To maximize the heuristic value of the proposed theoretical framework, the three modes of creativity are tied to their underlying mechanisms at three different levels of description: (A) neuroanatomy, (B) processes, and (C) EAs. This is not intended to be a hierarchical setup. Rather, the descriptions are complementary, often expressing the same thing but at a different explanatory level and in a different language. This also does not produce redundancies, but is instead explicitly meant to facilitate the identification of as many points of contact between creativity and the various areas of the psychological sciences. Also, no claim is made that the underlying mechanisms described here represent an exhaustive list. But the present theoretical framework does focus heavily on those aspects that also dominate the mainstream literature.
For (A) neuroanatomy, differences in the three modes are specified for the following key aspects: prefrontal cortex activity, neural network configuration or task sets, and explicit-implicit systems (for an overview of task sets and related concepts, see Monsell, 2003 and for the explicit-implicit distinction, see Dienes & Perner, 1999). For (B) processes, three sublevels – cognitive, computational, and physiological – could be considered, most of which change as a function of changes in (A) neuroanatomy. The focus here will be on cognitive processes, such as the various higher-cognitive functions – working memory, attention, cognitive flexibility, inhibitory control, memory retrieval, cognitive biases, functional fixedness, social norms, schemas, agency, etc. – spreading activation, task-set inertia, speed of processing, procedural memory, or processing efficiency. But, when applicable, computational processes, such as Bayesian inferences (for an introduction, see Wolpert, Doya, & Kawato, 2003), or physiological processes, such as levels of arousal or temporal synchrony, will also be highlighted.
For (C) EAs, a number of additional complexities need to be drawn out before the differences among the three modes of creativity can be appreciated. For all the heat in the debate on cultural evolution, there is a remarkable, broad consensus on key issues. To a first approximation, the following is unanimously agreed upon: (1) Culture is an evolutionary system (Dawkins, 1976; Gould, 1979; Smith, 2013); (2) Culture is a system with a variational or variation-selection pattern of change (Lewontin, 1991; Richerson & Boyd, 2005), and (3) Culture exhibits some coupling between variation and selection; that is, unlike biological EAs, the brain’s EAs are not blind, but rather fall on a continuum of degrees of sightedness (Dietrich & Haider, 2015; Kronfeldner, 2007; Simonton, 2013; for an introduction to sightedness, see Kronfeldner, 2010).
Recently, the brain’s prediction system (for an overview, see Bar, 2009) has been proposed as the brain mechanism of the sightedness upgrade that cultural EAs possess (Dietrich, 2015). The ability to compute a goal representation that contains fitness values with which to predict the location of solutions with a probability greater than zero, even when the problem space is unknown, causes unique properties that change the way cultural EAs operate and the type of artifacts they can bring into existence. Most importantly, predictive computations yield degrees of sightedness, enable cognitive scaffolding and generate the feeling of foresight and intention of human creativity (for details, see Dietrich, 2015). The brain’s predictive capabilities are known to differ for (1) explicit and implicit systems (Downing, 2009; Grush, 2004) and (2) engagement of higher-cognitive functions (Dietrich & Haider, 2015; Moulton & Kosslyn, 2009). In consequence, all three types of creativity can also be delineated on the basis of their EAs’ sightedness parameter, which has implications for the search heuristics, scaffolding, and sense of foresight and agency of each type. Strictly speaking, predictive computations should fall under (B) processes, but they are better discussed under (C) EAs, as they are the reason for the differences in degrees of sightedness of each type.
The deliberate mode
Evidence from connectionist modeling, psychology, and neuroscience have converged on the view that conscious processing and other higher-cognitive functions involve competition between widely distributed representations that are strongly dependent on top-down prefrontal activity (Baars, 1988; Dehaene & Changeux, 2011; Maia & Cleeremans, 2005; Miller & Cohen, 2001; Norman & Shallice, 1986). Based on this literature, the DM is characterized by strong prefrontal cortex activation that biases and thus directs information processing (Dietrich, 2004a). It follows that the DM can be associated with the explicit system in terms of the dual-system (explicit-implicit) architecture (Dienes & Perner, 1999; Haider & Frensch, 2005; Reber, 1993) and, to a limited extent, with the central-executive network (CEN) in terms of the brain’s large-scale networks (Bressler & Menon, 2010; Raichle et al., 2001). Finally, in connectionist terms, the DM can be characterized as a strong task set, in which the node weights of the knowledge structure are set to highly polarized frequency of occurrence values (Dietrich & Haider, 2017). Although task sets are standard explanations in cognitive psychology (Allport, Styles & Hsieh, 1994; Dreisbach & Haider, 2009; Monsell, 2003), the concept has not been applied to creative thinking, despite their obvious relevance to how we might conceptualize a given task, initially configure a task’s potential solution spaces, or functional fixedness.
From the conclusions for (A) neuroanatomy, the following corollaries for (B) processes can be derived. In the deliberate mode, the explicit control and the prefrontal activation bring the toolbox of the higher-cognitive function to bear on the creative task, including working memory, executive attention, inhibitory control, conscious memory retrieval, and agency. It is for this intentional and effortful way that it is called the deliberate mode. One should not conclude from this, however, that creativity depends on all the computational power and cognitive flexibility these higher-order processes make possible. In an effort to provide a constant stream of reminders against false category formations, they also bring disadvantages (Dietrich, 2015). Top-down prefrontal activity is associated with the strong activation of social norms, schemas, and cognitive biases (Damasio, 1994; Miller & Cohen, 2001), so that the effortful retrieval of knowledge from long-term memory and the recombinatorial shuffling of it in working memory is likely inherently biased towards general beliefs of what is considered true about the world. In other words, the search engine itself as well as the rearrangement of informational bits has built-in predispositions that are constrained by a number of parameters, such as biases, expectancies, preferences, schemas, experiences, or common sense. As such, the DM is more limited to creative insights that are more paradigmatic in type and rely on more close associations (Dietrich, 2004a).
Recasting the same tradeoff in network terminology, the DM’s strong top-down input changes how spreading activation ripples through the knowledge structure. The frequency of occurrence values and the strength of their connections are recalibrated so that the task set now reflects the extent to which certain aspects of the task or a person’s expertise and belief system are expected to contribute to a solution. These new calibrations then determine how information flows through that network. The more polarized the weight assignment, in either the plus or minus direction, the less likely it is for spreading activation to diffuse and establish remote associations. Such an opinionated knowledge structure is highly efficient – or highly inflexible, depending on the point of view – when it processes specific information (Dietrich & Haider, 2017).
The DM’s general pattern of activation also has consequences for the parameters of (C) EAs. The bringing online of the explicit system’s full arsenal of higher-cognitive functions includes prediction processes as well (Dietrich, 2015; Downing, 2009). In the DM, the brain’s neural simulators can thus compute highly informed goal representations that yield more degrees of sightedness when exploring unknown problem spaces. These superior EAs have strongly directional heuristics, the capacity to scaffold, and bring about the sense of foresight and agency (Dietrich, 2015). It is a powerful upgrade from the blind EAs of biological evolution and it can make short work of a given problem space. The tradeoff highlighted above can also be expressed at this evolutionary level of description. The advantage of advanced heuristics is efficiency (see Gigerenzer, & Gaissmaier, 2011, for details). By pre-empting, a priori, counterintuitive paths or ideas orthogonal to accepted wisdom, a vast search space is trimmed down to a more manageable region that is more likely to contain a solution. This only works well, however, if the solution is indeed located in the predicted region of the problem space. To quip, while the DM has the advantage of limiting the solution space, it has the disadvantage of limiting the solution space!
In keeping with efforts to build bridges between different levels of description, we can ask what might constitute a selection process – an aspect of (C) EAs – in a connectionist network for (A) neuroanatomy and (B) processes. How could a fitness function be implemented in such a network? One possible mechanism is the speed of processing (Whittlesea, 2004). A weighted network creates processing speed differentials for different kinds of information that are run on it. In evolutionary language, a strongly polarized network corresponds to a case of strong selection pressure. Depending on the configuration, some information is processed more expediently than others. That information is simply a better fit for the preset processing pathways. The speed-of-processing parameter could serve as a selection process because it can discern value (Dietrich & Haider, 2017). Individual differences in creativity can also be understood that way. A person with a different network setting, as determined by factors such as expertise or past experience, might process the same potential solution at a lower speed and would not select for it.
It should be evident at this point that the underlying mechanisms described here for the DM are incompatible with associating the whole of creativity with mind-wandering, right brains, defocused attention, remote associations, unconscious thinking, or the default mode network.
The spontaneous mode
For the SM, a very different picture emerges for all three levels of description. In removing the creative task from consciousness, there are significant changes for all aspects of (A) neuroanatomy. The SM has considerably weaker supervisory, top-down influences from the prefrontal cortex that guided the effortful thinking of the DM (Dietrich, 2004a). However, the SM must still be considered as part of the explicit system. It can draw on explicit long-term memory, use a scaffold – as evidenced by discontinuous insights – and, most importantly, the ongoing, unconscious processing has access to working memory as a creative idea. As we will see shortly, none of that is the case for processing in the encapsulated implicit system (Dienes & Perner, 1999; Haider & Frensch, 2005). Also, the SM could be loosely associated with the default mode network, but this association must be regarded as tentative at present due to the still limited understanding of the functional significance of that network. Finally, the SM could be characterized as a weak task set in which network nodes are reset to much less polarized frequency-of-occurrence values (Dietrich & Haider, 2017).
The adjustments are equally extensive for (B) processes. Once the mind is otherwise applied, the higher-cognitive functions go partially or fully offline. Processing for the creative task is now unconscious, lacks attentional control, and the highly selective retrieval of knowledge from long-term memory is as gone as is the sense of agency. The loss of all this computational power renders the SM a much less efficient information-processor. On the flipside, incubating the creative task also puts an end to the strong, top-down activation of schemas, expectations, and norms, and, with this prefrontal-imposed bias lifted, the SM has the potential to chance upon more paradigm-shifting ideas or remote associations (Dietrich, 2004a).
Expressing the same tradeoff in connectionist language, as soon as the task set is demoted into the unconscious, the activation in this network changes in strength and quality. Right away, this triggers a drop off in overall network intensity, but the network continues to process the task due to task-set inertia (for an introduction to this concept, see Allport, Styles, & Hsieh, 1994). With the loss of top-down influences from the prefrontal cortex, this general decline in activity is accompanied by the resetting of all relevant knowledge nodes to more moderate values. The task set representation of the creative task is now both weaker and more neutral. This has two effects for spreading activation. First, information processing is more diffuse, but this loss of sharpness and direction also comes with a wider reach in the network. Second, processing speed differentials are altered. The same information that might have run at low processing speeds in the DM’s network configuration could do better in a more evenly organized network. Also, there is the possibility that speed of processing works as a strengthening mechanism that can determine which particular combination of information breaks through to working memory and becomes a creative insight (Dietrich & Haider, 2017).
For (C) EAs, the SM must make due with greatly reduced prediction competencies (Dietrich, 2015). Without the benefit of higher-order, conscious thought, neural simulators are unlikely to have either full access to memory or the computational power to generate well-informed goal representations. The SM possesses, therefore, considerably less degrees of sightedness, its search heuristics are less directional, and the sense of foresight and agency can no longer be computed (Dietrich, 2015). The tradeoff here is the opposite from the one in the DM. Due to the reduced sightedness and inferior heuristics, the SM must contend with larger solution spaces and a more aimless search process, making it much less efficient. But if the solution is located outside the region of the solution space predicted by the DM’s strongly directional heuristics, the EAs of the SM are also more likely to find it. To quip again, while the SM has the advantage of larger solution spaces, it has the disadvantage of larger solution spaces!
The SM should be expanded to accommodate the larger set of altered states of consciousness (ASC). Although the daydreaming described above is only a mild sort, it is generally considered an ASC along with a few other putative ASC, such as dreaming, meditation, hypnosis, flow, drug states, and the runner’s high (Blackmore, 2005; Dietrich, 2007b). The transient hypofrontality theory (THT) has been proposed as a general brain mechanism that can account for a great number of phenomenological features common to all ASCs (Dietrich, 2003). It is based on the common conceptualization of brain areas and mental abilities into a functional hierarchy with the top layers in the prefrontal cortex contributing the most sophisticated elements of the conscious experience, such as self-reflection, working memory, executive attention, temporal integration, abstract thinking, cognitive control, volition, or the sense of agency. Analogous to peeling an onion, the THT simply postulates that alteration to consciousness involves the progressive downregulation of networks supporting the highest cognitive capacities, down the functional hierarchy, one phenomenological subtraction at a time, to those supporting more basic ones. Since all altered states share the gradual disappearance of mental faculties that depend on prefrontal input – with the exception of executive attention for some – ASCs include a transient state of hypoactivity, of various depths and extent for each altered state, in networks of the prefrontal cortex. As such, the consequences of the weakened, top-down prefrontal projections that apply to the creative thinking in the SM – little higher-cognitive functions, less sightedness, or more possibility for remote associations – can be extended to ASC in general.
Clearly, neither type of creativity can be said to be better. Each has its advantages and disadvantages depending on the specific situation. The critical factor seems to be where in the problem space the solution is located, which, needless to say, we do not know ahead of time. Generally speaking, if it is in line with the current thinking, the DM is better; if it is orthogonal to it, the SM is better. Also, this analysis underscores the value of decomposing creativity into valid types when making claims about possible cognitive and neural mechanisms. While creativity, as a whole, cannot be associated with decreased prefrontal activity, default mode network, sleep, defocused attention, mind-wandering, unconscious thinking, or remote associations, a particular type of creativity – the SM – might be. This cannot be done with divergent thinking. Notice also that the DM and SM cannot be conflated with convergent and divergent thinking as both the DM and the SM contain both convergent and divergent thinking Fig. 1.
The flow mode
The FM is a radically different way of generating creative behavior. Unlike the DM and SM, both of which do not necessitate physical motion and their processing eventually ends with an active representation in working memory – the creative idea – the FM does require motion and it bypasses consciousness altogether. One feature of the FM – the need for motor efficiency – produces a cascading series of corollaries that fundamentally changes the FM’s mechanisms at all three levels of description (Dietrich, 2015). For this discussion, it is more practical to combine the two levels of description of (A) neuroanatomy and (B) processes.
First, the FM is driven by the implicit system (Dietrich, 2004b). This follows, as a matter of consequence, from the flexibility-efficiency tradeoff between the explicit and implicit systems. The sophisticated and complex explicit system is based on higher-order mental representations and has evolved for cognitive flexibility. The simple and concrete-operational implicit system is based on a procedural representational format and has evolved for speedy and accurate motor execution (Dienes & Perner, 1999; Haider & Frensch, 2005; Reber, 1993). The key to understanding the flexibility-efficiency tradeoff is the computational fact that motor efficiency precludes cognitive flexibility and vice versa (Dietrich & Audiffren, 2011; Dienes & Perner, 1999). A system cannot be complex and fast at the same time, and the need for both at different times is likely the reason we have two distinct information-processing systems in the first place. This tradeoff must be rigorously applied to the FM, because the inherent processing efficiency of implicit knowledge is paramount to the fluid and effortlessness action of the FM.
Second, the FM requires the automatization of the motor skill (Dietrich, 2015). This is a direct corollary of implicit control. The implicit system is experience-based and can only develop a mental representation of the motor task by doing it. In other words, since the motor efficiency that defines the FM can only proceed from an implicit representation, the motor skill must be well practiced first. This is why novices learning a new skill do not report flow experiences (Csikszentmihalyi, 1996). In addition, given that the neural substrate of the implicit system is based on activity in the basal ganglia, cerebellum, and the supplementary motor area (Poldrack & Packard, 2003), so must be the neural substrate of the FM. As such, the creativity that arises in the FM cannot be associated with either the CEN or the DMN.
Third, the FM requires a general state of transient hypofrontality (Dietrich, 2003, 2004b). This, too, follows from the need for motor efficiency. Explicit interference in the control of an automated motor skill – consciously thinking about the movements – decreases the quality of the movement (Beilock & Carr, 2005; Ravizza, 1977), or, to state it the other way around, optimal performance of a real-time sensorimotor integration task is associated with maximal implicitness of its execution (Dietrich, 2004b). A graceful and fluid motor sequence cannot be micro-managed by the explicit system. Its representational format is abstract and higher-order and hence too slow to be applied to a specific situation in real time. A state of lower prefrontal activity, therefore, would minimize the chances that top-down, explicit processes compromise the smooth, implicitly-driven execution of a motor skill and ensure that a person can enter, or can stay, in flow. From an evolutionary perspective, a state of transient hypofrontality is beneficial in a pressure situation involving physical motion; it simply is not adaptive to engage higher-order analytical processes when the devil takes the hindmost (Dietrich & Audiffren, 2011).
Moreover, the THT fits with FM phenomenology. While the movements themselves are performed with ease, the flow experience is marked by phenomenological subtractions of exactly those higher-cognitive functions that depend on prefrontal activity. With the exception of focused attention, the FM lacks conscious awareness, a sense of self and agency, intentionality, abstract analysis, mental time travel, or the ability to consider possible long-term consequences of the ongoing action. Should any of these complex mental faculties return, flow is gone (Dietrich, 2004b). Dennett (2004) once quoted the painter Philip Guston as saying:
When I first come into the studio to work, there is this noisy crowed which follows me there; it includes all of the important painters in history, all of my contemporaries, all the art critics, etc. As I become involved in the work, one by one, they all leave. If I’m lucky, every one of them will disappear. If I’m really lucky, I will too.”
It might be helpful at this point to draw out two issues that are potential sources of confusion regarding the mechanisms of the SM and FM. First, the two types of creativity should not be confused just because both are unconscious. As said, the SM emanates from the explicit system, as it can draw on explicit long-term memory, use a scaffold, and access working memory. The FM can do none of that. Knowledge in the encapsulated implicit system cannot become conscious through an internal, bottom-up process (Dienes & Perner, 1999; Haider & Frensch, 2005; Reber, 1993). It must go, as it does for the FM, through the circuitous route of involving actual behavioral performance (Dietrich, 2015). Put another way, the unconscious creativity in the SM is explicit; the unconscious creativity in the FM is implicit. Second, while both types cannot occur without the muffling of prefrontal-dependent, explicit processes, the effect of transient hypofrontality on the creative process differs for each type. In the SM, the prefrontal hypoactivity changes the way the explicit system itself processes information. In the FM, it facilitates the way the implicit system processes information. Also, since the task itself is automatized, the transient hypofrontality in the FM can also be much more profound because explicit supervision is not critical – and, indeed, is detrimental.
Before the differences that arise from (C) EAs are fleshed out, this can be summarized as follows: Creativity in the DM and SM are offline, outside real time, and tied to consciousness, while in the FM it is online, inside real time, and outside consciousness. Because of this, scientists rarely make use of the FM. A creative act in science typically does not require motor efficiency and can thus be done offline and outside real time. This does not hold for artists, as there are many forms of artistic creativity that require online performance and others that do not.
The shift from explicit to implicit control in the FM makes for a change in (C) EAs that could not be any bigger. The reason for this are the fundamentally different prediction capacities of the implicit system (Dietrich, 2015; Downing, 2009; Grush, 2004). This is perhaps best illustrated by first distinguishing between known and unknown problem spaces (for an earlier discussion on problem spaces, see Boden, 1998).
The motor system does have a prediction mechanism. The system relies heavily on emulators and controllers, which are representations of the future (Wolpert, Ghahramani, & Jordan, 1995). Indeed, it is likely in motor control where the capacity for predictive computation first evolved (Wolpert et al., 2003). However, the motor system can only run these predictions in a known problem space. Its concrete-operational setup cannot handle hypothetical situations, that is, predict the sensory consequences for a movement the body has never made (Dienes & Perner, 1999). The system simply lacks the computational means to generate controllers that are imaginary. Unlike the explicit system, then, it has no way to render a prediction in an unknown problem space. The fact that the motor system cannot compute an internal model of the entire sensorimotor arc is the reason why a first-time movement feels so weird (Frith, 1992). We simply cannot anticipate what we are going to see or feel. In sum, the motor system must already know the goal or movement end states, so that a predictor-controller pair can be established and allowed to converge using Bayesian inferences (Wolpert, et al., 1995). It should be obvious that a known problem space does not count as creativity; the problem space is already mapped out.
The execution of a skilled motor sequence – a known problem space for the motor system – is not even a variation-selection EA. For all possible forward steps or predictors, there are principally known consequences or controllers. Instead, it is a mapping process in which no variation-selection method is needed for the individual movement steps. With all predictor-controller pairs established, the algorithm that needs to be solved is essentially known. In evolutionary language, this can be described by a Lamarckian EA in which the forward steps are directed or instructed by their known adaptive values. The movement sequence can unfold fully sighted, directed by a whole host of acquired and internalized controllers. The equivalent in the explicit system of prediction processes in a known problem space would be, for instance, the game of chess or the planning out of one’s morning chores. This is well-trotted territory, cases of strategic planning or effective decision-making in which the brain runs simulation chains that reason through a series of choice points in the future. In principle, there is (near total) sightedness of the adaptive landscape and the outcome of each step has (near total) predictability.
But creativity is an adventure into parts unknown, which, in motor control, only occurs if we try a totally new movement – snowboarding or dancing tango, for instance. In this case, the motor system has no way of making a prediction because it cannot compute a hypothetical controller. It does not know what the correct movement endpoints might be and without these controllers it would not even know the algorithm it needs to solve (Wolpert et al., 2003). The motor system must first acquire them and it can only accomplish that through action. Learning in the implicit system is stochastic and stepwise. The system tries out, by trial and error, solutions to environmental contingencies (Perruchet & Vinter, 2002). Unlike the explicit system, which forms long-term memories that can then be fed into neural simulators, the implicit system embeds the new information into the network itself by slowly and gradually shifting the weights in its knowledge structure. These weight changes represent predictions that are then used the next time the situation arises. Motor learning is basically a process of acquiring controllers or reducing prediction errors (Wolpert et al., 2003).
In evolutionary terms, one could also say that learning is the process of gaining degrees of sightedness of the fitness landscape. The initial acquisition of a motor skill, then, is best described by a Neo-Darwinian EA. The lack of any kind of prediction competencies to simulate possible movement end states means that the motor system must approach unknown problem spaces blind. It has no controllers or sight of the fitness function. This also explains the disappearance of foresight and agency in the FM (Dietrich, 2015), as evidenced by the quote of Philip Guston, since these experiences also depend on prediction processes (Frith, 1992; Grush, 2004). The great difference to the DM and the SM, of course, is that the explicit system can compute goal representations in such unknown problem spaces. The difference in cognitive systems is also the reason why the efficiency-unconventionality tradeoff that exists between the DM and SM cannot be extended to the FM. The FM’s implicit processing represents a non-linear break with the logic that less directional heuristics allow for more remote associations.
Incidentally, this is also why flow flows. This feeling of perpetual continuity is generated by three mechanisms. First, flow is a no-search, mapping process in which each step is automated and paired with a controller. Second, flow is a stepwise process since the implicit system cannot anticipate more than one step at the time. Without prediction, each step of the chain depends on the preceding one and triggers the next. This also makes it impossible for the FM to scaffold. Third, flow is free of meta-analytical, explicit interference that could disrupt the performance.
Given that the activity itself must occur in a known problem space, the question arises where in the FM is the creativity. In creating a painting, for instance, an artist’s actual hand movements are nothing new to the motor system, and each stroke of the brush is mapped. Likewise, a jazz musician’s technique must be skillful for improvisation to yield anything creative. Although there are differences between these two flow examples, as the former still allows for revisions while the latter does not, the creative act cannot lay in the steps themselves but must emerge from the overall pattern produced by the entire action-perception sequence (Dietrich, 2015). In other words, a series of individual, uncreative steps can produce a final, creative configuration. This is different from the DM and SM in which a single step can be a creative act (Table 1).