Spring

12

Mirror Neurons In Motor and Social Acts
The authors who ushered in what might be considered the most critical finding of 21st century neuroscience to date—the discovery of mirror neurons (MNs)—contend that MNs underlie simulation theories of action understanding and mindreading (Gallese & Goldman, 1998); Gallese et al., (2004). The mechanism that allows mirroring of action is referred to as the ‘direct-matching hypothesis,’ (Rizzolatti et al., 2001), and the claim is based on the response properties of MNs in monkeys. The discovery of MNs is important in that if the implications of their function are properly understood, it lays a foundation for explanations of not only understanding of action and other minds (Gallese, 2003), but for other big questions, such as the evolution of language (Rizzolatti & Arbib, 1998). The concern of this paper is MN function as a feature at the core of the stronger claims. Alternative theorists challenge the characterization of the MN role as ‘mirroring’ action, as their findings show that mirroring cannot be direct and must rely on an interpretation of the observed action. They claim: a) MNs do not simulate with enough precision to be referred to as ‘mirroring,’ in which they are said to simulate an agent’s motor program onto the very same motor repertoire of an observer not performing the act; b) mirror neuron activation is predictive of action, not replicative; and c) the function of these neurons supports a model that is generative of top-down action understanding that is emulative, and formed outside the motor system, not embodied in the simulation of the motor act (Csibra, 2007). While attributions made for MNs are compelling in the philosophical accounts by Gallese (2001) and others, on close inspection unequivocal empirical evidence is sparse. The alternative model proposed has unanswered questions for the human mirrormatching theorists. Yet mindreading, which requires identifying and understanding others as similar to ourselves is heralded as the key to our evolutionary success (Tomasello, 2000).

Introduction An essential pre-condition to successful social interactions is the intrinsic capacity to recognize other-intentions (Sartori et al., 2011). The claim that we do so through observing the movements of others is the prevailing view, and yet gives rise to a multilayered controversy. The dominant view purports the function of mirror neurons (MNs) underlies third-person mindreading; ‘mindreading’ being the cognitive capacity to represent psychological states of the self and other, such as desires, beliefs, intentions, and feelings. There are three essential components to this view: ‘strict motor resonance,’ which posits that MN brain activity resonates between agent and observer, creating precisely matching neural motor patterns as they interact; ‘internal forward models’ which generate a postdictive observer understanding of agentintentions; and ‘embodied simulation’ wherein the intentions of others are gleaned by the observation of the movement of an agent. These advocates of intention-embedded motor processes deny the computational-representational view of human cognition as an instrument that transforms symbols using syntactic rules. Contesting this view are proponents of a contrasting, alternative model who instead propose that the task of MNs is to assess a prior representation of an agent’s intention and to predictively compute the motor command best suited to ensure the agent’s goals are reached (Jacobs, 2008). In this paper I will restrict the focus to the essential, underlying question at hand: how does an MN function? Solving this illuminates the inquiry into the further questions of if and how MNs contribute to our understanding of other minds. I will start by giving a brief background on MNs and the context for the opinions, then present the two opposing views in condensed form, and from there consider the supporting evidence on the points of disagreement. Ultimately I conclude that new research is needed; research that is inspired by new theoretical models, and seeks to falsify MNs as strict replicators, if only to better understand their nature.

2

View 1 (Vittorio Gallese, 2011) The prevalent view of how we understand the intentions of others is rooted in action as a relational concept. In this view, in the forward model of motor action, MNs are said to source an efference copy—an internal copy of the signal of a motor program—to estimate sensory outcomes of an action. In an agent’s brain canonical neurons (object-activated) translate features of an object that the agent wishes to act upon (e.g. a
Image: Csidra, 2007

bottle he wishes to grasp) into the best-fit motor program to

execute the action. MNs are activated by an efference copy of the motor plan and simulate the action program to ensure the intended goal is reached. As we observe the agent’s actions, our neural resonance mechanisms model the observed behavior for us internally—matching the motor commands to those we recruit when we ourselves perform the action—mapping the same neural substrate we use to do so. The MNs retrodictively create a representation of the agent’s intent (Gallese, 2011: 40-41). View 2 (Pierre Jacobs, 2008) Alternatively Jacobs (2008) theorizes that in line with inverse models of action, MNs in both an agent and observer’s brain will compute the motor commands to achieve the agent’s prior goals and intentions. In the agent this is done from an internally generated representation of the prior intentions, while
Image: Csibra, 2007

in the observer the representation of prior intentions is derived from perceptual cues. In this view visual analysis makes a contribution to interpreting actions and sequences, as shown by cell activation in the superior temporal sulcus (STS), which has perceptual but no motor properties.

3

As Csibra (2005) points out, prior to being transformed into motor code, in the visual system mid/high level interpretations of observed actions may contain goal assumptions that get incorporated in the motor code. Likewise actions may be filtered at this level into units of relevancy (e.g. non-grasping hand). The STS responds not only to body movements but important eye and head motion, including the target of visual attention (gaze) for which to date no MNs show a response. So the STS, reciprocally connected to the inferior parietal lobe (thus having a path to the pre-motor cortex), in this view is thought to make a contribution to generating an inner representation of an agent’s intentions. Rather than a retrodictive ‘action mirroring’ process that recovers the intention from coded action, Csibra (2005) calls this ‘action reconstruction’ as here the MN predicts the motor command needed to get to the goal.

In the Research - View 1 Between 1985 and 2000 a string of single-neuron studies in monkeys revealed a discrete subset of pre-motor and parietal neurons, dubbed mirror neurons, were activated both when the animal executed a motor act and when it observed the same act performed by others (Gallese, 2001). These were designated ‘mirror-neurons,’ requiring an action-object interaction (by monkey or human) in order to be discharged. The reasoning that led Gallese and Goldman (1998) to the view that MNs might underlie mindreading began with this discovery along with additional data from some other studies. For example, Fadiga et al. (1995) demonstrated that the observation of hand actions performed by others enlivened the same neural hand and arm substrates that were used when observing agents performed the same.

4

Two Computational Systems are presented: one that receives information about an action and outputs a goal for it (action- to-goal), and another one that generates the right action for the goal emerging at its input (goal-to-action).

View 1 (embodied)
Action to Goal
Auto-­‐‑Pre-­‐‑Reflexive Modeling View(Strict motor resonance adherence model) Mirror Neurons simulate a representation of the agent'ʹs intention from an a priori representation of the agent'ʹs motor command in accord with an internal forward model. • How? • Canonical neurons translate features of the target into the most suitable motor program enabling a successful action to be produced. • An efference copy of this signal is fed to mirror neurons which faciliate a simulation of the action program. • This simulation of the action is used to predict its consequences (goal). • MNs in an observer’s brain is taken to constitute an automatic mental simulation of the agent’s observed movements, enabling the observer to match the agent’s observed movements onto her own motor repertoire (without executing the movements in question). • The activity of MNs is seen as enabling the observer to represent the agent'ʹs intentions • Simulation Postdictive • Learning Imitative

View 2 (cognitivist)
Goal to Action
Computational-­‐‑Representational View
(Uses logical chains of related MN activity) Mirror Neurons compute a representation of the agent’s motor command from an a priori representation of the agent’s intention, in accordance with an internal inverse model. • How? • Neurons in the STS respond to a vast array of eye and head movements and full body movements (including locomotion), a purely perceptual network of ‘social perception’ involving the STS, the amygdala and the orbito-­‐‑ frontal cortex • The representation of an agent’s prior intention, might arise in an observer’s brain as a result of this activity of the STS, which is reciprocally connected to the inferior parietal cortex, and is known to contain neurons with purely perceptual and no motor properties. • MNs in the observer’s brain predictively compute the best motor command suitable to satisfy the agent’s intention. • Simulation Predictive • Learning Emulative

Figure 1 Distillation of the two approaches. View 1 - Gallese, 2011, View 2 - Jacobs, 2008

5

Buccino et al. (2001) used fMRI in a study that showed that the observation of hand, mouth or foot during both transitive and intransitive actions created a corresponding engagement of somatotopic premotor cortex regions in the observer. This did not occur during observation of the same but motionless body parts. Corroborating this evidence that intentions or goals are gleaned though observation of motor action, distinct mental resonance mechanisms were revealed in two contrasting pathological examples. ‘Echopraxia’—the impulsive imitation of others’ movements—demonstrates that strict imitation uninhibited by thought may be possible. ‘Imitation behavior,’ a condition resulting from a lesion to the orbito-frontal cortex, led to the imitating of a goal by using an identical action rather than the one observed. This showed that a goal could be deduced from observation of motor action (Gallese, 2001). The hypothesis described as View 1 above was in part motivated by the need for a more comprehensive account for understanding “intended, mind-driven behavior,” as the researchers do not find that the visual coding of a stimulus accounts for it (Gallese, 2001: 35). A problem for visual analysis has been the question of how the neurons involved bind with sufficient complexity to promote understanding of the meaning of an action. By extending the visual hypothesis—a visual analysis of elemental action formation, with no motor involvement—to include motor synergistic action with sensory binding, a basis may be formed for action understanding, called the direct-matching hypothesis. Here, higher-order visual neurons combine the output of neurons responsive to both the observation of an arm reaching and the direction of gaze. For example, in studies with monkeys, the neurons activate when a monkey observes an actor reaching, but only when the actor’s gaze is also the apple (target) being reached for. Beyond helping code motor acts, MNs enable imitation in this view, and as imitation normally

6

happens for social reasons (e.g. for learning), the mirror system is further suggested as underlying other cognitive strategies, such as mindreading (Rizzolati, 2001)). To compound the earliest findings, a study by Meltzoff (1995) had adults execute and display goal-directed acts so that infant toddlers of eighteen months could re-create them. An example from the study was of an adult demonstrating trying to pull apart a toy dumbbell, but without success. The toddlers were able to re-enact the failed attempt. However when the children observed a mechanical device attempting to pull apart the dumbell they were unable to duplicate the effort. From this authors took it that in order to understand and mimic the intended goal of an observed action, a link must exist between the observed and observer. The proposal by these researchers is that this link is formed by the embodiment of the intended goal in the action, shared by the two parties. While the data from these experiments is considered valid, it is the construction of the View 1 hypothesis that has been criticized by the alternative theorists. Primarily this is because further studies have revealed evidence that questions the three primary assertions required to operationalize the view. In sum, the authors of View 1, the most widely accepted view, propose that the fundamental function of MNs is mindreading. This stronger claim represented by the direct-matching hypothesis leads to philosophical questions about whether a belief can be formed about another’s psychological state, via an experiential event that is not mediated by mental activity. Can a non-conceptual grasp of another’s mind (auto-simulated) even be categorized as third-person mindreading (Jacobs, 2008)? These claims have led to evidentiary disagreement of the three main features for the View 1 model by advocates of View 2. As discussed below these are: strict motor resonance, retrodictive representation in the internal forward models, and embodied simulation.

7

In the Research - View 2 I. Strict Motor Resonance Based on earlier evidence the most fundamental aspect of MNs was presumed to be the strict congruence between their motor and perceptual properties. Gallese and Goldman (1998: 498) had surmised that because MNs are a resonance tool they represent a neural basis and phylogenetic predecessor for mindreading. Initially, the same MNs had been found to fire both during execution and observation of an act of grasping (Jacobs, 2008). Several studies took place in 2005 with the goal of extending MN activity beyond recognizing the motor act to representing an agent’s intention. While this connection was made, the studies also brought new evidence to light in two primary areas seen as inconsistent with a strict resonance model of perceptual and motor properties. These are addressed below as: i) The Hierarchical Structure of Intentions; and ii) The Model of Chains of Logically Related MNs. i. Hierarchical Structure of Intentions A hierarchical structure of intentions distinguishes between a basic and non-basic act (Jacob, 2008, citing Goldman, 1970). The activation of MNs depends not only on the presence of goals, but may also fire preferentially according to the further, higher-level goals of an agent. In a Fogassi et al. (2005) study where monkeys were trained to perform two actions, some MNs showed more activation when a monkey grasped an object “to eat,” as compared to grasping the object “to place” into a container. Depending on the soon-to-be-taken action separate sets of MNs in the inferior parietal lobule fired. These neurons responded similarly when the monkey observed the same actions performed by an experimenter. Yet the first segment of the actions, the grasping, was shown to be without demarcation in kinematics—geometric motor positioning—indicating the MNs take into account the further goal and not just the perceived action (Csibra, 2007). In an fMRI experiment of like design but with humans, Iacoboni et al.

8

(2005) found that the MN firing in the right inferior frontal areas showed a far stronger bias for the “grip to drink” over the “grip to clean” condition. According to Searle (1983) agents have networks of “nested” intentions. Turning on a light when you want to read a book and don’t believe you have enough light is a non-basic act, as it will be attained through a series of basic discrete acts. Further, since your belief isn’t an actual fact, intentions can’t be anticipated. Therefore intentions are in relation to motor acts on a many-to-one basis. Perceiving a goal is more abstract that acting on a target. Jacobs (2008) discredits strict motor resonance finding it “lacks the resources to bridge the gap between representing an agent’s motor intention and representing her higher-level intentions” (Jacobs, 2008: 208). ii. Model of Chains of Logically Related MNs Rather than sheer replication (strict congruence of motor resonance), MN thinkers now subscribe to the idea of chains of ‘logically related’ MNs. As Iacoboni et al. (2005: 533) stipulate, the relation between pairs of MNs in a chain is inductive, not strictly logical, saying,
“... the present findings strongly suggest that coding the intention associated with the actions of others is based on the activation of a neuronal chain formed by mirror neurons coding the observed motor act and by ‘logically related’ mirror neurons coding the motor acts that are most likely to follow the observed one, given the context.”

Fogassi found 65% of MNs during grasping tasks are modulated by the more complex action of which the movement is only a part. The experiments do not show that motor contagion underlies action understanding, which as described is the representation of an agent’s prior intention. Rather than replicating the observed motor act we find MNs in the role of allowing the observer to predict an agent’s next act by combining the observation of the motor act and contextual clues (Jacobs, 2008). Csibra (2007) finds the most compelling evidence against strict motor resonance is the fact that there are only degrees of similarity between the act and observing the act (ie. when a monkey acts vs. when it observes an experimenter in the same act). In a review of the

9

studies, Csibra (2007) found the following: a) 19-41% of MNs—about one-third—show a oneto-one congruence between visual and motor cell properties for action category (e.g. grasp) and manner of execution (e.g. grasp type); b) 21-68% of MNs respond to two or more types of observed actions (e.g. grasping by hand or mouth), also termed “broadly congruent” MNs which make up the bulk of all MNs; and c) 10% of MNs with no relation being found between the two. Based on this meta-evidence, if monkeys relied on direct matching of action to goals, they would consistently misapprehend what they were seeing. This low visuo-motor congruence in the research can perhaps better be explained by action prediction which does not require strict replication, while extracting goals from action coding does. II. Retrodictive Representation Internal model systems are a critical concept to understanding motor control system

design and operation. Basically there are two types, forward and inverse models, which work together to compute the input that a system requires in order to attain the sought-after output. Forward models estimate the contribution needed to move to the next state from the current one, summing up the causal relationships amongst all the system inputs. By contrast, inverse models deliver the final motor command that will enact the change, controlling the desired transition (Wolpert et al., 1998). Gallese’s hypothesis has been framed using the logic of forward models. These are defined as ‘forward’ because they can capture the causal relationship of actions, as signaled by an efference copy, and from there estimate outcomes of motor commands. Essential to the success of the View 1 scenario, this idea conflicts with the typical flow of internal forward models for motor control. i. Internal Motor Models Essentially, forward models compute the sensory consequences of an action from a representation of the motor command while inverse models compute motor commands sufficient

10

to enact the agent’s prior intention (Jacobs, 2008). Blakemore and Decety (2001) hypothesize that an observer could use a forward internal model of action to retrodict an agent’s intention from his observed movements. Yet Csibra (2007) found that the retrodictive account has two difficulties when positioned in the forward model: first, it reverses the flow of information that is characteristic of computations executed by this model; and second, the account under-represents the complexity of computation needed to map an underlying intention from an observed movement. As discussed, the same movement can be applied toward many intentions. Of primary consideration is that an observer does not access an efference copy of the agent’s motor command as the agent does. So in this view the claimed tension is that two separate activations—the internally generated activation of MNs in an agent’s brain and the externally generated activation of MNs in an observer’s brain—make a contribution to the agent’s goal representation. In this event, the computation would therefore be predictive in the agent’s case and retrodictive in the observer’s case. So it becomes difficult to reconcile this asymmetry in a unified model. Jacobs (2008) asks us to resolve whether MNs function predictively as the new model of logically related MNs would imply, or whether MNs function retrodictively (in accordance with Gallese and Goldman’s views). In the case of non-human primates, researchers stress the value of anticipating a conspecific’s next move so the observing animal can prepare a reaction (Gallese and Goldman, 1998: 495-6). They don’t focus on the importance of retrodicting the conspecific’s former intention nor do they address the inhibitory skills that would be required for retrodiction, which has not yet been established in non-human primates.

11

III. Embodied Simulation View 1 authors distinguish between an engaged and a disengaged understanding of an agent’s action based on motor resonance. But as Jacobs (2008) points out, it is quite a jump from there to making the further—and much bigger—claim that motor resonance also automatically reveals the contents of an agent’s mind. View 2 proponents fault the latter assertion because simply matching an observed movement onto one’s motor substrate is not enough to know what an agent has in mind, nor have experimental results shown that MN activity generates an account of an agent’s prior intention. To accord with forward models of action it is proposed that an agent is able to predict the sensory consequences of his soon to be realized action via an efference copy of his own motor instructions. Yet an observer, unlike the agent, does not receive an efference copy of the agents agent’s motor program. So how could embodied simulation enable activity of MNs in an observer’s brain to yield the goal of the impending movement of an acting agent? i. Perceptual or Contextual Clues Researchers Jacob and Jeannerod (2005) and many others use evidence from developmental psychology to show that an observer is able to understand an agent’s prior intention without first mapping an agent’s motor command to her neural substrate. In experiments children from nine months to adult will assign and describe the emotions and social goals of a simple moving geometrical shape that has no human features (Jacob, 2008: 206-208). It is clear that infants show a consistent preference for “helping” as opposed to “hindering” shapes, yet ascertaining the intent of a triangle or square is clearly not due to use of motor simulation of non-biological movement (Kuhlmeier et al., 2003). In the Fogassi et al. (2005) and Iacoboni et al. (2005) experiments discussed above it was

12

shown that correlations exist between MN activity and the contextual cues of the scene. As noted, when the monkey sees an experimenter grasp a piece of food in order to eat it, neurons fire, yet MNs would not activate without any object involved when the experimenter mimicks the grasp to eat. Since goal hypotheses are only generated for meaningful actions, this indicates that the goal of the action is understood, without simulation. So, if only actions with clear outcomes are simulated by MNs, then simulation is not involved in the judgment of determining whether an action has a goal, which contradicts the evidence as interpreted in support of simulation theories. If MN simulation is involved in MN function, the simulation observed in the monkey is because of his having understood a goal, and not because he comprehends the goal due to his having simulated the action.

Conclusions First, I introduced the topic under conjecture from neuroscience about the role of motor action in mindreading, involving MNs. Next, I distilled these to a diagram showing two theoretic views, describing their discrete characteristics and how they claim to represent the function of MNs. In order to answer the question of whether motor action has a role in mindreading, from an evidence-based perspective I’ve explored the question of how MNs function, which is the critical underlying component for simulation-based theories. In doing so I’ve outlined the main components behind Gallese and Goldman’s (1998) influential idea that MNs simulate actions between individuals in order to yield intentions and goals. I’ve also represented the alternative view that arises in the most basic sense from the fact that MNs were first found in in macaque monkeys, yet we are short on proof of mindreading in monkeys. Until about 2005 most studies in the wake of the excitement following the discovery of MNs were designed with the original premise in mind, seeking to further the hypothesis of the

13

original authors of View 1 above. But studies designed in this way have done little to enhance actual knowledge. The results of these experiments have led the movement of theorists to endorse a new model, based on chains of logically related MNs as outlined in View 2. In considering both views complementary points were discussed in the research analysis of this paper. Here I summarized the evidence that shows the inconsistency of a strict resonance model of the activity of MNs based on the hierarchy of intentions, and the new findings for logically related MNs, shown to at best predict an agents motor intention, not social intention. The second point highlighted was the tension between computational resources afforded by the theory of internal forward models and a retrodictive process for MNs used to compute an agent’s intention. Instead, it has been proposed that, using an internal inverse model, MNs compute a representation of the agent’s motor command from a prior representation of the agent’s intention. The third point discussed in the research section of the paper is the experimental failure to provide an evidentiary basis for MN activity as a representation of the agent’s prior intention through embodied motor simulation. Rather, the evidence seems at least in part to agree with the alternative evaluations in which MN activity presupposes a representation of the agent’s intention using a perception of contextual cues. The alternative account does affirm the value of MNs in the social cognition of primates, as it gives weight to MN contribution to an onlooker's ability to estimate a conspecific’s next motor act (Jacobs, 2008). From this I conclude that before we can explain how we ascribe intention in a system that relies on MNs, new foundational research is needed to understand their function. It might be most fruitful to attempt to falsify the alternative view—View 2 above—since so much experimental work in support of View 1 has yielded few steps forward.

14

References Blakemore, S. and Decety, J. (2001). From the Perception of Action to the Understanding of Intention. Neuron, Vol. 42:323–334. Accessed online Feb 23, 2012. Gallese, Vittorio.Eagle, Morris N. Migone, Paolo. (2007). Intentional Attunement: Mirror Neurons and the Neural Underpinnings of Interpersonal Relations. Journal of the American Psychoanalytic Association. Sage Publications. Accessed online Feb. 26 2012.

15

Gallese, Vittorio. (2001). The ‘Shared Manifold’ Hypothesis From Mirror Neurons To Empathy. Journal of Consciousness Studies, Vol. 8, No. 5–7:33–50. Accessed online. < http://philpapers.org/rec/GALTSM > Iacoboni, Marco. Molnar-Szakacs, Istvan. Gallese, Vittorio. Buccino, Giovanni. Mazziotta, John C. Rizzolatti, Giacomo. (2005). Grasping the Intentions of Others with One’s Own Mirror Neuron System. PLoS Biology March 2005 Volume 3, Issue 3: 05290535.Accessed online Feb 23, 2012. Jacob, Pierre. (2007). Replies to our Critics. Psyche: Volume 13, Issue 2: 1-11. Accessed online Feb. 24, 2012. < http://psyche.cs.monash.edu.au/> Jacob, Pierre. (2008). What Do Mirror Neurons Contribute to Human Social Cognition? Volume 23, Issue 2:145–255. Mind and Language. Accessed online Feb 20, 2012. Kuhlmeier, V. Wynn, Karen. Bloom, Paul. (2003. Attribution of Dispositional States by 12Month-Olds. Psychological Science, Vol. 14, No. 5: 402-408. Accessed online Mar. 15, 2012. < http://www.jstor.org/stable/40064159> Pierno. Ansuini. Castiello. (2007). Motor Intentions versus Social Intentions: One System or Multiple Systems? Psyche: Volume 13, Issue 2: 1-10. Accessed online Feb. 24, 2012. Sartori, Luisa. Becchio, Cristina. Castiello, Umberto. (2011). Cues to intention: The role

16

of movement information. Cognition: Volume 119, Issue 2, May 2011: 242–252. Elsevier Ltd. Accessed online Feb. 23, 2012. Rizzolati, G. Arbib, M. (1998). Language Within Our Grasp. Trends in Neurosciences, Volume 21:188–194. Accessed online March 10, 2012. < http://www.sciencedirect.com.ezpprod1.hul.harvard.edu/science/article/pii/S0166223698012600> Rizzolatti, Giacomo. Fogassi, Leonardo. Gallese,Vittorio. (2001). Neurophysiological Mechanisms Underlying the Understanding and Imitation of Action. Nature Reviews. Neuroscience. Volume 2(9): 661-70. Accessed online Feb 22, 2012. < http://ukpmc.ac.uk/abstract/MED/11533734> Rizzolatti, Giacomo. Fogassi, Leonardo. Gallese,Vittorio. (2006). Mirrors In the Mind. Scientific American 295: 54-61. Accessed online Feb 24, 2012. Wolpert, D.M. Kawato, M. (1998). Multiple paired Forward and Inverse Models for Motor Control. Neural Networks Volume 11: 1317–1329 Accessed online Feb 24, 2012 < http://www.sciencedirect.com.ezpprod1.hul.harvard.edu/science/article/pii/S0893608098000665>

17