R2.1. Associative learning is sufficient for the development of MNs

Taken at face value, this claim is so obviously false that no one could imagine it was the thrust of our article. Even when one takes “the development of MNs” to be shorthand for “the development of the matching properties, or mirrorness, of MNs” – as we did throughout O2F – it is clear that the developmental system draws on a multitude of resources in addition to associative learning, starting with basic nutrients and the genetically inherited potential to develop neurons at all. The useful version of the “associative learning is sufficient” claim is laid out by Bonaiuto, who suggests that the associative account assumes “every neuron is either directly or indirectly connected with every other neuron in the brain.” This is useful because it gives us a new and clearer way of saying: “The associative hypothesis assumes that gene-based natural selection has played a significant background role with respect to the development of MNs; for example, in shaping the anatomy of visual and motor cortex for visual guidance of action” (O2F, sect. 3.2, para. 1). With Bonaiuto's help, we can now rephrase this: The associative hypothesis does not assume complete or random connectivity between visual and motor cortex. Rather, it assumes that gene-based natural selection has produced partial connectivity as an adaptation for the development of visual control of actions, not for mirroring. For example, visual areas detecting highly action-relevant properties of objects may have the potential to be more richly or readily connected with premotor cortex than visual areas processing less action-relevant properties of objects. It is also possible that genetic evolution has produced some preferential connectivity between visual areas processing observed action and motor areas. However, the associative account suggests that, if this is the case, these predispositions do not favor mirror over non-mirror or counter-mirror connections. Therefore, the issue that separates the associative account and Bonaiuto's view from the genetic account (e.g., Fogassi) is whether the pattern of partial connectivity, or “primary repertoire” (Kilner & Friston), was favored by natural selection because it tends to produce sensorimotor neurons that are good for guiding instrumental action or because it produces MNs that are good for action understanding or another social cognitive function.

R2.2. The genetic account assumes that learning does not contribute to the development of MNs

According to Behme (see also Martin & Santos), our discussion of MN responses to unnatural stimuli – such as plastic crumpling and paper tearing – made clear that, although we said that “the genetic hypothesis does not necessarily assume that experience plays a minimal role” (O2F, sect. 3.1, para. 4), the genetic account is really a straw man who insists that learning does not contribute to the development of MNs. By reaffirming its basic tenets, Fogassi's commentary confirms that the genetic account is not a straw man. More specifically, we agree with Behme that the mere fact that some MNs respond to unnatural stimuli is problematic only for the starkest version of the genetic account. However, as we emphasized (O2F, sect. 4.1), many MNs do not merely respond, they respond maximally, to unnatural stimuli. This is hard to reconcile even with versions of the genetic hypothesis that assign important roles to sensory and/or motor learning, because these assume that links among sensory neurons, or among motor neurons, account for responses to unnatural stimuli. In contrast, maximal responding to unnatural stimuli implies that the learning that produces tool-use and audiovisual MNs is sensorimotor; it connects sensory neurons coding unnatural stimuli directly with motor neurons coding action (Cook Reference Cook2012).

R2.3. Genetically inherited traits are insensitive to experience

We are grateful to Kilner & Friston and Lingnau & Caramazza for mentioning some of the classic studies showing unequivocally that genetically inherited traits are sensitive to experience.

R2.4. Genetically inherited traits are insensitive to associative learning

The fourth claim deserves closer scrutiny. Several commentators assumed not only that we are making this claim, but also that, without it, the evidence that sensorimotor learning can radically change MNs does not support the associative over the genetic account (Bertenthal; Bonaiuto; Gervais; Kilner & Friston; Newman). This misconception is largely our fault. The argument underlying our tests of the associative account has been explained previously (Heyes Reference Heyes2010), but was not made explicit in the target article. To rectify this, here is a short, focused version of the argument: Our training experiments (reviewed in O2F, sect. 7) indicate that (A) counter-mirror sensorimotor experience (e.g., moving an index finger whenever you see little finger movement) reverses the matching properties of MNs, and (B) sensorimotor experience with inanimate objects (e.g., making a fist whenever you see a certain shape) induces MNs to respond to inanimate objects.Footnote ¹ As predicted by the associative account, effects A and B show that sensorimotor experience perturbs but does not damage MNs: it prevents MNs from selectively encoding similar observed and executed actions, but does not stop them from encoding systematic relations among stimuli and responses. This is hard for the genetic account to accommodate because (1) if a trait is a gene-based adaptation, its development tends not to be perturbed by environmental variations that were present when the trait evolved, and (2) sensorimotor experience of the kind that reverses and induces MN activity is likely to have been present in the period when, according to the genetic account, MNs evolved. Specifically, it is likely that the common ancestors of extant monkeys and humans experienced contingencies between observation and execution of nonmatching actions (e.g., when one individual countered a blow from another) and between objects and actions (e.g., when distinctive actions were made on distinctive objects).Footnote ² Thus, our key assumption is not that “genetically inherited traits are insensitive to associative learning,” but that “if a trait is a gene-based adaptation, its development tends not to be perturbed by environmental variations that were present when the trait evolved.”

R2.5. Associative learning cannot be adaptively specialized

Several commentators have suggested that we were wrong to contrast the genetic account with the associative account because associative learning can be adaptively specialized (aka “canalized” and “exapted”) (Behme; Bertenthal; Gervais; Giese; Keysers, Perrett, & Gazzola [Keysers et al.]; Lotem & Kolodny; Martin & Santos; Simpson et al.). For example, as indicated by Garcia's classic experiments on food aversion learning (e.g., Garcia et al. Reference Garcia, Kimeldorf and Koelling1955), genetic evolution can alter input mechanisms or learning rate parameters so that contingencies between some pairs of events (e.g., flavor and illness) are learned more readily than others (e.g., flavor and shock), a phenomenon known as “selective association” (Gemberling & Domjan Reference Gemberling and Domjan1982; Heyes Reference Heyes2003; Pearce Reference Pearce2008). To be clear: We have no doubt that associative learning can be adaptively specialized. Indeed, if we had any doubts, we would not have put so much effort into experiments that were explicitly designed to test for signs of adaptive specialization in the development of MNs. We have tested for adaptive specialization – for example using counter-mirror and induction training (O2F, sect. 7) – and failed to find it. That is why the associative account is not a hybrid model; why it is “extreme” (Lotem & Kolodny) in suggesting that MNs are forged by domain-general processes of learning that have not been specialized by genetic evolution for the development of MNs.

The associative hypothesis would be less plausible if it had been shown that adaptive specialization of associative learning is not merely possible, but normal or typical (Lotem & Kolodny). However, the evidence does not support typicality. There are quite a few examples of selective association, but selective association can be due to prior learning (Mackintosh Reference Mackintosh, Hinde and Stevenson-Hinde1973) rather than genetic processes – “ontogenetic inflection” rather than “phylogenetic inflection” (Heyes Reference Heyes2003) – and in many cases it has merely been assumed that the source of selective association is phylogenetic; that it constitutes adaptive specialization.

The commentaries on our article offer a rich array of hybrid genetic-associative possibilities. For example, commentators point out that MNs could be built by a form of associative learning that is adaptively specialized for empathy (Gervais) or self-monitoring (Bonaiuto) rather than action understanding; via epigenetic as well as genetic processes (Simpson et al.); by the modification of memory parameters as well as input mechanisms (Bonaiuto; Keysers et al.; Lotem & Kolodny); and where the modified input mechanisms involve mother-infant face-to-face interaction, infant-directed speech, imitation by adults, and the face preference (Gervais; Keysers et al.). We agree that these are all logical possibilities and that there is “no a priori reason” (Behme), such as lack of parsimony (Lotem & Kolodny), to reject hybrid models. Indeed, as we said in the target article, we see hybrid modeling as “a promising direction for future research” (O2F, sect. 8.2). We just want to see some evidence. It is not enough to generate hypotheses; they also need to be tested.

Reader's commentary confirms that it is possible to test adaptive specialization hypotheses by discussing research on social learning, rather than MNs, in which this has been done successfully. In contrast, very few commentators have identified empirical research on MNs that, in their view, supports a hybrid, adaptive specialization model over the associative and genetic accounts. Oberman et al. cite a study by Howlin and Moss (Reference Howlin and Moss2012) which reported that people with autism spectrum conditions (ASC) are at a disadvantage, compared with neurotypical controls, in relation to employment, social relationships, health, and quality of life. To count this as evidence for a hybrid model, it would be necessary to make three unjustified assumptions: that people with ASC have abnormal MNs; that their disadvantages are due to this abnormality; and that the disadvantages would not occur if MNs developed in the manner described by either the genetic or the associative account.

Similarly, citing Meltzoff (Reference Meltzoff1995), Bertenthal suggests that infants more readily imitate human than mechanical movements and that this animacy bias supports a hybrid model over the associative account. However, a replication of Meltzoff's (Reference Meltzoff1995) study found evidence that his results were due to stimulus enhancement – a phenomenon that does not implicate MNs – rather than imitation (Huang et al. Reference Huang, Heyes and Charman2002). More broadly, animacy bias in imitation – which has been reported in other studies – is fully compatible with the associative account. Human developmental environments typically offer many more opportunities to associate the observation of human action, than the observation of mechanical action, with action execution (Press Reference Press2011). Therefore, one does not need to invoke adaptive specialization to explain animacy bias. Furthermore, as predicted by the unadorned associative account, animacy bias is abolished when action execution is correlated with observation of mechanical movements (Press et al. Reference Press, Gillmeister and Heyes2007).

Taking a broader approach, Gervais (see also Martin & Santos and Oberman et al.) suggests that hybrid models are better able to explain why monkeys and humans, but not rats, have MNs. The first problem with this argument is that we do not know whether rats or free-living monkeys have MNs; they have not been tested. The second problem is that the associative account predicts species differences based on the amount and kind of sensorimotor experience received in the course of typical development, and – borrowing Bonaiuto's terminology again – the pattern of partial connectivity between sensory and motor areas that has been favored by genetic evolution for visual guidance of action.

We hope our article will encourage those who back hybrid models to test them against the genetic and associative accounts in carefully designed experiments. This was, to a large degree, our purpose in writing O2F. The resulting experiments may well show that the associative account is wrong or needs to be augmented (Heyes, in press). For example, in common with Gervais, we have suggested elsewhere that cultural evolution, rather than genetic evolution, may have favored social practices that facilitate the development of imitation and MNs through associative learning (Heyes Reference Heyes2012a; Reference Heyes, Sterelny, Joyce, Calcott and Fraser2013). But even if future experimental research shows that the associative account is plain wrong, in our view it would be a good result. It would mean that research on MNs has moved from exciting speculation to exciting experimentation: from ideas to tests and testability.

R4.1. Hebbian and/or associative?

In O2F we offered two interpretations of what Keysers et al. describe as the Hebbian hypothesis: On one reading, it is “identical in substance to the associative hypothesis [and] helpful in providing a more detailed neuronal model of how sensorimotor experience makes MNs out of motor neurons” (sect. 8.1, para. 1), and on the other reading – which we thought “less likely to represent the authors' intentions” – “‘Hebbian learning’ differs from ‘associative learning’ in depending on contiguity alone, rather than both contiguity and contingency” (sect. 8.1, para. 2). In their commentary, Keysers et al. seem to confirm that our first interpretation was correct. They certainly emphasize the neuronal specificity of their model, and stress that, however Hebbian learning may have been understood in the past, Keysers et al. understand it to be a process that depends on contingency. However, if, as they suggest, the associative and Hebbian accounts are merely “descriptions at different levels,” and it is “idle” to argue that one is better than the other, it is not clear why Keysers et al. also suggest that the two accounts make different predictions; that the Hebbian account would, and the associative account would not, predict the formation of associations between as well as within action phases – for example, between “reach-STS” (observation of reaching) and “grasp-PM” (execution of grasping), as well as between “grasp-STS” (observation of grasping) and “grasp-PM” (execution of grasping). Two theories are either equivalent, and it is idle to ask which is correct, or they make different empirical predictions, and it is the basic stuff of science to test them against one another.

In this case, we think that Keysers et al. are right about the equivalence rather than the differential predictions. Tracking of predictive relationships between events (rather than event co-occurrence), via prediction error, is at the heart of the formal models of learning that underwrite the associative account (e.g., Rescorla & Wagner Reference Rescorla, Wagner, Black and Prokasy1972; Schultz & Dickinson Reference Schultz and Dickinson2000). Therefore, the associative account is wholly compatible with “sensory prediction” (Keysers et al.), “predictive coding” (Brass & Muhle-Karbe; Kilner & Friston; Press et al. Reference Press, Heyes and Kilner2011), and, most specifically, with the formation of associations between, as well as within, action phases. The challenge for both the associative and the Hebbian views – assuming they are equivalent in substance – is to define the conditions in which learning produces “classical” MNs, which fire during the observation and execution of the same action (e.g., grasping), and those in which it yields nonmatching MNs such as logically related MNs, that may encode relations between successive components of an action sequence. The advantage of the associative/Hebbian account over the genetic account is that it can readily accommodate both kinds of MNs because associative learning is all about prediction rather than matching. However, because associative learning does prediction on a variety of timescales, those who support an associative account need to work on predicting the timescale on which it will operate for different types of action and of sensorimotor experience.

R4.2. Ideomotor and/or associative

In their thoughtful commentary, Brass & Muhle-Karbe suggest that MNs are a neural substrate of ideomotor representations, established through the formation of response-effect associations, and contrast this with our associative account, which identifies MNs as “sensorimotor” representations and refers to “sensorimotor” learning (see also Ho, MacDonald, & Swain [Ho et al.]). They may well be right. There may be a real, empirically testable difference between their ideomotor account and our associative model. But, as with Keysers et al., where Brass & Muhle see conflict, we see harmony. We suspect that the two views seem to be different because what we call “stimuli” they call “response effects,” but the two accounts are in fact united in suggesting that the critical associations – the associations that give MNs their cardinal properties – link neuronal representations of “R” and “E,” not “S” and “R,” in their diagram (see Fig. 1 in Brass & Muhle-Karbe's commentary). If this is correct, our training experiments (O2F, sect. 7.1) contain the right ingredients for the ideomotor account – they expose participants to contingent relationships between R and E, execution and observation of action – but not in the order in which these events are typically experienced in everyday life. In each trial of our training regimes, participants observe an action and then execute an action. We have used this kind of regime because it is easier to control the contingencies experienced by participants when the observed action comes first, and because there is compelling evidence that associative learning is bidirectional (Ayres et al. Reference Ayres, Haddad and Albert1987; Elsner & Hommel Reference Elsner and Hommel2001; Reference Elsner and Hommel2004; Mahoney & Ayres Reference Mahoney and Ayres1976). It is important to note that, confirming the bidirectionality assumption, other groups have replicated and extended our findings using training regimes in which response execution precedes action observation (e.g., Wiggett et al. Reference Wiggett, Hudson, Tipper and Downing2011).

Of course, the proof of the ideomotor/associative pudding will be in the testing, but the examples mentioned by Brass & Muhle-Karbe do not yet convince us that it will be possible to generate differential predictions. They suggest that two empirical findings are inconsistent with the associative account. In the first, Kunde (Reference Kunde2001) found that responses were executed faster when followed by compatible, rather than incompatible, sensory effects – evidence for anticipatory excitation of the sensory consequences of action. However, as discussed in the previous paragraph, the associative account is compatible with predictive motor-to-sensory propagation, following exposure to either S-R or R-E contingencies. In the second study, Liepelt et al. (Reference Liepelt, von Cramon and Brass2008b) found a larger index/middle automatic imitation effect when fractional finger movements were presented inside, rather than alongside, metal clamps. Applying the associative account, Heyes (Reference Heyes2011) agreed with the view (proposed by Brass previously and given by Brass & Muhle-Karbe in their present commentary) that this could be described as an effect on mirroring of “high-level beliefs about the intentionality of the observed action,” but suggested that the effect was mediated by enhanced attention to the clamped movements (see also Heil et al. and Brown & Brüne).

R4.3. Reasoning and/or association

The associative account identifies the learning processes involved in the development of MNs with those that produce conditioning in a variety of animal species. However, Heil et al. argue that this type of “low-level” associative process is ill-suited to producing MNs. Appealing to evidence that beliefs about causal structure can modulate human associative learning, Heil et al. raise the possibility that another type of associative process is responsible for the development of MNs, one that is shaped by and integrates higher-level cognitive processes. This is an interesting but, in its current form, largely speculative hypothesis. To make it testable, Heil et al. would need to say more about the kind of high-level processes they take to be involved in the development and online control of MNs (e.g., the processes that yield contingency awareness may or may not be reasoning processes) and about the way in which the high- and low-level processes are supposed to interact. In humans, language-mediated high-level processes can certainly influence associative learning by modulating input mechanisms – for example, task instructions can direct participants' attention towards or away from the computer screen on which stimuli are presented – but this kind of interaction is entirely compatible with our associative hypothesis (e.g., Lotem & Kolodny). To motivate their hypothesis, Heil et al. would also need to cite findings from the MN literature that they regard as consistent with a high-level associative process, but inconsistent with a low-level mechanism. At present they suggest only that a high-level process would be needed if MNs are to mediate action understanding. This may be so, but it is not a compelling argument in favor of a high-level process given the ambiguity of the term “action understanding” (O2F, sect. 3.1), and the corresponding lack of clear evidence that MNs mediate action understanding (O2F, sect. 9.2).

In section R4 we have argued that both the associative and Hebbian accounts are compatible with predictive coding; that the associative account is complementary to, not in conflict with, ideomotor theory; and that there are currently no data from MN research suggesting that their development or operation is guided by a “higher-level” associative process. We therefore maintain that phylogenetically ancient associative processes are “the right kind of learning” to yield the field properties of MNs.

R7.1. Understanding “action understanding”

In O2F we pointed out that there is “no consensus about exactly what is meant by ‘action understanding,’ or how it differs from cognate functions such as ‘action perception,’ ‘action recognition,’ and ‘action selection’ ” (sect. 3.1, para. 2). This is a conceptual problem that has generated a major empirical problem. When it is not clear what a mechanism is supposed to be doing, it is all but impossible to design experiments finding out whether the supposition is correct. We also suggested that the best way to solve this problem would be to focus on the development of system-level theories (sect. 9.1.2), in which MNs feature as one component of a system defined by its outcome or typical effect and in which the role of MNs is clearly distinguished from, and related to, the roles of other components (Heyes, in press).

In the commentaries, no one denies that understanding “action understanding” is a problem, and many endorse the need for system-level theory, explicitly or by discussing particular system-level theories with enthusiasm (Behme; Heil et al.; Holt & Lotto; Kilner & Friston; Lingnau & Caramazza; Saygin & Dick). (We particularly like Kilner & Friston's slogan, “no neuron is an island.”) However, confirming that we have not constructed a straw man (cf. the claim by Oberman et al.), members of the Parma group eschew the system-level approach. Instead, these commentators seek to elucidate the “action understanding” function of MNs by (1) repeating the claim that they encode or interpret action “goals” (Fogassi; Gallese & Sinigaglia), and are therefore “bearers of intentionality” (Gallese & Sinigaglia); (2) noting that MNs transform “sensory information into a motor format” (Rizzolatti); and (3) describing MN-mediated action understanding as “action understanding ‘from inside’ (Rizzolatti & Sinigaglia Reference Rizzolatti and Sinigaglia2010): ‘a first person process, where the self feels like an actor, rather than a spectator’ ” (Rizzolatti).

The first of these strategies is not promising because, as we argued in O2F (sect. 4), the term “goal” is at least as ambiguous as “action understanding,” and on at least two natural interpretations of “goal” – as an object of action and as a high-level intention – the evidence suggests that goals are not selectively encoded by MNs. The second strategy is more straightforward, but it is clearly insufficient because it does not refer to the cardinal, matching properties of MNs. They can certainly be said to transform sensory information into a motor format, but that is also true of, for example, canonical neurons. The third strategy is phenomenological; it distinguishes MN-mediated action understanding from other sorts of action understanding with reference to the actor's conscious experience – whether he or she “feels” like an actor or a spectator. This suggestion is coherent and intuitively appealing, but it is unlikely to be helpful as a guide for experimental research. Without a return to the methods of introspectionism, researchers in psychology and neuroscience are “spectators”; we are “on the outside,” trying to work out the function of MNs, not by examination of our own conscious experience, but by studying the brains and behavior of other people. Therefore, to use the phenomenological strategy we would need to be told how, from the outside, to distinguish “inside” and “outside” action understanding; how these types of action understanding differ in terms of the observable behavior they produce.

Rizzolatti also explains with helpful clarity that he now believes that mirror neurons contribute to social cognitive functions such as imitation, empathy, and song recognition, as well as action understanding, and regards “the problem of how MNs originate [as] utterly irrelevant as far as their function is concerned.” We have no quarrel whatever with these views; they are wholly in accord with the “functionally permissive” character of the associative hypothesis (O2F, sect. 9) and the way in which it splits questions about the origin of MNs from questions about their function. However, it is important to note that these views do not make the genetic hypothesis a straw man – Fogassi's commentary and the quotations in O2F (sect. 3.1), attest to its reality – or obviate the need for system-level theory. Indeed, they increase that need. There are just two widely recognized methods of defining the function of a trait in a biological system: one does it with reference to the trait's history, typically its evolutionary history, and the other, “functional analysis,” does it in the context of what we have described as system-level theory (Cummins Reference Cummins1975; Godfrey-Smith Reference Godfrey-Smith1994). Given that these are the two paths available, rejection of the history route – denial that the function of MNs relates to their origins – makes pursuit of the alternative, system-level theory, absolutely essential; and multiplying the putative functions of MNs means that we need system-level theories defining the role of MNs, not only in action understanding, but also in imitation, empathy, and song recognition.

Gallese & Sinigaglia are absolutely right: the mere fact that “action understanding” is a new term is not a reason to reject it. All we are saying, here and in O2F, is that those who use the term need to tell us what it means, and to do so in the context of system-level theory.

R7.2. Experimental approaches

Few commentators (except perhaps Fogassi) disagree with our proposal that MN research should report participants' developmental history: that is, their sensory, motor, and sensorimotor experience with the actions for which MNs are being tested. Indeed, Gervais comments on the “value of the enterprise” of documenting the effects that variation in developmental environment may have on the ontogeny and function of MNs. In this vein, the developmental data described by Krogh-Jespersen, Filippi, & Woodward (Krogh-Jespersen et al.) are a promising start. These data demonstrate how infants' motor experience (specifically, in this case, their tendency to produce goal-directed actions) relates to their neural responses to observation of others' actions; to their ability to understand others' actions; and to their ability to imitate. These findings are consistent with the predictions of the associative account; however, they serve to highlight the importance of assessing participants' previous sensorimotor (not just motor; cf. Gallese & Sinigaglia; Rizzolatti) experience with specific actions when investigating neural responses to, understanding of, or imitation of those actions (see earlier response to Bertenthal for further explication of this point).

Several other commentators appear to embrace our suggestion that MN research pay closer attention to developmental history: Simpson et al. propose examining the effects of early sensorimotor experience on imitation, while Ho et al. note that it would be informative to consider the development of MNs in infants who have atypical experience, for example, those who cannot walk themselves or whose parents suffer from depression. We wholeheartedly concur that research with these populations could prove highly informative in dissociating the relative roles played by sensory, motor, and correlated sensorimotor experience in the development of MNs, and perhaps also shed some light on their function.

We are pleased that most commentators have endorsed our call for a more rigorous, experimental approach to the investigation of MN function (Behme; Gervais; Holt & Lotto; Krogh-Jespersen et al.; Lingnau & Caramazza; Newman; Oosterhof et al.; Reader; de la Rosa & Bülthoff). Here we address some criticisms of our proposals and discuss the commentators' additional suggestions for future research.

Orban's commentary illustrates the constraints that have been placed on MN research to date by the genetic account. It suggests that MNs have only, and will only, be found for grasping actions; that MNs can only be considered such when found in classical areas; and that only a primate model holds any value for understanding the human brain. Such suggestions are in accord with an account that considers MNs to be a genetic adaptation for representing others' grasping actions in one's own motor system (see also Fogassi). They illustrate the risk that the genetic account will confine future research to certain actions, brain areas, and species. However, as we indicated in the target article, the associative account predicts (and data support these predictions) that mirror neurons will be found for a range of actions (e.g., Ferrari et al. Reference Ferrari, Gallese, Rizzolatti and Fogassi2003), in a range of brain areas (e.g., Dushanova & Donoghue Reference Dushanova and Donoghue2010; Mukamel et al. Reference Mukamel, Ekstrom, Kaplan, Iacoboni and Fried2010; Tkach et al. Reference Tkach, Reimer and Hatsopoulos2007), and across a range of species (e.g., Prather et al. Reference Prather, Peters, Nowicki and Mooney2008). We therefore see no reason to reject the possibility that MNs could be induced in other species through sensorimotor training of species-appropriate actions.

We are not aware of any research that has investigated whether MNs are present in rodents. However, those commentators who deny that MNs are (Gervais), or could be (Orban), present in rodents (or in other species; Oberman et al.) seem to have misunderstood our point. The associative account suggests that whether MNs are likely to exist in laboratory (or indeed free-living) rats will depend on their prior learning history; but it makes the strong prediction that regardless of whether they are present or not, appropriate experience in which the sight or sound of an action is paired contingently with execution of that action should produce visuo- or auditory-motor MNs for that action, in rodents and indeed in any other species. This approach is also endorsed by Reader's commentary, in which he proposes the use of insect or fish models to investigate MNs. Building on the previous point, Newman's commentary provides an interesting description of behaviors consistent with the existence of audio-vocal mirror-like circuits in a variety of species. We agree with Newman that it would be interesting to test for MNs in these species – indeed the presence of audio-vocal MNs in swamp sparrows has already been demonstrated (Prather et al. Reference Prather, Peters, Nowicki and Mooney2008).

The preceding commentators have focused on experimentation into MN properties. Other commentators propose investigation of the possible role of MNs in contagious behavior (Provine) and in higher-level social cognition tasks: Martin & Santos advocate carrying out single-unit recording while macaques undergo tests assessing their awareness of others' knowledge. We believe that such an endeavor would benefit from the kind of system-level analysis we have outlined in the target article (see also Holt & Lotto). Such an analysis would specify, for example, the role that MNs are predicted to play in these tests and whether or not this role must be carried out by MNs.

Several commentaries touch on the methodological challenges facing future MN research. Lingnau & Caramazza understood us to be suggesting that research into the modulation of MN responses by associative learning is sufficient to provide evidence of the role of MNs in social cognition. This is certainly not the case. Instead, the associative account implies that the function of MNs cannot be discovered solely by research on their field properties; tests of the kind described in O2F (sect. 9.1.3), and promoted by Lingnau & Caramazza, are also needed. We agree, therefore, that temporary and permanent lesion studies are an important component of research on MN function; indeed, we presented some evidence from such studies in the target article. However, such studies are vulnerable to what we termed the “localization problem”: not all neurons in “classical” MN areas are mirror neurons and thus lesion studies may overestimate the involvement of MNs in the cognitive function of interest.

Oosterhof et al. point out that MVPA, although not a causal technique, can help to overcome the localization problem. This method, unlike standard fMRI approaches, allows measurement of spatially distributed responses to specific observed and executed actions. Oosterhof et al. provide an interesting description of how this technique could be used to test further the predictions of the associative account, and we agree that MVPA may also prove a promising additional tool to investigate the functional properties of MNs.

Bonaiuto argues that computational models can provide a convincing demonstration of the concepts underlying the associative account. We agree and have therefore previously presented such a model (Cooper et al. Reference Cooper, Catmur and Heyes2013a; Reference Cooper, Cook, Dickinson and Heyes2013b). As Bonaiuto suggests, this model requires appropriate input representations of actions. It demonstrates how sensorimotor associative learning can modulate the development of mirror properties (Cooper et al. Reference Cooper, Cook, Dickinson and Heyes2013b) and how mirror networks operate, once developed, to produce imitative behavior (Cooper et al. Reference Cooper, Catmur and Heyes2013a). Other computational models which use domain-general learning processes and demonstrate the development of MN-like properties in the domain of speech processing are described by Holt & Lotto and Saygin & Dick. Giese also summarizes how insights from visual pattern recognition may be applied to computational modeling of MNs. Thus, we consider computational modeling to be important not only for testing accounts of how MNs develop, but also for providing novel predictions regarding the functional properties of mature mirror systems.