Conference PaperPDF Available

System-1 and System-2 realized within the Common Model of Cognition

Authors:

Abstract

Attempts to import dual-system descriptions of System-1 and System-2 into AI have been hindered by a lack of clarity over their distinction. We address this and other issues by situating System-1 and System-2 within the Common Model of Cognition. Results show that what are thought to be distinctive characteristics of System-1 and 2 instead form a spectrum of cognitive properties. The Common Model provides a comprehensive vision of the computational units involved in System-1 and System-2, their underlying mechanisms, and the implications for learning, metacognition, and emotion.
System-1 and System-2 realized within the Common Model
of Cognition
Brendan Conway-Smith 1*, Robert L. West1
1Department of Cognitive Science, Carleton University, 1125 Colonel By Dr, Ottawa, ON, Canada
Abstract
Attempts to import dual-system descriptions of System-1 and System-2 into AI have been
hindered by a lack of clarity over their distinction. We address this and other issues by
situating System-1 and System-2 within the Common Model of Cognition. Results show that
what are thought to be distinctive characteristics of System-1 and 2 instead form a spectrum
of cognitive properties. The Common Model provides a comprehensive vision of the
computational units involved in System-1 and System-2, their underlying mechanisms, and
the implications for learning, metacognition, and emotion.
Keywords
1
System 1, System 2, dual-process, cognitive modeling, common model, cognitive architecture
1.0. Introduction
Significant progress has been made in Artificial Intelligence (AI) by studying and attempting to copy
human cognition. An important aspect of human intelligence is that humans tend to choose different
types of cognitive processes to match the demands of a situation. The concept of System-1 and System-
2 captures this idea by positing a dual-systems model where System-1 provides fast, heuristic-based
thinking, and System-2 allows for slower, more rational thought. This maps onto folk psychological
notions of rationality that contrast deliberate rational thought with fast impulsive thinking.
However, while the System-1 and System-2 dichotomy has provided a meaningful way to study
different styles of thought, there is no agreed upon computational framework that integrates these
two modes of thinking into a unified view of cognition. In this paper, we argue that the Common
Model of Cognition, originally the Standard Model [1] provides a unified framework for
understanding System-1 and System-2. We will also show how recent criticisms of System-1 and
System-2 can be clarified using the Common Model of Cognition to ground discussion.
2.0. System-1 and System-2
The terms System-1 and System-2 refer to a dual-system model that ascribes distinct characteristics
to what are thought to be opposing aspects of cognition. Dual-systems theories (or dual-process
theories) are common in psychology and posit that cognitive processes can be divided into two
contrasting categories. A variety of theories have arisen due to different proposed ways of
characterizing the division [2, 3, 4]. However, roughly speaking, dual-systems theories have divided
thinking between the intuitive and the rational [5].
System-1 is considered to be evolutionarily old and characterized as fast, associative, emotional,
automatic, and not requiring working memory. System-2 is considered to be more evolutionarily
recent and characterized as slow, declarative, rational, effortful, and relying on working memory.
System-1 and System-2 are often used in fields such as psychology, philosophy, neuroscience, and
AAAI 2022 FALL SYMPOSIUM SERIES, Thinking Fast and Slow and Other Cognitive Theories in AI, November 17-19, Westin Arlington
Gateway in Arlington, Virginia, USA
brendan.conwaysmith@carleton.ca (B. Conway-Smith); robert.west@carleton.ca (R. L. West)
0000-0003-3936-6125 (B. Conway-Smith)
© 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISS N 1 6 1 3 - 0 0 7 3
artificial intelligence as a means for ontologizing the functional properties of human cognition. The
neural correlates of System-1 and System-2 have also been examined [6].
Recently, however, this dual-system model has been criticized for lacking precision and conceptual
clarity [7], leading to significant misconceptions [8] [9], and obscuring the dynamic complexities of
psychological processes [10]. Much of this criticism stems from controversy over the alignment
assumption. The alignment assumption refers to the claim that cognitive functions must align with
either System-1 or System-2 [9]. From an AI perspective, the alignment assumption would be
convenient, however, this assumption has been criticized as overly simplistic and some dual-systems
theorists do not endorse it, referring instead to “typical correlates” rather than “defining features”[8].
Researchers in need of greater specificity have developed more detailed definitions of Systems-1
and 2. For example, Proust [11] has argued that a more precise computational definition is needed to
understand the role of System-1 and 2 in metacognition (the use of higher level, or meta-level,
processes to control cognition. Proust defined these systems in terms of their distinctive informational
typologies, where System-1 metacognition is implicit, non-symbolic, and non-conceptual, while
System-2 metacognition is explicit, symbolic, and conceptual.
Computational models of System-1 and System-2 have also been built. For example, Thomson et al.
[12] argued that the expert use of heuristics (System-1) could be defined in terms of instance-based
learning in ACT-R. In fact, there are numerous ways that cognitive models and cognitive architectures
can and have been mapped onto the System-1 and 2 distinction. In some cases, the dual-process
approach has been built directly into the architecture. For example, the CLARION architecture [13]
and the LIDA architecture [14] have instantiated components that map directly onto characteristics
of System-1 and System-2 type thinking.
However, while it is useful to work on modelling different aspects of System-1 and 2, the larger
question is, in what sense is System-1 and System-2 a valid construct? What are the necessary and
sufficient conditions that precisely define System-1 and 2? And what are the cognitive and neural
alignments to System-1 and System-2?
2.1. Cognitive architectures
Evans [15], one of the originators of dual-system theory, has stated that an important issue for future
research is the problem that “current theories are framed in general terms and are yet to be developed
in terms of their specific computational architecture.”
Following Dennett [16], we argue that a computational description is essential for clarifying high-
level, psychological characterizations, such as System-1 and System-2. At the time, Dennett received
significant pushback on his view from psychology and philosophy, however, in our opinion, this was
due to it being too early in the development of cognitive models to fully appreciate their value.
As Newell [17] noted, creating one-off computational models of individual psychological
phenomena is also problematic, as it leads to a plethora of isolated “micro models”, unconstrained by
considerations of forming an integrated agent architecture. Newell’s solution to this was the concept
of a unified computational cognitive architecture. A cognitive architecture is a computationally
implemented agent architecture that models the components of cognition and how they interact to
produce psychological phenomena.
2.2. The Common Model
The Common Model of Cognition, originally the Standard Model[1], is a consensus architecture
that integrates decades of research on how human cognition functions computationally. The
Common Model represents a convergence across cognitive architectures regarding the modules and
components necessary for human-like intelligence. That is, the Common Model is a higher-level
architectural specification that describes most, if not all, cognitive architectures capable of modeling
the full range of human cognition. The Common Model has also been investigated by way of
correlating its modules with their associated brain regions [18]. Neural imaging across a range of
tasks strongly supports the Common Model as a leading candidate for modeling the functional
organization of the human brain [19].
The Common Model has five components working memory, perception, action, declarative
memory, and procedural memory. Procedural memory is integral to how the Common Model
operates, as it reacts and provides instruction to other modules based on working memory content.
The simplest way of building procedural memory is as a production system acting on production
rules (if-then rules). This is how it is implemented in ACT-R, yet not in other Common Model
Architectures such as SOAR or Sigma. Since a production system is an easy way to describe and
captures the essential action that we are interested in for this paper, we will describe procedural
memory as a production system and procedural knowledge as productions. However, it is important
to keep in mind that this is for communication purposes and that a production system represents the
simplest possible mechanism for implementing procedural memory in the Common Model.
In terms of processing, the Common Model can be driven in two different ways. The first is
through task specific productions. These are fast and automatic, in the sense that the system does not
consider extra information - it knows how to proceed. The second way is through information stored
in declarative memory. In this case, the necessary productions are not available, so information is
requested from declarative memory, which is then used by productions to advance. While this maps
onto System-1 and Sytem-2, as we will see, the picture is more complex.
3.0. System-1
Researchers generally describe System-1 using a constellation of characteristics. Specifically, System-
1 is described as fast, associative, emotional, automatic, and not requiring working memory [20, 21,
22]. System-1 is considered to be evolutionary old and present within animals. It is composed of
biologically programmed instinctive behaviors and operations that contain innate modules of the kind
put forth by Fodor [23]. System-1 is not comprised of a single system but is an assembly of sub-
systems that are largely autonomous [24]. Automatic operations are usually described as involving
minimal or no effort, and without a sense of voluntary control [20]. Researchers generally agree that
System-1 is made of parallel and autonomous subsystems that output only their final product into
consciousness (often as affect), which then influences human decision-making [15]. This is one reason
the system has been called intuitive[25].
System-1 relies on automatic processes and shortcuts called heuristics - problem solving operations
or rule of thumb strategies [26]. The nature of System-1 is often portrayed as non-symbolic, and has
been associated with reinforcement learning [27] and neural networks [28]. Affect is integral to
System-1 processes [29]. Affect based heuristics result from an individual evaluating a stimulus based
on their likes and dislikes. In more complex decision-making, it occurs when a choice is either weighed
as net positive (with more benefits than costs), or as net negative (less benefits than costs) [30].
System-1 can produce what are called cognitive illusions” that can be harmful if left unchecked.
For example, the "illusion of validity” is a cognitive bias where individuals overestimate their ability
to accurately predict a data set, particularly when it shows a consistent pattern [31]. This is closely
related to feelings-of-knowing that can provide both accurate and inaccurate epistemic signals [32].
Biases and errors within System-1 operate automatically and cannot be turned off at will. However,
they can be offset by using System-2 to monitor System-1 and correct it.
3.1. System-1 in the Common Model
System-1 can be associated with the production system, which is the computational instantiation of
procedural memory in the Common Model [33]. Procedural knowledge is represented as production
rules (“productions”) which are modeled after computer program instructions in the form of
condition-action pairings. They specify a condition that, when met, will perform a prescribed action.
A production can also be thought of as an if-then rule. If it matches a condition, then it fires an action.
Productions transform information to resolve problems or complete a task, and are responsible for
state-changes within the system. Production rules fire automatically off of conditions in working
memory [34]. They are considered to be automatic due to the fact that they are triggered without
secondary evaluation. Neurologically, production rules correlate with the 50ms decision timing in the
basal ganglia [35].
The Common Model production system has many of the properties associated with System-1 such
as being fast, automatic, implicit, able to implement heuristics, and reinforcement learning. However,
the Common Model declarative memory system also has some of the properties associated with
System-1. Specifically, associative learning and the ability to implement heuristics that leverage
associative learning [12]. Here, it is important to understand that the Common Model declarative
memory cannot operate without the appropriate productions firing, and without the use of working
memory. Therefore, from a Common Model perspective, System-1 minimally involves productions
firing based on working memory conditions. However, it can also involve productions directing
declarative memory retrieval, which also relies on working memory.
Based on this, System-1 cannot be defined as being uniquely aligned with either declarative or
procedural memory. System-1 activity must necessarily involve production rules and working
memory, and can also include declarative knowledge. A partial exception to this may be the direct
links between perception and action. It is possible, in the CMC, to have responses that are directly
triggered by perception. These connections, it turns out, are important for fitting to brain data [19].
4.0. System-2
Researchers generally understand System-2 in terms of a collection of cognitive properties
characterized as slow, propositional, rational, effortful, and requiring working memory [4, 20, 35].
System-2 involves explicit propositional knowledge that is used to guide decision-making [37].
Propositional knowledge is associated with relational knowledge [38] which represents entities (e.g.:
John and Mary), the relation between them (e.g.: loves) and the role of those entities in that relation
(e.g.: John loves Mary). Higher level rationality in System-2 is also said to be epistemically committed
to logical standards [6]. System-2 processes are associated with the subjective experiences of agency,
choice, and effortful concentration [35]. The term “effortful” encompasses the intentional, conscious,
and more strenuous use of knowledge in complex thinking. Higher level rationality is considered
responsible for human-like reasoning, allowing for hypothetical thinking, long-range planning, and
is correlated with measures of general intelligence [15].
Researchers have studied various ways in which System-2’s effortful processes can intervene in
System-1 automatic operations [5]. Ordinarily, an individual does not need to invoke System-2 unless
they notice that System-1 automaticity is insufficient or risky. System-2 can intervene when the
anticipated System-1 output would infringe on explicit rules or potentially cause harm. For example,
a scientist early in their experiment may notice that they are experiencing a feeling of certainty.
System-2 can instruct them to resist jumping to conclusions and to gather more data. In this sense,
System-2 can monitor System-1 and override it by applying conceptual rules.
4.1. System-2 in the Common Model
Laird [39] draws on Newell [40], Legg and Hutter [41] and others to equate rationality with
intelligence, where an agent uses its available knowledge to select the best action(s) to achieve its
goal(s).” Newells Rationality Principle involves the assumption that problem-solving occurs in a
problem space, where knowledge is used to navigate toward a desired end. As Newell puts it, an
agent will use the knowledge it has of its environment to achieve its goals” [42]. The prioritizing of
knowledge in decision-making corresponds with the principles of classical computation involving
symbol manipulation and transformation.
The Common Model architecture fundamentally distinguishes between declarative memory and
procedural memory. This maps roughly onto the distinction between explicit and implicit knowledge;
while declarative knowledge can be made explicitly available to working memory, procedural
knowledge operates outside of working memory and is not directly accessible. However, declarative
knowledge can also function in an implicit way. The presence of something within working memory
does not necessarily mean it will be consciously accessed [43].
Higher level reasoning involves the retrieval of declarative information, representing
propositional information, into working memory to assist in calculations and problem-solving
operations. This appears to correlate with what System-2 researchers describe as effortful”, as this
requires more computational resources (i.e., more production cycles) to manage the flow of
information through limited space in working memory. As Kahneman points out, System-1 can
involve knowledge of simple processes such as 2+2=4. However, more complex operations such as 17x16
require calculations that are effortful, a characteristic that is considered distinctive of System-2 [20].
Effort, within the Common Model, involves greater computational resources being allocated
toward a task. Moreover, the retrieval and processing of declarative knowledge requires more steps
and more processing time when compared to the firing of productions alone. This longer retrieval
and processing time can also account for the characteristic of “slow” associated with System-2.
5.0. Effort in System-1 and 2
The concept of “effort” makes up a significant and confusing dimension of System-1 and System-2.
While it is mainly associated with System-2 rationality, a precise definition of “effort” remains elusive
and is largely implicit in discussions of System-1 and System-2. Because System-2 is considered to
have a low processing capacity, its operations are associated with greater effort and a de-prioritizing
of irrelevant stimuli [44].
Effort can be associated with complex calculations in System-2 to the extent that it taxes working
memory. Alternatively, effort can be associated with System-2’s capacity to overrule or suppress
automatic processes in System-1 [20]. For example, various System-1 biases (such as the “belief bias”)
can be subdued by instructing people to make a significant effort to reason deductively [45]. The
application of formal rules to control cognitive processes is also called metacognition the
monitoring and control of cognition [46, 47]. Researchers have also interpreted metacognition
through a System-1 and System-2 framework [48, 49]. System-1 metacognition is thought to be
implicit, automatic, affect-driven, and not requiring working memory. System-2 metacognition is
considered explicit, rule-based, and relying on working memory.
While the concept of “effort” is considered to be the monopoly of System-2, a computational
approach suggests that effort is instead a continuum with low effort cognitive phenomena being
associated with System-1, and high effort cognitive phenomena being associated with System-2.
5.1. Effort in the Common Model
The Common Model helps to elucidate how “effort” can be present in System-1 type operations in
the absence of other System-2 characteristics. While neither dual-system theories nor the Common
Model contain a clear definition of “effort”, computational characteristics associated with effort can
be necessitated by System-1. For instance, “effort” is often associated with the intense use of working
memory. However, the Common Model requires working memory (along with its processing
limitations) for both System-1 and System-2 type operations. There is little reason why System-1
should necessarily use less working memory than System-2 in the Common Model. Instead, it would
depend on the task duration and intensity.
System-1 and System-2 can also be clarified by importing Proust’s [11] more precise account.
Proust attempted to elucidate these two systems by claiming that they should be distinguished by
their distinctive informational formats (System-1 non-conceptual; System-2 conceptual). In this sense,
System-1 metacognition can exert effortful control while simultaneously being implicit and non-
conceptual. For example, consider a tired graduate student attending a conference while struggling
not to fall asleep. An example of System-1 metacognition would involve the context implicitly
prompting them to feel nervous, noticing their own fatigue, and then attempting to stay awake. This
effort is context-driven, implicit, non-conceptual, and effortful. Alternatively, System-2
metacognition can exert effort by way of explicit concepts, as in the case of a tired conference-
attendee repeating the verbal instruction “try to focus”. Both scenarios could be modelled using the
Common Model, and to reiterate, there is hardly any reason why System-1 should require less effort.
One way to think about effort is in terms of the expense of neural energy. In this sense, effort can
be viewed as the result of greater caloric expenditure in neurons. The neural and computational
dynamics responsible for the effortful control of internal states have shown to be sensitive to
performance incentives [50]. Research also indicates that the allocation of effort as cognitive control
is dependent on whether a goal’s reward outweighs its costs [51]. Both of these relate to
reinforcement learning, which is associated with System-1. Additionally, effort in the Common Model
can also be thought of in terms of the number of production cycles, where more cycles equate to more
effort. In these terms, high effort is synonymous with slow processing, in that it takes longer to
process a task. However, the number of production cycles would remain constant per unit of time.
Finally, it may be the case that overall effort is not the appropriate metric. Instead, System-2
thinking may be associated with a particular type of effort. Models of rational thinking in the Common
Model architectures tend to use productions to manipulate information in working memory. For
example, a production might take information from one chunk, modify it, and insert it into another
chunk. This sort of read/write activity may in fact be what humans find effortful. Another possibility
is that effort is related to the number of modules in use.
6.0. Emotion in System-1 and 2
Emotion and affect play a vital role in the distinction between System-1 and System-2 processes [20,
52]. Decisions in System-1 are largely motivated by an individual’s implicit association of a stimulus
with an emotion or affect (feelings that something is bad or good). Behavior that is motivated by
emotion or affect is faster, more automatic, and less cognitively expensive. One evolutionary
advantage of these processes is that they allow for split-second reactions that can be crucial for
avoiding predators, catching food, and interacting with complex and uncertain environments.
Emotions can bias or overwhelm purely rational decision processes, but they can also be
overridden by System-2 formal rules. While emotions and affect have historically been cast as the
antithesis of reason, their importance in decision-making is being increasingly investigated by
researchers who give affect a primary role in motivating decisions [53, 54]. Some maintain that
rationality itself is not possible without emotion, as any instrumentally rational system must
necessarily pursues desires [55].
6.1. Emotion in the Common Model
Feelings and emotions have strong effects on human performance and decision-making. However,
there is considerable disagreement over what feelings and emotions are and how they can be
incorporated into cognitive models. However, while philosophical explanations of emotion have been
debated, numerous Common Model accounts of emotional phenomena have been created. Somatic
markers have been modeled as emotional tags attached to units of information [56]. Low-level
appraisals have been modeled as architectural self-reflections on factors such as expectedness,
familiarity, and desirability [57]. Core affect theory has been modeled to allow agents to prioritize
information using emotional valuation [58]. Alarm has been modeled as productions operating in the
amygdala [59]. Affect and feelings have been modelled by treating them as non-propositional
representations in working memory, or “metadata” [60].
Overall, the question of how to model emotion and affect in the Common Model remains
unresolved. However, a recent (2022) workshop on emotion in the Common Model has reached some
level consensus:
1. Emotion acts in parallel to evaluate the threat levels and the desirability of both external states
(e.g., an approaching tiger) and internal states (e.g., noticing you’re fatigued and losing focus
in a meeting).
2. Emotion can access information from working memory.
3. Emotion can input information into working memory.
4. Emotion can influence other modules through parameter adjustments (e.g., raise noise levels,
lower thresholds, etc.).
From this viewpoint, emotion plays an important role in System-1 and System-2 thinking. One reason
why System-2 and rational thinking have been thought of as separate from emotion is that rational
thought requires sustained focus, which often means ignoring emotional distractions. However, from
a Common Model perspective, the desire to attain a goal associated with the outcome of rational
thought is also emotional. This emotion arguably remains in the background as it is required by
System-2 processes.
7.0. Learning in System-1 and 2
Researchers have associated System-1 with learning that is automatic, fast, and implicit, while
System-2 learning is considered to be deliberate, slow, and explicit [36]. System-1 implicit learning
usually involves subjects being unaware of what they are learning [61]. In contrast, System-2 explicit
learning entails the intentional learning of information such as memorizing a list of word pairs.
Research supports a distinction between implicit and explicit learning. Evidence shows that while
subjects with amnesia have a reduced capacity for explicit learning, their implicit learning can remain
[62]. The distinction between implicit and explicit learning is strongly related to learning in procedural
memory and declarative memory, which are fundamental to psychology and neuroscience [63].
In the case of skill learning, Fitts & Posner [64] advanced a three-stage skill acquisition model
whereby slow explicit knowledge, when repeatedly practiced, becomes converted into fast implicit
knowledge. From a dual-system perspective, this is a process by which cognitive operations migrate
from System-2 to System-1 as greater skill is developed [65].
7.1. Learning in the Common Model
As Laird and Mohan [66] noted, the Common Model can be thought of as having two levels of
learning. Level 1 learning includes all automatic architectural learning mechanisms. This includes
associative learning and episodic learning in declarative memory, as well as procedural compilation
and reinforcement learning in procedural memory. Level 2 learning involves knowledge-based
metacognitive strategies that create experiences for Level 1 mechanisms to learn from, such as
deliberate practice and studying. Metacognitive strategies can be encoded into procedural memory
for more efficient use of Level 1 mechanisms, but can also be encoded as declarative knowledge that
is interpreted by procedural knowledge.
The Common Model also compiles stored instructions from declarative memory into procedural
memory through repeated interpretation and practice [67]. When applied to Proust’s [11] dual-system
framework for metacognition, this allows System-2 metacognition to be compiled into System-1
metacognition. It provides a source for the production-based metacognitive strategies discussed above
(although they can also be learned through implicit, automatic mechanisms). This lends insight into
the possible mechanisms that underlie metacognitive skill learning. Productions for higher-level
logical thinking are likely also acquired this way (although some may be innate).
8.0. Conclusion
The following insights arise from grounding System-1 and System-2 in the Common Model:
1. Both System-1 and System-2 rely on procedural memory. While System-1 is more directly driven
by the fast, automatic actions of the production system, System-2 is also reliant on the
production system. Even when System-2 is driven primarily by explicit declarative knowledge,
it requires the production system to retrieve and act on that knowledge.
2. System-2 can involve emotion. System-2 goal-directed rationality requires affect in (at least) the
form of a preferred desired end. Further, according to the Common Model, there are multiple
routes for System-2 rationality to be influenced by System-1 affective biases.
3. Both System-1 and System-2 require working memory. While the conventional view is that
System-1 does not require working memory, the constraints of the Common Model necessitate
it. Production rules (procedural knowledge) are activated by the content of working memory.
Hence working memory is required for even the simplest System-1 processing.
4. Effort is not well defined. There is little evidence that System-2 requires more effort than
System-1. Rather, humans seem to be conscious of the effort required to maintain focused
rational thought over longer periods of time.
Regardless of whether one adopts the Common Model architecture, researchers should be cautious of
assuming that System-1 and System-2 can be treated as separate, dichotomous modules. The
framework is far from agreed upon and deep issues continue to be unresolved.
Since Descartes, dualism has continually been reimagined as mind and soul, reason and emotions,
and opposing modes of thought (e.g.: System-1 and System-2). These represent our attempts to make
sense of our own minds, its processes, and how this understanding maps onto our personal
experiences. Clearly, System-1 and System-2 capture something deeply intuitive about the
phenomenology of cognition.
Interpreting System-1 and System-2 within the Common Model results in our concluding that the
“alignment assumption” (that the two systems are opposites) is a false dichotomy. There are, of
course, cases where all properties of System-1 and System-2 are cleanly bifurcated on either side.
However, between these two extremities lies a spectrum where the characteristics are mixed. In fact,
from a Common Model perspective, it would be more accurate to say that the relationship is
hierarchical, with System-2 built on top of System-1. In this sense, a “Levels” characterization might
be more appropriate. A good candidate for this is Newell’s distinction between the Cognitive Level
and the Knowledge Level [40], which conceptually maps onto System-1 and System-2.
Interpreting System-1 and System-2 within the Common Model suggests that System-2 is
emergent from the dynamic interactions within System-1, in humans. This raises the question of
whether this is the best way to build System-1 and System-2 in an AI. Possibly, building them as
separate systems would avoid certain shortcomings of the human system. Certainly, it can be more
desirable from a design and tractability perspective. However, it is also possible that a two systems
approach would be less flexible and less adaptive than an emergent levels approach.
The Common Model of Cognition provides a comprehensive view of the computational units
involved in System-1 and System-2 type processes. By grounding dual-process models within the
framework of the Common Model we gain a clearer understanding of the underlying mechanism
involved.
References
[1] J. E., Laird, C. Lebiere, P.S. Rosenbloom, A standard model of the mind: Toward a common
computational framework across artificial intelligence, cognitive science, neuroscience, and
robotics." Ai Magazine 38, no. 4 (2017): 13-26. https://doi.org/10.1609/aimag.v38i4.2744
[2] P. C. Wason, J. S. B. Evans, Dual processes in reasoning?. Cognition 3, no. 2 (1974): 141-154.
[3] K.E. Stanovich, Who is rational?: Studies of individual differences in reasoning. Psychology Press,
1999. https://doi.org/10.4324/9781410603432
[4] F. Strack, R. Deutsch, Reflective and impulsive determinants of social behavior. Personality and
social psychology review 8, no. 3 (2004): 220-247.
[5] D. Kahneman, A perspective on judgment and choice: mapping bounded rationality. American
psychologist 58, no. 9 (2003): 697. https://doi.org/10.1037/0003-066X.58.9.697
[6] T. Tsujii, S. Watanabe, Neural correlates of dual-task effect on belief-bias syllogistic reasoning: a
near-infrared spectroscopy study. Brain research 1287 (2009): 118-125.
[7] G. Keren, Y. Schul, Two is not always better than one: A critical evaluation of two-system theories.
Perspectives on psychological science 4, no. 6 (2009): 533-550.
[8] G. Pennycook, W. De Neys, J.S.B. Evans, K.E. Stanovich, V.A. Thompson, The mythical dual-
process typology. Trends in Cognitive Sciences 22, no. 8 (2018): 667-668.
[9] J. De Houwer, Moving beyond System 1 and System 2: Conditioning, implicit evaluation, and
habitual responding might be mediated by relational knowledge. Experimental Psychology 66, no.
4 (2019): 257. https://doi.org/10.1027/1618-3169/a000450
[10] A. Moors, Automaticity: Componential, causal, and mechanistic explanations. Annual review of
psychology 67, no. 1 (2016): 263-287.
[11] J. Proust, The philosophy of metacognition: Mental agency and self-awareness. Oxford, 2013.
[12] R. Thomson, C. Lebiere, J.R. Anderson, J. Staszewski, A general instance-based learning
framework for studying intuitive decision-making in a cognitive architecture. Journal of Applied
Research in Memory and Cognition 4, no. 3 (2015): 180-190.
[13] R. Sun, S. Paul, T. Chris, The interaction of the explicit and the implicit in skill learning: a dual-
process approach." Psychological review 112, no. 1 (2005): 159.
[14] U. Faghihi, C. Estey, R. McCall, S. Franklin, A cognitive model fleshes out Kahneman’s fast and
slow systems. Biologically Inspired Cognitive Architectures 11 (2015): 38-52.
[15] J.S.B. Evans, In two minds: dual-process accounts of reasoning. Trends in cognitive sciences 7,
no. 10 (2003): 454-459. https://doi.org/10.1016/j.tics.2003.08.012
[16] D.C., Dennett, Brainstorms: Philosophical essays on mind and psychology. MIT press, 2017.
[17] A. Newell, You can't play 20 questions with nature and win: Projective comments on the papers
of this symposium. (1973).
[18] Z. Steine-Hanson, N. Koh, A. Stocco, Refining the common model of cognition through large
neuroscience data. Procedia computer science 145 (2018): 813-820.
[19] A. Stocco, C. Sibert, Z. Steine-Hanson, N. Koh, J.E. Laird, C.J. Lebiere, P. Rosenbloom,
Analysis of the human connectome data supports the notion of a “Common Model of Cognition”
for human and human-like intelligence across domains." NeuroImage 235 (2021): 118035.
[20] D. Kahneman, Thinking, fast and slow. Macmillan, 2011.
[21] J. S. B. Evans, K.E. Stanovich, Dual-process theories of higher cognition: Advancing the debate.
Perspectives on psychological science 8, no. 3 (2013): 223-241.
[22] F. Strack, R. Deutsch, Reflective and impulsive determinants of social behavior. Personality and
social psychology review 8, no. 3 (2004): 220-247.
[23] J. Fodor, The modularity of mind. Scranton. (1983).
[24] K. E. Stanovich, R. F. West, Individual differences in reasoning: Implications for the rationality
debate?. Behavioral and brain sciences 23, no. 5 (2000): 645-665.
[25] D. Kahneman, A perspective on judgment and choice: mapping bounded rationality. American
psychologist 58, no. 9 (2003): 697. https://doi.org/10.1037/0003-066X.58.9.697
[26] H. A. Simon, Herbert, A behavioral model of rational choice." The quarterly journal of economics
69, no. 1 (1955): 99-118. doi:10.2307/1884852
[27] A. G. Barto, S. S. Richard, S. B. Peter, Associative search network: A reinforcement learning
associative memory. Biological cybernetics 40, no. 3 (1981): 201-211.
[28] P. McLeod, K. Plunkett, E. T. Rolls, Introduction to connectionist modelling of cognitive
processes. Oxford University Press, 1998.
[29] D. G. Mitchell, The nexus between decision making and emotion regulation: a review of
convergent neurocognitive substrates. Behavioural brain research 217, no. 1 (2011): 215-231.
[30] P. Slovic, M.L. Finucane, E. Peters, D.G. MacGregor, Risk as analysis and risk as feelings: Some
thoughts about affect, reason, risk and rationality. In The feeling of risk, pp. 21-36. Routledge,
2013. doi:10.1111/j.0272-4332.2004.00433.x
[31] D. Kahneman, A. Tversky, On the psychology of prediction. Psychological review 80, no. 4
(1973): 237. https://doi.org/10.1037/h0034747
[32] J. T. Hart, Julian, Memory and the feeling-of-knowing experience. Journal of educational
psychology 56, no. 4 (1965): 208.
[33] M. K. Singley, J. R. Anderson, The transfer of cognitive skill. No. 9. Harvard Uni. Press, 1989.
[34] J. R. Anderson, Knowledge representation. Rules of the mind (1993): 17-44.
[35] A. Stocco, C. Lebiere, J. R. Anderson, Conditional routing of information to the cortex: A model
of the basal ganglia’s role in cognitive coordination. Psychological review 117, no. 2 (2010): 541.
[36] K. Frankish, Dual‐process and dual‐system theories of reasoning. Philosophy Compass 5, no. 10
(2010): 914-926. https://doi.org/10.1111/j.1747-9991.2010.00330.x
[37] S. Epstein, R. Pacini, Some basic issues regarding dual-process theories from the perspective of
cognitiveexperiential self-theory. (1999).
[38] G. S. Halford, W. H. Wilson, S. Phillips, Relational knowledge: The foundation of higher
cognition. Trends in cognitive sciences 14, no. 11 (2010): 497-505.
[39] J. Laird, Intelligence, knowledge & human-like intelligence. Journal of Artificial General
Intelligence 11, no. 2 (2020): 41-44. 10.2478/jagi-2020-0003
[40] A. Newell, Unified theories of cognition, Harvard University Press, Cambridge, MA (1990).
[41] S. Legg, M. Hutter, Universal intelligence: A definition of machine intelligence. Minds and
machines 17, no. 4 (2007): 391-444.
[42] A. Newell, The knowledge level. Artificial intelligence 18, no. 1 (1982): 87-127.
[43] D. Wallach, C. Lebiere, Implicit and explicit learning in a unified architecture of cognition.
Attention and implicit learning (2003): 215-250.
[44] K. E. Stanovich, Who is rational?: Studies of individual differences in reasoning. Psychology
Press, 1999. https://doi.org/10.4324/9781410603432
[45] J. S. B. Evans, T. Julie, L. Barston, P. Pollard, On the conflict between logic and belief in
syllogistic reasoning. Memory & cognition 11, no. 3 (1983): 295-306.
[46] J. H. Flavell, Metacognition and cognitive monitoring: A new area of cognitivedevelopmental
inquiry. American psychologist 34, no. 10 (1979): 906.
[47] L. Fletcher, P. Carruthers, Metacognition and reasoning. Philosophical Transactions of the Royal
Society B: Biological Sciences 367, no. 1594 (2012): 1366-1378.
[48] S. Arango-Muñoz, Two levels of metacognition. Philosophia 39, no. 1 (2011): 71-82.
[49] N. Shea, A. Boldt, D. Bang, N. Yeung, C. Heyes, C.D. Frith, Supra-personal cognitive control
and metacognition. Trends in cognitive sciences 18. (2014):186-193.
[50] S. W. Egger, E. D. Remington, C. J. Chang, M. Jazayeri, Internal models of sensorimotor
integration regulate cortical dynamics. Nature neuroscience 22, no. 11 (2019): 1871-1882.
[51] A. Shenhav, S. Musslick, F. Lieder, W. Kool, T. L. Griffiths, J.D. Cohen, M. M. Botvinick,
Toward a rational and mechanistic account of mental effort. Annual review of neuroscience 40,
no. 1 (2017): 99-124. https://doi.org/10.1146/annurev-neuro-072116-031526
[52] S. Chaiken, and Y. Trope, Dual-process theories in social psychology. Guilford Press, 1999.
[53] R. B. Zajonc, Feeling and thinking: Preferences need no inferences. American psychologist 35,
no. 2 (1980): 151. https://doi.org/10.1037/0003-066X.35.2.151
[54] L. F. Barrett, P. Salovey, The wisdom in feeling: Psychological processes in emotional
intelligence. Guilford Press, 2002.
[55] J. S. B. Evans, Spot the difference: distinguishing between two kinds of processing. Mind &
Society 11, no. 1 (2012): 121-131.
[56] A. R. Damasio, Descartes' error and the future of human life. Scientific American 271, no. 4
(1994): 144-144.
[57] P. S. Rosenbloom, J. Gratch, V. Ustun, Towards emotion in sigma: from appraisal to attention. In
International conference on artificial general intelligence, pp. 142-151. Springer, Cham, 2015.
[58] I. Juvina, O. Larue, A. Hough, Modeling valuation and core affect in a cognitive architecture: The
impact of valence and arousal on memory and decision-making. Cognitive Systems Research 48
(2018): 4-24. https://doi.org/10.1016/j.cogsys.2017.06.002
[59] R. L. West, J. T. Young, Proposal to add emotion to the standard model. In 2017 AAAI Fall
Symposium Series. 2017.
[60] R. L. West, B. Conway-Smith, Put Feeling into Cognitive Models: A Computational Theory of
Feeling. In Proceedings of ICCM 2019 17th International Conference on Cognitive Modelling,
2019 iccm-conference.neocities.org/2019/proceedings/papers/ICCM2019_paper_47.pdf
[61] P. A. Frensch, D. Rünger, Implicit learning. Current directions in psychological science 12, no. 1
(2003): 13-18. https://doi.org/10.1111/1467-8721.01213
[62] B. Milner, Les troubles de la memoire accompagnant des lesions hippocampiques bilaterales.
Physiologie de l’hippocampe 107 (1962): 257-272.
[63] L. R. Squire, Declarative and nondeclarative memory: Multiple brain systems supporting learning
and memory. Journal of cognitive neuroscience 4, no. 3 (1992): 232-243.
[64] P. M. Fitts, M. I. Posner, Human performance. brooks. Cole, Belmont, CA 5 (1967): 7-16
[65] D. Kahneman, S. Frederick, Attribute substitution in intuitive judgment. Models of a man: Essays
in memory of Herbert A. Simon (2004): 411-432.
[66] J. Laird, S. Mohan, Learning fast and slow: Levels of learning in general autonomous intelligent
agents. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1. 2018.
[67] J. R. Anderson, Acquisition of cognitive skill. Psychological review 89, no. 4 (1982): 369.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
The Common Model of Cognition (CMC) is a recently proposed, consensus architecture intended to capture decades of progress in cognitive science on modeling human and human-like intelligence. Because of the broad agreement around it and preliminary mappings of its components to specific brain areas, we hypothesized that the CMC could be a candidate model of the large-scale functional architecture of the human brain. To test this hypothesis, we analyzed functional MRI data from 200 participants and seven different tasks that cover a broad range of cognitive domains. The CMC components were identified with functionally homologous brain regions through canonical fMRI analysis, and their communication pathways were translated into predicted patterns of effective connectivity between regions. The resulting dynamic linear model was implemented and fitted using Dynamic Causal Modeling, and compared against six alternative brain architectures that had been previously proposed in the field of neuroscience (three hierarchical architectures and three hub-and-spoke architectures) using a Bayesian approach. The results show that, in all cases, the CMC vastly outperforms all other architectures, both within each domain and across all tasks. These findings suggest that a common set of architectural principles that could be used for artificial intelligence also underpins human brain function across multiple cognitive domains.
Article
Full-text available
Sensorimotor control during overt movements is characterized in terms of three building blocks: a controller, a simulator and a state estimator. We asked whether the same framework could explain the control of internal states in the absence of movements. Recently, it was shown that the brain controls the timing of future movements by adjusting an internal speed command. We trained monkeys in a novel task in which the speed command had to be dynamically controlled based on the timing of a sequence of flashes. Recordings from the frontal cortex provided evidence that the brain updates the internal speed command after each flash based on the error between the timing of the flash and the anticipated timing of the flash derived from a simulated motor plan. These findings suggest that cognitive control of internal states may be understood in terms of the same computational principles as motor control.
Conference Paper
Full-text available
Feelings are potentially conscious experiences that inform us about brain/body states related to drives (e.g., feeling hungry), emotions (e.g., feeling angry) and knowledge states (e.g., feeling unsure). In this paper we propose a unified computational definition of feelings that can be used to add feelings to cognitive models.
Article
Full-text available
It is generally assumed that relational knowledge is the foundation of higher cognition such as (analogical and conditional) reasoning, language, the use of relational categories, and planning. Dual-system models (e.g., Kahneman, 2011) that divide the realm of cognition into two systems with opposing properties (e.g., fast vs. slow, intentional vs. unintentional, conscious vs. unconscious, associative vs. propositional) foster the view that other psychological phenomena are not relational in nature. In this paper, I argue that the impact of relational knowledge is more widespread than dual-system models imply. More specifically, I review evidence suggesting that also Pavlovian conditioning, implicit evaluation, and habitual responding are mediated by relational knowledge. Considering the idea that relational knowledge underlies also fast, unintentional, unconscious, and seemingly associative psychological phenomena is not only theoretically important but also reveals new opportunities for influencing thinking and behavior.
Article
Full-text available
The Common Model of Cognition (CMC) is an effort to highlight the commonalities between multiple studies of human-like intelligent minds and cognition, and bring these architectures and functions together into one common model. The CMC is still relatively new, and while it has substantial theoretical evidence in its favor, there is little empirical evidence confirming the theory. The CMC is fundamentally informed by human cognition, and must at least hold true for human brains. As such, this paper uses a large fMRI dataset from the Human Connectome Project (HCP) to refine the CMC and test it as a reasonable model for human cognition. We tested three models in the CMC family and two alternative models against two tasks from the HCP dataset. We found that one model from the CMC family explained the HCP dataset best, providing further empirical evidence in favor of the CMC, whilst also suggesting a slight modification to the CMC itself that may improve the model’s generalizability.
Article
Full-text available
The purpose of this article is to begin the process of engaging the international research community in developing what can be called a standard model of the mind, where the mind we have in mind here is human-like. The notion of a standard model has its roots in physics, where over more than a half-century the international community has developed and tested a standard model that combines much of what is known about particles. This model is assumed to be internally consistent, yet still have major gaps. Its function is to serve as a cumulative reference point for the field while also driving efforts to both extend and break it.