupting the construction of an object
representation.
•
Change Detection in ARCADIA
The initial version of ARCADIA’s visual pipeline was designed for object localization, identification, and tracking,
but with minor extensions, it can also detect changes in one
or more properties of an object. To this end, the attentional
strategy reflects priorities both for constructing objects from
visual data and for the task of detecting (visual) changes in
objects in the world. ARCADIA differs from most cognitive
architectures in that it operates over pixel-level, video data.
Further, the system takes multiple cycles to construct object
representations from low-level features (e.g., closed contours, color profiles, “retinal” location). Once constructed,
the representations must be tagged as referring either to new
object instances or to previously seen objects. Because objects take multiple cycles to encode into representations, interruptions in the normal object-construction process may
result in errors that are characteristic of human perception,
which we discuss in the next section.
As shown in figure 4, ARCADIA’s visual pipeline is
served by a variety of components. Space precludes discussing the computational details of each component, however we give succinct descriptions below.
• Bottom-Up Feature Computation polls a camera component for the current frame of visual input. Feature computation is carried out by an implementation of the Itti-Koch
approach to visual saliency calculations (Itti, Koch, and
Niebur 1998). The output of this component is an interlingua element that has a visual saliency map and the point
of maximum bottom-up saliency in its argument list.
• Iconic Memory also takes frames directly from a camera
component. These frames are segmented to extract closed
contour regions from the image along with location and
hue-saturation-value information. The output of iconic
memory is an interlingua element that has a list of protoobjects (Rensink 2000), one for each detected closed contour, with associated location and hue-saturation-value information.
• Bottom Up Fixation Generation produces requests for visual foveation guided by attention capture. This com-
•
•
•
ponent scans accessible content for proto-objects and
saliency maps and attempts to find the maximally salient
proto-object. The resulting region containing that protoobject is encoded as a fixation request and passed to accessible content.
Top Down Fixation Generation also produces foveation
requests, but it is guided by the task at hand. In this case,
the task is change detection, which plausibly involves reexamining objects for alterations to their properties. This
component scans accessible content for elements of visual
short term memory (vSTM) and requests fixations at the
last-reported location of one of these (currently selected at
random). As with the bottom-up component, the resulting
region is encoded and passed to accessible content.
Early Binder responds to a fixation in the focus of attention. This component takes the proto-object target of
that fixation and stores it for one extra cycle. Simultaneously, various property detection components (classifiers for color, shape, etc.) produce classifications that are
delivered to accessible content. In the next cycle, early
binder scans accessible content for these classifications,
gathers the positive instances and creates a new interlingua element that represents the object, populating the argument list with the positive classifications. The new object representation is passed to accessible content.
Identity responds to an object representation in the focus
of attention and determines whether the object was previously encountered. This task is carried out by scanning
accessible content for other objects that were previously
encoded item in vSTM. For this particular implementation, the identity component checks to see if the new object is the same size and at a spatial location within a small
threshold with respect to the vSTM items in accessible
content. If so, the identity component produces an interlingua element that marks the object as being an update,
and if not, marks it as new.
Visual Short Term Memory scans accessible content on
each cycle for interlingua elements produced by the identity component. When it encounters an interlingua element marked as new, it inserts the element into its list.
Updates are handled by finding the closest existing element in its list, and updating property values. Internally,
vSTM is capacity-limited at six objects. On each cycle,
vSTM outputs each of its object representations into accessible content.
Change Detection (not pictured) is a task specific component that looks for color differences. This component
examines elements in accessible content for identity relationships between old and new objects (along size and
location dimensions) that differ in color and produces an
element that stores the old and new state of the object,
representing the detection when a discrepancy is found.
Other components in ARCADIA include a change reporter
that spots changes in the focus of attention and outputs
the discovery, a working memory that stores change events
among other things, and modules for moving cameras and
monitoring fixations.
Attention and Failures to Perceive
The hybrid attentional strategy for change detection and object construction is shown in in figure 3, and duly prioritizes
reporting changes when they appear in accessible content. If
no changes are found, then the focus is given to new objects.
If there are no changes or new objects, ARCADIA will focus on (accepted and processed) fixation requests, then candidate fixations, and finally on a random element. Because
certain components involved in early vision operate independently from the focus of attention, fixation candidates are
generated in parallel alongside feature-binding and encoding. While this is usually an unproblematic sequence, certain forms of perceptual interruptions can occur, as they do
in change blindness.
Change Blindness in ARCADIA
Change blindness is a perceptual phenomenon wherein large
visual changes occurring in full view go unnoticed (Simons
and Rensink 2005). In a typical change blindness experiment, subjects view an initial image followed by a fleeting disruption in visual continuity: an blank screen, localized blotches superimposed on the image, or short film-cuts
(on the order of tens of milliseconds), during which time
the image is modified in some way. When visual continuity resumes, subjects can only rarely identify where the
change occurred. Researchers conjecture that disruptions occurring across the whole of the visual field nullify the ability of bottom-up saliency calculations to guide attention to
the region where the alteration occurs. For a detailed walkthrough of change blindness in ARCADIA, see the work by
Bridewell and Bello (2015).
Inattentional Blindness in ARCADIA
Whereas change blindness embodies a failure of bottom-up
attentional capture, a top-down, strategic focusing of attention may also lead to failures to perceive. Inattentional blindness is one particularly well known instance of the latter phenomenon that results from capacity limitations on the focus
of attention. Research on inattentional blindness suggests
that people consciously experience only those objects and
events to which they directly attend (Mack and Rock 1998).
As such, failures to attend entail failures to consciously experience, and corresponding failures to verbally report. Because attentional focus is widely considered to be extremely
limited, strategic focus on one particular task or aspect of
the environment leaves the perceiver open to missing much
of what is going on around them. In a poignant demonstration, Hyman and colleagues found that subjects involved in a
conversation on their mobile phones while walking through
a park failed to notice a unicycling clown along their route
(Hyman et al. 2010). Notably, inattentional blindness is not
an all-or-none phenomenon. The degree to which perceivers
are unaware of a particular stimulus depends on a variety of
factors: most notably its visual similarity to the current focus
and its relative distance from the object or event on which
the focus of attention is currently trained (Most et al. 2001;
Newby and Rock 1998).
We predict that because ARCADIA is equipped with a
limited focus of attention, it too should fall prey to inattentional blindness when strategically focused on a particular
object or task. To this end, we designed a video stimulus
comprising a single ball that moves on a white background
and occasionally changes direction. While moving, the ball
also changes from blue to green multiple times. ARCADIA’s
task is to count the number of times that the ball changes
color. During the video, a highly salient stimulus, a bright
red star, zips across the bottom of the display. Because ARCADIA’s attention is trained solely on the ball for purposes
of tracking and counting color changes, we predict that attention will never be captured by the red star, even though
it may be pre-attentively processed by certain components
(bottom-up feature detection, iconic memory, etc.). Accordingly, we predict that removal of the counting task will result
in ARCADIA’s attention flickering back and forth between
the star and the ball.
To be clear, ARCADIA does not have explicit goals or
established task structures at this time. Modeling a task in
the system involves implementing components that provide
the required functionality and an attentional strategy that can
produce the desired behavior. For counting color changes,
we start with the model of change detection that we described earlier in the paper. To this, we add two components:
working memory and a change counter. The attentional strategy resembles the one in figure 3 with an added distinction.
We discuss these efforts before discussing the resulting system behavior.
Added Components: Counting and Remembering
• Working Memory provides storage for information that is
actively used by the system, operating like a cache. This
component responds to memorization requests, which
have type “action,” in the focus of attention. These may
take the form of “remember this new element” or as updates of the form “replace an earlier version of this element with this new one.” As with vSTM, working memory makes its elements available as accessible content.
• Change Counter responds to a change event in the focus of attention. To this end, the component looks for a
number associated with counting changes in accessible
content. If there is no such number, the change counter
produces a request to memorize the number one, with
the properties that it is a “counter” and that it updates
on a “change.” Otherwise, if a number exists with those
properties, then the component requests a memory update
action that replaces the old number with its incremented
value. As pointed out when discussing working memory,
these requests must receive the focus of attention to be
processed in ARCADIA’s cycle.
Attentional Strategy
An attentional strategy for this task requires the system to
preferentially attend to an object, detect color changes, and
increment a counter accordingly. Because memory operations have type “action,” the additional components combined with the attentional strategy in figure 3 already enable
Figure 5: ARCADIA visualizations while (left) untasked and (right) tasked with counting changes. The top left window in
each image shows objects stored in vSTM. The bottom left window shows an iconic representation of ARCADIA’s view where
objects are picked out. Red boxes indicate where working memory representations were last updated, the pink box indicates
the current visual fovea, green boxes are candidate fixations. The window to the right is the bottom-up saliency map. In both
instances, the star is more salient than the circle.
the system to detect changes and increment a counter. However, if a salient distractor impinges on ARCADIA’s visual
field, will be drawn away from the object that it is tracking,
potentially missing changes. To indicate the importance of
the task, we make one minor modification to the previous
strategy.
Recall that the change detection strategy involved, at level
5, a task-general preference for a candidate fixation. When
the strategy uses this preference to select the next focus of
attention, the system chooses a candidate fixation randomly
from those that are available. In that case, ARCADIA may
be drawn away from a task-relevant object to foveate on a
salient distractor. To prevent this behavior, we modify the
attentional strategy to prefer fixating on objects represented
in vSTM over those regions that are suggested by bottom-up
feature computations. To restate, we use the attentional strategy to encode that it is more important to continue looking at
the task-relevant object (i.e., the moving circle) than to look
toward anything else (i.e., the red star).
Resulting Behavior
As illustrated in figure 5, ARCADIA behaves as predicted.
The image on the left, which was produced using the attentional strategy in figure 3, shows ARCADIA’s focus
squarely on the distractor and two objects in vSTM. In contrast, the image on the right, which used the modified attentional strategy, shows that ARCADIA visually attends only
to the circle throughout the scenario. Importantly, the associated saliency maps (where brighter pixels represent greater
saliency) indicate that the red star coincides with the most
salient region in both cases. Notice in both cases that the
distractor is processed by low-level visual components, but
never attains an object representation due explicitly to a lack
of attention.
Future Directions and Concluding Remarks
These results are only the first steps in building a general,
taskable cognitive-system that perceives, deliberates about,
and acts within the world. To this end we are implementing an auditory saliency map (Kayser et al. 2005), taking the
first step toward interpreting audio signals within ARCADIA with an eye toward multi-sensory integration. We also
intend to develop a theory of long-term memory and attentive recollection where memories are multi-representational
in nature, containing perceptual properties (Morsella et al.
2009). Further, recollection will be modulated by internally
directed attention (De Brigard 2012). At the higher-level
end of cognitive activity, we will be implementing models
of deliberative reasoning informed by mental model theory
(Khemlani, Barbey, and Johnson-Laird 2014).
Readers of this paper may find themselves mystified about
why anyone would build such a capacity-limited system and
then portray the expression of those limitations as markers
of success. We address the first part of the question by making the uncontroversial claim that regardless of the available computational resources, there will always be too much
data to process everything in full. This is true whether we
are talking about autonomous systems operating within the
world or cognitive aids that assist people in making sense
from mounds of data. To handle this problem, people design
systems with resource limitations in mind, hard coding their
expectations about what must or must not be processed. In
contrast, we suggest a better solution: develop a computational model of attention. A cognitive system with an attentional mechanism makes explicit the processing trade-offs
that are sensible for specific tasks.
To address the second part of the question, we claim that if
one treats attention as a general solution to manage computational resources, then the trade-offs observed in other systems with attentional mechanisms (e.g., humans) provide informative markers for any implementation. Here, we demonstrated that an emphasis on task completion produces a sort
of tunnel vision that blocks out otherwise salient stimuli. We
suggest that this finding applies to any system that can control its focus of attention. That is, part of what it means to
attend to one thing is to ignore everything else.
References
Baars, B. J. 1997. In the theatre of consciousness. Global
Workspace Theory, a rigorous scientific theory of consciousness. Journal of Consciousness Studies 4:292–309.
Block, N. 1995. On a confusion about a function of consciousness. Behavioral and Brain Sciences 18:227–247.
Bridewell, W., and Bello, P. F. 2015. Incremental object perception in an attention-driven cognitive architecture. In Proceedings of the Thirty-Seventh Annual Meeting of the Cognitive Science Society, 279–284.
De Brigard, F. 2012. The role of attention in conscious
recollection. Frontiers in Psychology 3:1–10.
Dehaene, S.; Changeux, J. P.; Naccache, L.; Sackur, J.; and
Sergent, C. 2006. Conscious, preconscious, and subliminal
processing: a testable taxonomy. Trends in Cognitive Sciences 10:204–211.
Hyman, I. E.; Boss, S. M.; Wise, B. M.; McKenzie, K. E.;
and Caggiano, J. M. 2010. Did you see the unicycling
clown? Inattentional blindness while walking and talking on
a cell phone. Applied Cognitive Psychology 24:597–607.
Itti, L.; Koch, C.; and Niebur, E. 1998. A model of saliencybased visual attention for rapid scene analysis. IEEE
Transactions on Pattern Analysis and Machine Intelligence
20:1254–1259.
Kayser, C.; Petkov, C. I.; Lippert, M.; and Logothetis, K.
2005. Mechanisms for allocating auditory attention: an auditory saliency map. Current Biology 15:1943–1947.
Khemlani, S.; Barbey, A.; and Johnson-Laird, P. N. 2014.
Causal reasoning with mental models. Frontiers in Human
Neuroscience 8:1–15.
Kieras, D. 2010. Modeling visual search of displays of many
objects: the role of differential acuity and fixation memory. In Proceedings of the Tenth International Conference
on Cognitive Modeling, 127–132.
Koch, C., and Tsuchiya, N. 2012. Attention and consciousness: related yet different. Trends in Cognitive Sciences
16:103–105.
Mack, A., and Rock, I. 1998. Inattentional Blindness. Cambridge, MA: MIT Press.
Morsella, E.; Lanska, M.; Berger, C. C.; and Gazzaley, A.
2009. Indirect cognitive control through top-down activation
of perceptual symbols. European Journal of Social Psychology 39:1173–1177.
Most, S. B.; Simons, D. J.; Scholl, B. J.; Jimenez, R.; Clifford, E.; and Chabris, C. F. 2001. How not to be seen: the
contribution of similarity and selective ignoring to sustained
inattentional blindness. Psychological science 12:9–17.
Newby, E. A., and Rock, I. 1998. Inattentional blindness as
a function of proximity to the focus of attention. Perception
27:1025–1040.
Nyamsuren, E., and Taatgen, N. A. 2013. Pre-attentive and
attentive vision module. Cognitive Systems Research 24:62–
71.
Rensink, R. A. 2000. Seeing, sensing, and scrutinizing.
Vision Research 40:1469–1487.
Simons, D. J., and Rensink, R. A. 2005. Change blindness:
past, present, and future. Trends in Cognitive Sciences 9:16–
20.