The primary hypothesis set forth by Cook et al. is that mirror neurons (MNs) are the result of generic associative learning processes, rather than the result of evolutionary selection for action understanding. They claim that the standard view of mirror neurons, what they call the “genetic account,” suggests that the predisposition to develop MNs is heritable and was selected for on the basis of their role in action understanding. However, in their characterization of the genetic account, Cook et al. do allow for the role of experience in shaping MNs. Computational models that simulate the development of MNs through experience show that this is possible through associative learning mechanisms, but that the connectivity to form these associations must already be in place and that this connectivity must be somewhat specialized for control of hand actions.
Cook et al. describe the “exaptation hypothesis” as claiming that MNs require a special kind of sensorimotor learning. However, a closer look at several of the computational models developed under this hypothesis, such as the Mirror Neuron System (MNS) model (Oztop & Arbib Reference Oztop and Arbib2002), reveals that they do in fact use standard learning algorithms completely compatible with the associative learning account. What makes these models work is the structure of their input representations and their connectivity. The pure associative learning account seems to assume that every neuron is either directly or indirectly connected with every other neuron in the brain. Such architecture would require significantly more trials of action and observation in order to correctly associate visual stimuli with the relevant motor representations.
The simplest version of the genetic account would predict that MNs would be found in different areas of the brain, depending on the unique history of each individual. This is not the case, at least in monkeys, and this seems to be due to a genetic influence on the patterns of connectivity expressed by each brain region. Indeed, as Cook et al. claim, there is a “wealth of the stimulus” – so much that the space of possible hand–object interaction representations in the visual and motor domains makes the associative learning account computationally intractable. What makes the “exaptation hypothesis” models able to handle such a space is the fact that the inputs are constrained to represent the hand–object relationships appropriate for performing manual actions. This is thought to occur throughout motor development as the infant learns to extract the relevant features from visual stimuli for controlling the hand relative to the object (Oztop et al. Reference Oztop, Bradley and Arbib2004). Once the inputs are restricted to those necessary to control transitive actions, “domain-general learning processes” can proceed to associate the visual representation with the motor program at various levels of abstraction.
Although the learning algorithm in the MNS model was compatible with the associative learning account, the network required extensive pre-processing of its input. Mirror neurons respond to observation of dynamic hand actions and therefore must process trajectories in the space of hand–object relationships. Mirror neurons will often respond to observation of a grasp before the hand contacts the object. In order to predict the outcome of a grasp before its completion, the MNS model transformed a temporal sequence of hand–object relations into a spatial pattern of neural activity for input to the network. A subsequent version of the MNS model, MNS2, discarded this preprocessing step by using a recurrent neural network and a modified learning algorithm to handle raw input sequences (Bonaiuto et al. Reference Bonaiuto, Rosta and Arbib2007). These models show that although MNs may acquire their properties through associative-style learning processes, extra circuitry is required to perform the computations necessary for processing dynamic visual input from objected-directed hand actions.
The MNS2 model additionally proposed that audiovisual MNs develop their auditory properties through simple associative learning. However, in this model, extra mechanisms such as working memory and dynamic remapping were required to handle the case where MNs correctly predict the outcome of a grasp when the final portion was obscured. It is not clear how these functions could be developed through pure associative learning.
Giese and Poggio (Reference Giese and Poggio2003) present a model of visual tuning in the mirror system that is the most compatible with the associative learning account. This model currently does not include a learning mechanism, but it does address the existence of view-dependent and -independent mirror neurons and does not require reconstruction of the arm and hand shape. However, it still requires extensive processing to transform visual input into a reduced space such that it can be associated with motor signals.
The Augmented Competitive Queuing (ACQ) model embeds a network such as those in the MNS and MNS2 models in a larger network that learns self-actions (Bonaiuto & Arbib Reference Bonaiuto and Arbib2010). In this model, MN activity signals recognition of successful completion of one's own actions. Their output is used as an eligibility trace in reinforcement learning algorithms that modify the recognized action's desirability – how likely an action is to lead to a reward; and executability – how likely an action can be successfully performed in the current context regardless of reward. This model shows how mirror systems can have evolved for the purposes of monitoring one's own actions and fit within a reinforcement learning framework for action selection.
A mechanistic model of MNs with random or full connectivity and pure associative learning has never been developed. Current computational models suggest that appropriate coarse-grained connectivity and input representations are required to make the space of possible hand–object relation trajectories tractable. While the associative learning account is compatible with these models at a first approximation, it does not offer any detailed explanations as to how networks of MNs acquire their properties in development and operate in the adult. Conceptual models such as the associative learning theory of mirror neuron origins which do not provide a proof of concept in the form of a computational model, are unconvincing.
The primary hypothesis set forth by Cook et al. is that mirror neurons (MNs) are the result of generic associative learning processes, rather than the result of evolutionary selection for action understanding. They claim that the standard view of mirror neurons, what they call the “genetic account,” suggests that the predisposition to develop MNs is heritable and was selected for on the basis of their role in action understanding. However, in their characterization of the genetic account, Cook et al. do allow for the role of experience in shaping MNs. Computational models that simulate the development of MNs through experience show that this is possible through associative learning mechanisms, but that the connectivity to form these associations must already be in place and that this connectivity must be somewhat specialized for control of hand actions.
Cook et al. describe the “exaptation hypothesis” as claiming that MNs require a special kind of sensorimotor learning. However, a closer look at several of the computational models developed under this hypothesis, such as the Mirror Neuron System (MNS) model (Oztop & Arbib Reference Oztop and Arbib2002), reveals that they do in fact use standard learning algorithms completely compatible with the associative learning account. What makes these models work is the structure of their input representations and their connectivity. The pure associative learning account seems to assume that every neuron is either directly or indirectly connected with every other neuron in the brain. Such architecture would require significantly more trials of action and observation in order to correctly associate visual stimuli with the relevant motor representations.
The simplest version of the genetic account would predict that MNs would be found in different areas of the brain, depending on the unique history of each individual. This is not the case, at least in monkeys, and this seems to be due to a genetic influence on the patterns of connectivity expressed by each brain region. Indeed, as Cook et al. claim, there is a “wealth of the stimulus” – so much that the space of possible hand–object interaction representations in the visual and motor domains makes the associative learning account computationally intractable. What makes the “exaptation hypothesis” models able to handle such a space is the fact that the inputs are constrained to represent the hand–object relationships appropriate for performing manual actions. This is thought to occur throughout motor development as the infant learns to extract the relevant features from visual stimuli for controlling the hand relative to the object (Oztop et al. Reference Oztop, Bradley and Arbib2004). Once the inputs are restricted to those necessary to control transitive actions, “domain-general learning processes” can proceed to associate the visual representation with the motor program at various levels of abstraction.
Although the learning algorithm in the MNS model was compatible with the associative learning account, the network required extensive pre-processing of its input. Mirror neurons respond to observation of dynamic hand actions and therefore must process trajectories in the space of hand–object relationships. Mirror neurons will often respond to observation of a grasp before the hand contacts the object. In order to predict the outcome of a grasp before its completion, the MNS model transformed a temporal sequence of hand–object relations into a spatial pattern of neural activity for input to the network. A subsequent version of the MNS model, MNS2, discarded this preprocessing step by using a recurrent neural network and a modified learning algorithm to handle raw input sequences (Bonaiuto et al. Reference Bonaiuto, Rosta and Arbib2007). These models show that although MNs may acquire their properties through associative-style learning processes, extra circuitry is required to perform the computations necessary for processing dynamic visual input from objected-directed hand actions.
The MNS2 model additionally proposed that audiovisual MNs develop their auditory properties through simple associative learning. However, in this model, extra mechanisms such as working memory and dynamic remapping were required to handle the case where MNs correctly predict the outcome of a grasp when the final portion was obscured. It is not clear how these functions could be developed through pure associative learning.
Giese and Poggio (Reference Giese and Poggio2003) present a model of visual tuning in the mirror system that is the most compatible with the associative learning account. This model currently does not include a learning mechanism, but it does address the existence of view-dependent and -independent mirror neurons and does not require reconstruction of the arm and hand shape. However, it still requires extensive processing to transform visual input into a reduced space such that it can be associated with motor signals.
The Augmented Competitive Queuing (ACQ) model embeds a network such as those in the MNS and MNS2 models in a larger network that learns self-actions (Bonaiuto & Arbib Reference Bonaiuto and Arbib2010). In this model, MN activity signals recognition of successful completion of one's own actions. Their output is used as an eligibility trace in reinforcement learning algorithms that modify the recognized action's desirability – how likely an action is to lead to a reward; and executability – how likely an action can be successfully performed in the current context regardless of reward. This model shows how mirror systems can have evolved for the purposes of monitoring one's own actions and fit within a reinforcement learning framework for action selection.
A mechanistic model of MNs with random or full connectivity and pure associative learning has never been developed. Current computational models suggest that appropriate coarse-grained connectivity and input representations are required to make the space of possible hand–object relation trajectories tractable. While the associative learning account is compatible with these models at a first approximation, it does not offer any detailed explanations as to how networks of MNs acquire their properties in development and operate in the adult. Conceptual models such as the associative learning theory of mirror neuron origins which do not provide a proof of concept in the form of a computational model, are unconvincing.