The central hypothesis put forward by Cook et al. is that mirror neurons (MNs) are not the direct result of a genetic adaptation per se, but rather, they are a result of associative learning between visual and motor modalities. Within this framework, it is the mechanism underlying associative learning that has been selected for – not MNs. One of the crucial differences between the genetic adaptation account and the associative account is that the adaptation account assumes that the origin and function of MNs are co-dependent, whereas the associative account dissociates the questions about origin and function, such that they can be treated as independent processes. Any discussion about the ontogeny of MNs, therefore, depends inextricably upon their proposed functional role. Cook et al. adopt the view that the function of MNs is to encode the goal of an observed action. They then go on to argue that the neurophysiological data do not clearly support this functional role and should be considered as evidence against the genetic adaptation account.
Another possibility though, is simply that the mirror neurons do not encode the “goal” of the observed action – rather than encoding the goal of an action, MNs may be part of a distributed network that predicts the sensory (exteroceptive and proprioceptive) consequences of an action (Clark Reference Clark2013; Kilner et al. Reference Kilner, Friston and Frith2007b). Within this framework, MNs discharge during action observation not because they are driven by the visual input, but because they are part of a generative model that is predicting that input. In this predictive coding account, the motor system is active when observing an action because it is the best model of the observed action. The generative model entails a representation of – or probabilistic belief about – the goal or intention of an observed action. Given this belief, MNs, in concert with other brain areas, generate a prediction of the sensory consequences of the most likely action that would achieve the goal or intention: for example, the kinematics of the action. By comparing the predicted sensory information with the actual sensory information, the system can infer the goal of the observed action by minimising prediction error. Crucially, in this framework MNs play the same role during both action execution (active inference) and action observation (predictive coding) (Friston et al. Reference Friston, Mattout and Kilner2011). In predictive coding, neuronal representations are used to make predictions, which are optimised during perception by minimizing prediction error. In active inference, action tries to fulfill these predictions by minimizing sensory (e.g., proprioceptive) prediction error. This enables intended movements (goal-directed acts) to be prescribed by predictions, which action is enslaved to fulfill. In this framework, the “mirrorness” of MNs simply reflects the fact that the content of the representations, the action, remains the same in action execution and observation. What changes is the context, or agency – whether the action was produced by the self or another. Therefore, whatever account, genetic or associative, best explains the ontogeny of mirror neurons, it must hold for both action observation and action execution. Within the active inference framework, any selective pressure must operate at the level of agency (self or other) and not at the level of the mirror neurons.
Cook et al.'s article highlights the important point that it is incredibly hard to disambiguate the genetic and associative contributions to the ontogeny of a specific neuronal population. This is because all neurons show associative plasticity, and their response profiles can be modified through interactions with the environment – where these modifications depend upon heritable (genetic) synaptic (associative) plasticity. For example, orientation-tuned responses in neurons in primary visual cortex can be elicited in kittens as soon as they open their eyes – suggesting that the orientation maps are innate. However, depending on the environment, the orientation-tuning can be optimised during development to reflect the observed world (Blakemore & Mitchell Reference Blakemore and Mitchell1973). If a kitten is raised in an environment with only vertical stripes, the response properties of the kitten's neurons in the primary visual cortex will reflect this and responses to horizontal stimuli will be lost. In addition to this, many responses of neuronal populations that we think of as being a result of evolutionary adaptations – for example, binocular disparity responses and direct cortico-motoneuronal cells – are not present at birth but develop postnatally. This is in distinction to the formal phenotypes that contextualise the function of these neurons; for example, having two eyes and opposable thumbs. Indeed, for visual responses, the consensus view is that the primary repertoire of connections that underlie vision are present at birth and are fundamentally refined by early postnatal experience. In other words, it is not the neurons that are the genetic adaptation, but rather, how they form connections. In this light, it is tempting to propose the same for the visuomotor responses of MNs. In other words, mirror neurons arise as a result of domain-general mechanisms of associative learning, as proposed by Cook et al., but in the context of cortical connections between visual and motor systems selected by genetic adaptation. From this point of view, with respect to the ontogeny of MNs, perhaps we should consider that no neuron is an island?
The central hypothesis put forward by Cook et al. is that mirror neurons (MNs) are not the direct result of a genetic adaptation per se, but rather, they are a result of associative learning between visual and motor modalities. Within this framework, it is the mechanism underlying associative learning that has been selected for – not MNs. One of the crucial differences between the genetic adaptation account and the associative account is that the adaptation account assumes that the origin and function of MNs are co-dependent, whereas the associative account dissociates the questions about origin and function, such that they can be treated as independent processes. Any discussion about the ontogeny of MNs, therefore, depends inextricably upon their proposed functional role. Cook et al. adopt the view that the function of MNs is to encode the goal of an observed action. They then go on to argue that the neurophysiological data do not clearly support this functional role and should be considered as evidence against the genetic adaptation account.
Another possibility though, is simply that the mirror neurons do not encode the “goal” of the observed action – rather than encoding the goal of an action, MNs may be part of a distributed network that predicts the sensory (exteroceptive and proprioceptive) consequences of an action (Clark Reference Clark2013; Kilner et al. Reference Kilner, Friston and Frith2007b). Within this framework, MNs discharge during action observation not because they are driven by the visual input, but because they are part of a generative model that is predicting that input. In this predictive coding account, the motor system is active when observing an action because it is the best model of the observed action. The generative model entails a representation of – or probabilistic belief about – the goal or intention of an observed action. Given this belief, MNs, in concert with other brain areas, generate a prediction of the sensory consequences of the most likely action that would achieve the goal or intention: for example, the kinematics of the action. By comparing the predicted sensory information with the actual sensory information, the system can infer the goal of the observed action by minimising prediction error. Crucially, in this framework MNs play the same role during both action execution (active inference) and action observation (predictive coding) (Friston et al. Reference Friston, Mattout and Kilner2011). In predictive coding, neuronal representations are used to make predictions, which are optimised during perception by minimizing prediction error. In active inference, action tries to fulfill these predictions by minimizing sensory (e.g., proprioceptive) prediction error. This enables intended movements (goal-directed acts) to be prescribed by predictions, which action is enslaved to fulfill. In this framework, the “mirrorness” of MNs simply reflects the fact that the content of the representations, the action, remains the same in action execution and observation. What changes is the context, or agency – whether the action was produced by the self or another. Therefore, whatever account, genetic or associative, best explains the ontogeny of mirror neurons, it must hold for both action observation and action execution. Within the active inference framework, any selective pressure must operate at the level of agency (self or other) and not at the level of the mirror neurons.
Cook et al.'s article highlights the important point that it is incredibly hard to disambiguate the genetic and associative contributions to the ontogeny of a specific neuronal population. This is because all neurons show associative plasticity, and their response profiles can be modified through interactions with the environment – where these modifications depend upon heritable (genetic) synaptic (associative) plasticity. For example, orientation-tuned responses in neurons in primary visual cortex can be elicited in kittens as soon as they open their eyes – suggesting that the orientation maps are innate. However, depending on the environment, the orientation-tuning can be optimised during development to reflect the observed world (Blakemore & Mitchell Reference Blakemore and Mitchell1973). If a kitten is raised in an environment with only vertical stripes, the response properties of the kitten's neurons in the primary visual cortex will reflect this and responses to horizontal stimuli will be lost. In addition to this, many responses of neuronal populations that we think of as being a result of evolutionary adaptations – for example, binocular disparity responses and direct cortico-motoneuronal cells – are not present at birth but develop postnatally. This is in distinction to the formal phenotypes that contextualise the function of these neurons; for example, having two eyes and opposable thumbs. Indeed, for visual responses, the consensus view is that the primary repertoire of connections that underlie vision are present at birth and are fundamentally refined by early postnatal experience. In other words, it is not the neurons that are the genetic adaptation, but rather, how they form connections. In this light, it is tempting to propose the same for the visuomotor responses of MNs. In other words, mirror neurons arise as a result of domain-general mechanisms of associative learning, as proposed by Cook et al., but in the context of cortical connections between visual and motor systems selected by genetic adaptation. From this point of view, with respect to the ontogeny of MNs, perhaps we should consider that no neuron is an island?
ACKNOWLEDGMENTS
This work was funded by the Wellcome Trust, United Kingdom.