Introduction

Imitation and mindreading are distinctively human social-cognitive skills, which contribute in fundamental ways to being a person. The shared circuits model explains how they can be enabled by functional mechanisms of control, mirroring and simulation. It is strongly influenced by Prinz’s work on common coding of perception and action and by Gallese’s shared manifold hypothesis. The point of the model is partly to unify a wide body of evidence and theorizing about social cognition, and partly to illustrate the philosophical view that embodied cognition can emerge from active perception, by avoiding a ‘classical sandwich’ architecture in which central cognition is insulated from the world between twin buffers of perception on the input side and action on the output side (Hurley 1998, 1991). 

Philosophers distinguish the description of contentful actions and mental states attributed to persons from subpersonal descriptions of information being processed and passed between subsystems (Dennett 1969, 1978, 1991; McDowell 1994; Philosophical Explorations III:I, 2000; Hurley 2003 on a similar distinction for animals). The mental lives of persons depend on and are enabled by subpersonal information processes, though the latter need not correspond directly to people’s conscious mental processes or reasons. Subpersonal processes can be described functionally, or in terms of their neural implementations. Two related types of question arise about personal/subpersonal relations:  1) How are specific personal-level capacities in fact enabled by subpersonal-level processes? 2) What kinds of subpersonal processes could possibly do the enabling work?  That is, are there isomorphism constraints between the levels? Views about latter questions can influence answers to the former.

The shared circuits model addresses the second type of question about social cognition, using subpersonal resources associated with an active perception/embodied cognition approach and without assuming interlevel isomorphism. The model is described at the subpersonal functional level, not the personal or neural levels, though it aims to show how certain personal-level capacities could be informationally enabled and to raise empirical questions about neural implementation. It is a higher-order theoretical model that provides generic heuristic materials for framing more specific first-order hypotheses, for example, about how specific layers map into phylogenetic or ontogenetic stages; it also makes some general predictions. While the model predicts neural mirror systems, it’s a further question exactly how mirror neurons contribute to implementing it.

The model describes how associations that underwrite predictive simulation of the effects of an agent’s own movement, for instrumental control functions, could also yield mirroring and ‘reverse’ simulation of similar perceived actions by others. Mirroring allows means/ends associations with instrumental control functions to be accessed for simulative functions bilaterally, so that causes of observed actions can be simulated, as well as effects of intended acts. Such bilaterally accessible simulations of instrumental structure can provide information for various forms of social learning and for understanding the instrumental acts of others. The ends/means distinction isn’t absolute;  ends/means links can form chains. Multiple instances of the shared circuits structure can be linked to enable the flexible recombination of means and ends in instrumental control, in imitation, and in understanding others’ actions. The model shows how a shared information space for action and perception can be the basis of a shared information space for self and other, and of self/other and possible/actual distinctions. (Elsewhere I consider how the model can be adapted to perception of expressive actions like facial expressions and to emotional understanding.)

‘Simulation’ here has a generic sense applying both to simulation of effects (in predictive ‘forward models’) and of causes (in mirroring) (see Gallese in Frith & Wolpert (eds), 2003). It implies that the same processes that can generate or result from actual action (‘on-line’) are used to generate related information by producing a disengaged (‘off-line’) version of an associated effect or cause. Simulation uses certain processes to generate related information, rather than theorizing about them in separate meta-processes.

Layer 1: Basic Adaptive Feed-Back Control

The model starts from dynamic on-line control with feedback, which establishes an instrumental association between system outputs and their results: a control system produces outputs that are means to a targeted result. A comparator system for feedback control, such as a thermostat, compares a target signal with an input signal. If they don’t match, it changes the system’s output, tracks the resulting change in the input signal, or feedback, and readjusts output to minimize the mismatch to target. Feedback control is adaptive; output can be adjusted to compensate for changing exogenous influences, to bring sensed input closer to the target. Given different exogenous influences, different output is needed to achieve the target; when the weather changes, a thermostat adjusts heat output to maintain a target temperature.

Net sensed input results from the system’s own output plus independent environmental influences. In organisms, reafferent feedback carries input to the system resulting from the organism’s own activity; exafferent input results from exogenous events. Reafference includes visual and proprioceptive inputs resulting from movements of one’s hands, movement through space, manipulation of objects, etc. Exafference from environmental events includes visual inputs resulting from movements by other creatures in a social group. However, at layer 1 information for distinguishing reafference from exafference is not yet available.

The control process is cyclical and dynamic, with no nonarbitrary start, finish, or discrete steps; input is as much an effect as a cause of output (Powers 1973; Marken 1975,2002). In this sense, there is a shared information space: control depends on dynamic relations among inputs and outputs, but information about inputs is not segregated from information about outputs. This shared information space will be preserved and extended in further layers; if perception and action arise out of a system with this feature, they share a fundamental information space.

Means/ends associations can be chained:  output A leads to controlled result B; B can be viewed in turn as an output that leads to controlled result C; and so on. There are independently determined evolutionary, developmental, and individual differences in the grain and complexity of the potential control hierarchies of different creatures.

Layer 2: Predictive simulation of effects

Real time feedback can be slow and produce overshooting, as when a room is slow to warm up after the heat is turned up. Control functions can be speeded and smoothed by adding predictive simulations to the comparator system:  instrumental output-result associations can be activated predictively, simulating the effects of certain outputs for informational purposes (as in efference copy, or forward models). Over time an association is established between output and subsequent input in certain contexts, so that a copy of the motor output signals comes to evoke the associated input signal.  This process simulates feedback--predicts the consequences of output on input. Prediction can occur during actual action, to smooth a behavioral trajectory anticipating feedback, or prior to action, to provide information about alternative possible actions (see layer 4).

A control system is improved by predictive simulation partly because it no longer needs to wait for actual feedback (Miall 2003). The thermostat can switch off the heat before the target is achieved, and avoid overshooting; hand movement can be initiated in accord with predictions of retinal signals based on eye movement.  When real and simulated results don’t match, a local switch can default back to actual feedback control while the predictive simulation is fine-tuned to improve subsequent predictions. For instrumental control purposes, the system doesn’t need to monitor continuously or to access globally whether it’s using actual or simulated feedback.

Comparisons can now be made not just between a target and the actual results of action, but between anticipated results and actual subsequent input. This provides the resources to distinguish reafference from exafference:  information about goal-directed behavior of the organism from information about events in the world. This subpersonal information could be part of what enables the personal level distinction between action of the self and perception of the world. If so, perception and action would share the information space and processing resources described in layer 1; information about perception and about action would both depend on dynamic relations between inputs and outputs  (Hurley 1998, 1991). However, the system does not yet provide information about similarities between the agent’s actions and similar actions by others; nor does it provide information for a distinction between the agent’s actions and the actions of other agents (as opposed to: between the agent’s actions and the rest of the world at large). This suggests that that a self-world distinction associated with instrumental control functions can be available to creatures who lack intersubjective information or mindreading capacities.

A functional architecture like that of layer 2 predicts that cells or cell populations that mediate associations between copies of motor signals and actual input signals might acquire both motor and sensory fields.  Suppose an animal typically acts in a certain way on the perceived affordances of a certain kind of object:  eating a certain food in a certain way, eg. Copies of motor signals for the eating movements will come to be associated with a multimodal class of exafferent and reafferent inputs from such objects and the agent’s eating of them. Cells that mediate this sensorimotor association could thus have both sensory and motor fields, which between them capture the affordances of the objects. Canonical neurons are candidates for such predicted sensorimotor affordance neurons (Rizzolatti 2005; Miall 2003).

Again, the order of layers is heuristic and does not necessarily represent the order of evolution, development, or learning. For example, it may be natural to understand prediction of feedback in an instrumental control context, but in some cases prediction may precede instrumental control (Flanagan et al 2003; thanks to Marco Iacoboni here). The layers can to some degree be re-ordered to frame hypotheses for particular contexts. However, the progression from layer 2 to layer 3 cannot be reversed in the shared circuits model, since the model explains mirroring as an exaptive reversal of predictive simulation. Whether this functional progression maps in explanatory ways onto data about phylogenetic or ontogenetic progressions is a matter for first-order theorizing (though I mention suggestive pieces of evidence along the way).

Predictions deriving from layer 2 include:  Neural mechanisms implementing sensorimotor affordance associations (canonical neurons?). Associations between deficits in predictive simulation functions and in information distinguishing self from world and action from perception (see Frith 1992, pp. 81-3, 93). Information distinguishing self from world prior phylogenetically to information for social learning and understanding other agents.

Layer 3: Mirroring for Priming, Emulation and Imitation

The model now postulates that instrumental output-result associations can be activated bilaterally, from effect to cause as well as from cause to effect. As well as copies of motor signals predictively simulating input signals, as described so far, input signals can also evoke motor signals. (Gallese and Goldman, 1998, p. 498, suggest something like this reversal; thanks to Vittorio Gallese for discussion here; a related reversal is also suggested in Blakemore and Decety 2001, p. 564.)  So when an action is observed, motor activations occur in the observer that would generate a result similar to the observed action.  Observed actions are thus mirrored in the observer. (‘Mirroring’ is used here as a functional rather than neural description, of a default behavioral tendency produced by observing similar action;  functional mirroring depends on implementation in neural mirror systems.)  Mirroring at various grains or levels of control would enable various forms of copying capacity, as observed across various social species:  mirroring of more basic movements in response priming (cf. Rizzolatti’s “low-level resonance”), mirroring of goal-directed action or emulation (cf. Rizzolatti’s “high-level resonance”), and even full-fledged imitation (if the mirrored elements of a control hierarchy are sufficiently articulated and flexibly linked to provide the information needed for social learning of novel means to a goal).

Mirroring makes information about the instrumental structure of action accessible bilaterally, both from observing others act and from intending to act. But it does not itself distinguish the agent’s own action from observed action; it provides what Gallese calls a primitive intersubjective ‘we’-centric information space, or what Prinz describes as a ‘common code’ for action and perception of action. Mirroring brings together information about the basic causes of one’s own and others’ similar observed actions, so that informational dynamics of instrumental control can predictively simulate further effects for both one’s own and others’ similar actions. By uniting information about own and others’ acts, mirroring allows existing means/ends associations to be simulated bilaterally for both:  that (observed) act retrodicts this (own) motor activation via mirroring of causes, which is associated with further results via simulative prediction of effects.  However, this does not yet provide a distinction between own acts and others’ acts. In this sense, intersubjective information is subpersonally prior to the self/other distinction (as Gallese holds).

How could such mirroring arise?  Consider first movements that produce visual reafference for their agent. When a creature watches her own hand movements, an association is formed between copies of motor signals for such movement and visual reafference from such movements. Cells or cell populations mediating this association can acquire similar sensory and motor fields. If the first creature watches another perform similar hand movements and receives similar visual inputs, these will also activate her sensorimotor ‘matching’ cells with their motor fields. Such a functional architecture predicts neural mirror systems:  that cells or cell populations mediating this association would acquire similar sensory and motor field and that their sensory fields would not distinguish the observer’s own action from others’ similar actions. Evidence for neural mirror systems is strong and growing (though more work is needed to understand what computational work neural mirror systems require of individual mirror neurons).  They would provide the neural implementation of intersubjective information for functional mirroring--a shared information space for self and other in that information about own actions is not segregated from information about others’ similar actions. Note the intimate relationship between the sharing of circuits for action and perception and for self and other:  the shared intersubjective information space of layer 3 is a specification of and presupposes the shared information space for perception and action of layer 2, which in turn builds on the generic sensorimotor information space of layer 1.

Consider next how mirroring might arise for movements not seen by their agents. How can a correspondence be established between one’s own acts and others’ similar acts, without reafferent feedback from one’s own actions in the same modality as observations of others’ acts?  For example, facial movements would normally produce proprioceptive rather than visual feedback; how can this be associated with visual information about similar observed facial movements?  Several answers are compatible with the shared circuits model. Perhaps some supramodal correspondences are innate (as in newborn copying, Meltzoff, 2005). Perhaps some are acquired through experience with mirrors, or with being imitated (Heyes, 2005). Or perhaps associations between one’s own and others’ similar acts can be established via stimulus enhancement (see Heyes 2005 for closely related suggestions). Suppose a creature repeatedly sees conspecifics act a certain way; their actions draw its attention to the typical objects of their actions, and these evoke in the observer an innate or otherwise acquired response to those objects. A resulting indirect association is formed between visual observations of others’ actions and one’s own similar action. This is not initially copying; the object independently evokes others’ and one’s own acts. But the indirect, object-mediated association between own and others’ similar acts may become direct with repetition. Cells that mediate this association could thus acquire similar sensory and motor fields, so that observing another’s act primes similar action by the observer.

It may be helpful to clarify the functional distinction between instrumental control and mirroring. Instrumental control functions to achieve a target:  given certain inputs, it selects motor outputs that will in turn produce inputs matching a target. Analogously, if fleeing prey changes direction, its predator also changes direction.  By contrast, while mirroring also produces motor outputs given certain inputs, it doesn’t in itself have an instrumental function. Mirroring doesn’t select outputs as means to an end, but merely for similarity to input: The priming of my movement by observing another’s similar movement needn’t be a means to a target. Rather, on this model, it’s a by-product (via reversal) of predictive simulation; the latter does have instrumental control functions (see layer 2). However, mirroring may in due course be exapted for other functions, such as those associated with imitation, simulation for action understanding, or signalling. These in turn may enable capacities for higher, social (‘Machiavellian’) forms of control. Whether specific mirroring capacities are preserved by evolution will depend on their potential functions for different species facing different evolutionary problems.

The shared circuits model doesn’t describe one all-inclusive structure, but rather can be multiply implemented for various specific movements and results, at various points along means/ends chains (cf. Wolpert et al 2004). Consider a series of control spectra with an ultimate result at one end and detailed movements at the other, and a series of means/ends links between. The model could apply at successive linked points along such spectra, or between them (with the right neural connections); the means to one circuit’s target could be the next circuit’s target. A network of such linked circuits would support hierarchical control and permit flexible recombination of means and ends.

Varying forms of social learning could result:  response priming, emulation, imitation. These will vary across different creatures and over development along with (1) the grain and complexity of instrumental control capacities and (2) which of these have associated mirroring capacities and how richly and flexibly they are linked (see related discussion of Csibra’s paper in this forum):

(1) Different animals are equipped to different degrees with the potential for instrumental control and associated predictive simulations, via potential means/ends chains of different grain and lengths. Animals suitably equipped with control functions by evolution and development could form chains of simulative predictions, resulting in information such as:  this tail movement will have that effect on weight over legs, facilitating that movement trajectory in relation to that gazelle, and so on, with eating the prey further on in the predictive chain. Or, imagine a similar chain from knee movement to winning a slalom race.  Mirroring would operates such associations in reverse; mirroring the cause of another’s movements, or resulting relationship to an object, could enable movement priming, goal emulation, or even full-fledged imitation (if instrumental control and associated mirroring are sufficiently articulated and flexible). Combined with further information distinguishing self from other (see layer 4), simulative mirroring can provide information to enable understanding of others’ observed movements as instrumental actions with means/end structure. The simpler an animal’s control and predictive simulation capacities, and the shorter the means/ends chains it can sustain, the more limited the potential for mirroring and related functions. 

Mirroring associated with control of details of basic movement would predict priming of basic movements; mirroring associated with higher levels of control would predict priming of less basic movements.  Again, the distinction between ends and means is relative, not absolute. When I am learning to use chopsticks, finger movements are a means to chopstick deployment; when I can control them effortlessly, chopstick deployment is a means to sushi eating.  Suppose I eat sushi by deploying chopsticks by moving my fingers a certain way. You watch. If seeing me move my fingers generates motor activation associated with similar finger movements in you, then mirroring results:  you are primed to move your fingers in a similar way. If you can already control chopsticks, mirroring may occur at a less basic level, and prime chopstick deployment. Such priming could be goal-mediated:  your chopstick deployment could mirror mine when sushi eating results (even if no sushi is actually in sight), but not when the results are unrelated to sushi (similar to the well-known demonstration of mirror neurons tuned to movements aimed at apple grasping, even if the apple is hidden behind a screen, but not similar movements in the absence of an apple; Rizzolatti 2005).

Whether a movement is recognized as goal-directed at all may depend on, among other things, the instrumental capacity that is potentially available for simulative and mirroring functions. Capacities to simulate and mirror could provide information about the goal-directness of certain observed movements given relatively fine-grained, complex means/ends associations but not given relatively crude control capacities. Use of instrumental associations in mirroring can provide information about specific goals of observed movements, but mirroring doesn’t itself determine which instrumental associations, or predictive simulations thereof, are in place and hence potentially available for mirroring.

(2) Control processes may be neurally distributed so that components of an articulated mean/ends chain are processed in different areas—control of fine movements vs. gaze vs. posture vs. whole body movement vs. external objects, etc.  Some such neural areas may have mirroring as well as simulative predictive capacities, while others don’t; this will vary across species and development, also producing different capacities to derive information about whether various observed actions are goal-directed.. Prior (evolutionary or developmental) processes determine whether mirror systems are associated with particular control capacities, and thus which predictive simulations can be mirrored. Those which can be, can also provide information about specific goals of observed actions.

Imitative learning is a phylogenetically rare capacity requiring flexible associations between ends and means. The degree of articulation and corresponding linkages between mirror systems may determine how flexibly mirroring can be adapted for imitative learning. An animal that can mirror at various points along a chain of means/ends associations may never have used a certain means to a given goal. But if he observes and mirrors another’s novel way of achieving that goal, and neural links exist so that specific mirroring activations can be associated flexibly with targets, that goal and those mirrored means may be newly associated, enabling the creature to learn imitatively. You might learn to use chopsticks by watching me; sheer trial and error would be bettered by mirroring that primes your finger movements towards those you see me make, which are associated with the target of effective chopstick deployment. Imitative learning should be found in fewer species than either response priming or goal emulation separately—as is the case--since it additionally requires linkages supporting flexible instrumental associations between mirrored means and ends at reasonably fine grain.  

Flexible links between multiple mirror circuits for means and goals would enable imitative learning by capturing information about novel instrumental structure in observed action. Flexibly linked mirror circuits could generate behavioral building blocks combined in program level imitation of sequential or hierarchical structure (Byrne 2005; Whiten 2005).  They could allow an infant to form three-way associations between observed behavior by its parents (who have survived to reproduce, so may have adaptive behaviors not all of which are heritable), observed circumstances in which its parents perform such behavior, and its own similar behavior, supporting contextual imitation:  act like that, when the environment is like this (Byrne, 2005).  Given a capacity for monitored inhibition of the observer’s own action, information about the instrumental structure of observed action can be used in understanding others’ actions through simulative mirroring (see layer 4).

Predictions deriving from layer 3 include:  The existence of neural mechanisms that implement mirroring functions, which do not distinguish own and others’ actions, associated with neural systems that implement comparator control and predictive simulation functions (involving mirror neurons?). Automatic behavioral mirroring tendencies at varying grains, as permitted by the complexity and articulation of control functions across species and development (as in response priming, goal emulation, human infant copying, human perceptual induction effects, imitative interference and reaction time effects, chameleon effects). The phylogenetic rarity of capacities for imitative learning as opposed to response priming and goal emulation. Associations between deficits in predictive simulation control functions and mirroring functions, reflecting shared circuits for these functions.

Layer 4: Monitored Output Inhibition and Simulative Mirroring

Layer 4 introduces the capacity to activate instrumental associations while inhibiting actual output and while monitoring the inhibition of output. This capacity could be added to layer 2’s predictive simulations, to layer 3’s mirroring, or both.

Consider first how layers 2 and 4 would combine functionally. As described at layer 2, simulative predictions were made during on-going action, improving on-line control of actual action. For this function it isn’t essential that the system monitor whether it’s currently using actual or simulated feedback, as long as the target is achieved. However, simulative predictions of results could also function off-line, with actual motor output inhibited. Multiple simulative predictions could provide information about the results of alternative possible actions, rather than anticipating results for ongoing action. The simulated results of alternative possible actions could be compared with a target prior to actual action, providing information about which produces the closest match. Such information could enable decision-theoretic intelligence, instrumental deliberation and planning.

However, enabling these further capacities would require not just comparing the simulated results of different possible acts with a target, but also monitoring whether motor output is inhibited, so that the distinction between possible and actual actions is tracked. Layer 4’s added capacity for monitored inhibition provides a basis for this distinction:  simulated results given output inhibition would provide information about possible actions, simulated results without output inhibition would provide information about actual actions. So, multiple predictive simulations could provide information about the consequences of alternative actions by the agent, while monitoring of output inhibition could provide information that such actions are possible rather than actual.

Consider next how layers 3 and 4 would combine functionally. A creature observes another’s act, which primes a similar act by itself through mirroring. However, its own similar act is inhibited; the observed behavior is not actually copied. While various forms of behavior copying may have beneficial functions, unselective overt copying of an action could sometimes have disastrous results for the copier; we should expect development of a capacity to inhibit mirroring. When used off-line, mirroring simulates in the observer the causes of the observed action. Simulative mirroring operates in the reverse direction from simulative prediction; instead of simulating the feedback that would result from motor activations, mirroring simulates motor activation that would produce results similar to those observed, with actual motor output inhibited. Simulative mirroring could thus provide information for understanding the instrumental structure of observed actions (and perhaps for emotional empathy—a further matter I consider elsewhere). At layer 3 it was noted how the length and grain of chains of means/ends associations and flexibility of linkages between them should affect the character of copying behavior enabled by mirroring. These factors will similarly affect the kinds of understanding of observed actions enabled by simulative mirroring.  Goal-mediated simulative mirroring would provide information about goals of others’ observed movements, enabling an early stage in understanding others as intentional agents (hence in mindreading). If mirror circuits are sufficiently articulated and flexibly linked to enable imitative learning, then monitored inhibition of imitative mirroring would capture the instrumental structure of novel observed action more fully and flexibly, enabling more sophisticated mindreading.

For simulative mirroring to enable understanding another’s action would require not just simulation of means/ends associations for observed actions, but also monitoring whether the motor output of mirroring is inhibited, to separate information about the actions of others from information about one’s own actions. At layer 3, own and other’s actions are not distinguished. But use of information about actions to understand others has different consequences, and makes different demands on subsequent behavior, than use of information about actions to copy them. The added capacity for monitored inhibition of mirroring at layer 4 provides an informational basis for the distinction between self and other agents, which comes to overlay the shared information space for own and others’ actions at layer 3:  simulative mirroring with monitored inhibition would provide information about another’s action rather than about one’s own. So, simulative mirroring could provide information about the causes and instrumental structure of observed action, while monitoring of output inhibition could provide information that such actions are another’s, not one’s own.

Whether or how such subpersonal informational structures correspond to the personal level sense of being able to do otherwise, or of empathy with others’ goals, or of being the agent of an action, are further questions. The present point is to show how aspects of the personal level could be informationally enabled by subpersonal resources described by the shared circuits model.The actual/possible and self/other distinctions are necessary (if not sufficient) for much explicit theorizing and for aspects of the normativity and intersubjectivity that characterize the personal level. Understanding how information for these distinctions can arise subpersonally helps to understand how subpersonal processes can enable the personal level. The shared circuits model suggests that these two distinctions have a common informational basis in monitoring of motor inhibition; in this way, theoretical informational resources arise from practical.

In particular, layer 3’s subpersonal ‘first person plural’ is informationally prior to layer 4’s distinction between self and other (as in Gallese’s view).  At the level of subpersonal information, the problem of ‘knowledge’ of other minds is reconfigured:  it’s neither one of starting from information about the self and constructing a bridge across a gulf to information about other persons, nor one of starting from information about other persons and from the resources it provides somehow generating information about the self. The similarity of own and others’ acts comes first, with mirroring. The job that remains is not to bridge a gap between oneself and other agents, but to track distinctions among them, especially when multiple other agents are in play.

Philosophers may be tempted to think that the shared circuits model can ‘provide an account’ of understanding other agents that avoids both behaviorism and the problematic argument from analogy. However, there are further questions about whether and how the subpersonal priority of intersubjective information is reflected at the personal/animal level, questions about phylogeny, development, the structure of a mature capacity to understand other agents, the epistemology of such understanding. The role of subpersonal information in the epistemology of other minds in turn raises philosophical issues about the roles of reliable information and justification in knowledge. The shared circuits model does not itself answer these further important questions. Rather, it provides generic, adaptable tools for framing more specific hypotheses. Care is needed in doing so—we shouldn’t just assume an isomorphic projection from the subpersonal to the personal level, but nor should we assume that the structure of subpersonal information processing has no implications for the personal level.

Predictions deriving from layer 4 of the model include:  Close association between capacities for prediction and inhibition in action control and for understanding others and distinguishing them from self--and between deficits in these capacities (see Frith & Frith 1992, The Neuropsychology of Schizophrenia, pp. 81-83; 93). Dissocations between copying and inhibitory capacities (as in Lhermitte’s imitation syndrome patients). People with intact inhibitory capacities retain underlying copying tendencies that may generate interference or reaction time effects (Prinz; Bargh). Capacities for various forms of copying of others’ acts interleave with capacities for various levels understanding of others, in a bootstrapping process leading to full imitation and mindreading. Copying without inhibition should be found earlier, and capacities to inhibit copying and to distinguish others’ perspectives from one’s own should develop together.

Layer 5: Counterfactual Input Simulation for Strategic Social Intelligence

Finally, the system can be taken off-line on the input as well as output side. Monitored simulation of inputs to control systems equipped with simulative prediction and mirroring functions can provide a distinction between others’ actual and possible acts. Counterfactual inputs can now simulate different possible acts by others and their results. This social extension of counterfactual information combined with simulation of different possible acts by self and their results provides information needed to track how the results of various possible acts by others may in turn result from various possible acts of one’s own, and vice versa. This combination of functions provides enabling information for strategic or ‘Machiavellian’ social intelligence, game-theoretic deliberation, coordination and cooperation.  These simulative informational resources for instrumental and strategic functions may in turn provide a practical foundation for capacities to manipulate counterfactual information and modes of counterfactual theorizing more generally.

At this level, the demands of differentiating and tracking interacting means/ends relations for multiple other agents and multiple possible acts are acute. Meeting these demands, and further demands in differentiating the epistemic states of multiple others, may well require the simulative informational basis for understanding other agents to be supplemented by language-dependent theorizing capacities. Mindreading, like social learning and instrumental control, is a graded phenomenon, not all or nothing (Tomasello 1999). Capacities for interpretative theorizing about others may build on layer 4’s fundamental self-other distinction to fine-tune differentiations and identifications of multiple other intentional agents. The shared circuits model allows that simulation and theorizing may both be needed for mature mindreading with all the bells and whistles, though it describes the foundations of mindreading in terms of simulative mirroring of means/ends relations.

Predictions deriving from level 5: Understanding the instrumental structure of others’ actions is foundational for mindreading, and prior to understanding others’ epistemic attitudes (on phylogeny: Tomasello & Call; on ontogeny: Rakoczy, Warmeken & Tomasello). Mature mindreading may require simulation-based information about the instrumental structure of observed action to be supplemented with language-dependent theorizing, so that strategic and epistemic mindreading capacities develop with language, though understanding the goals of observed action might be present prior to language.

Conclusion

The shared circuits model of social cognition describes a functional progression of layers from (1) comparator feedback control to (2) simulative predictions to (3) mirroring to (4) simulative mirroring with monitored inhibition of output to (5) monitored simulation of input (with some scope for reordering the layers). Each layer provides informational resources that can enable certain personal-level capacities, as well as raising questions of neural implementation (see table). Predictive simulations of ongoing actions improve control of instrumental actions and provide information for an action/perception or limited self/world distinction. Mirroring provides intersubjective information about instrumental action that does not distinguish self and other; it can enable varieties of social learning, empathy, and automatic copying. Monitored inhibition of output combined with simulative prediction provides a distinction between one’s own actual and possible actions and information about results of alternative possible actions for deliberation. Monitored inhibition of output combined with simulative mirroring provides a distinction between one’s own and others’ actions and information about the instrumental structure of observed actions. Monitored simulation of input extends counterfactual informational about actions socially, adding a distinction between the actual and possible acts of others and information for strategic social deliberation.    

This layering of informational resources in the shared circuits model shows how cognitively significant resources (such self/world, self/other and possible/actual distinctions, intersubjective information for social learning and mindreading, and counterfactual instrumental and strategic information) can emerge from a shared information space for perception and action in instrumental control, which is preserved as higher functions are added. Social cognition is enabled as a specialization of active perception:  I perceive your action enactively, in a way that engages my own capacity for similar action, enabling me to copy or understand your action. Shared processing of actions of other and self is a special aspect of the shared processing of perception and action. The distinction between self and other is overlaid on the prior information space they share, which is again preserved as more advanced capacities are built on it. Tracking multiple possible acts of self and multiple others, with their interacting means/ends relations, plausibly requires that simulative informational resources be supplemented by language-dependent theoretical resources. However, capacities for manipulating possible/actual distinctions and counterfactual theorizing more generally may, like mindreading, have practical foundations in simulative functions.

The shared circuits model has three noteworthy aspects. First, it distils a shared information space for self and other from a shared information space for action and perception, based on control processes. Second, it shows how distinctions important to enabling the mental lives of persons--between action by self and perception of world, between self and other, and between the actual and the possible--can be overlaid on these prior shared information spaces. Finally, the model avoids the traditional conception of cognition as sandwiched between separate, peripheral systems for perception and action. Instead, it develops the implications of an active view of perception for the perception of action, showing how informational resources for embodied social cognition can be built on those for active perception.

Full references are omitted for reasons of space but can be provided in web discussion; or see ‘Active perception and perceiving action’, in press and at www.warwick.ac.uk/staff/S.L.Hurley; and S. Hurley & N. Chater, eds., Perspectives on Imitation, 2 vols. MIT Press, 2005.

Click to visualize a diagram summarizing the shared circuits model