Evolution has devoted by far and away the largest part of its history to building organisms that can move around in a dynamic environment, sensing their environments in ways conducive to their own survival and reproduction (Brooks Reference Brooks1991). The challenge to cognitive scientists is to explain how the strategies organisms use to solve these basic problems of perception and action scale up to the strategies humans use in solving abstract higher-level problems. I call this the “bootstrapping challenge.” Embodied cognitive science offers a programmatic response to the bootstrapping challenge that attempts to show how high-level problem solving might have been built upon the foundation of a substrate of perception and sensorimotor control. The ascent from sensing and moving to thinking, planning, and language understanding is an incremental and gradual one, and a key strategy may have been the redeployment of sensorimotor capacities to perform high-level cognitive tasks.
Anderson has done the embodied cognition community the enormous service of framing a global hypothesis about how these incremental changes might have taken place in our brains over the course of evolution. The central claim of his massive redeployment hypothesis (MRH) is that more recent cognitive functions such as those involved in abstract problem solving might have their origin in the reuse of evolutionarily older neural circuits that served biologically basic functions. In this commentary, I want to take up Anderson's claim that the principle guiding reuse is “functional inheritance” and not “semantic inheritance.” By “semantic inheritance,” I mean the kind of relation that concept empiricists and conceptual metaphor theories take to hold between concepts and sensorimotor representations. What connects both theories is the use of our experience and competence in one domain to guide our thinking in a distinct domain. Anderson describes very many instances of neural reuse that do not obviously involve the sensorimotor system, and hence do not involve semantic inheritance. He takes this to show that semantic inheritance may be a “side effect” (see sect. 4.6) of neural reuse. I will argue that it is only when reuse is accompanied by semantic inheritance that you find any bootstrapping from low-level cognitive functions to high-level cognitive functions. This follows from an argument Anderson himself makes against Susan Hurley's (2008) shared circuits model (SCM). Therefore, the question of what kinds of reuse support semantic inheritance (a question Anderson himself raises in sect. 7) becomes a particularly pressing issue for the embodied cognition research programme. I will finish up by suggesting that neural reuse and semantic inheritance may actually be much more closely tied than Anderson suggests.
We can see how semantic inheritance is required for bootstrapping by considering Anderson's discussion of Susan Hurley's (2008) shared circuits model (SCM). The model is complex, and I shall restrict my discussion to layer 3 of SCM, which describes how functional mechanisms used to predict sensory feedback in the control of motor behaviour might be reused to “simulate” the motor processes that stand behind the observed behaviour of another. This simulation is hypothesised to take the form of “mirroring” that can underwrite the copying of instrumental behaviour either in the form of priming, emulation, or imitation. Anderson worries that the inputs and outputs required for mirroring are “impoverished” and “abstract” when compared to those inherited from layer 2. When I perform an action myself, for instance, the action is represented from my own point of view. Anderson supposes that when I observe another's action, I must represent the other's action from a third-person point of view. Hence, the progression from layer 2 to layer 3 would seem to require a translation of a first-person representation of action into a third-person representation. Without some explanation of how this translation gets effected, we will not have shown how high-level cognitive abilities like imitative learning can have their basis in the reuse of low-level sensorimotor representation.
This problem Anderson has identified for SCM would however seem to apply equally to MRH. What allows the control mechanisms found at layer 2 to be reused at layer 3 are the functional properties of those control mechanisms. According to MRH, it is a neural region's functional properties that allow a region used in one domain to get reused in a distinct domain. The inheritance of functional properties falls some way short of guaranteeing semantic inheritance. Functional inheritance doesn't on its own explain the abstraction and informational impoverishment you find as you move from lower-level sensorimotor behaviour to higher-level cognition. If this is right, it seems to follow that neural reuse won't suffice for bootstrapping.
Hurley's SCM may however have resources for responding to this problem that are different from those outlined by Anderson in his target article. What is missing from Anderson's framing of the problem is any mention of the sensorimotor associations that drive the predictions at layers 2 and 3 of SCM. Predictions of the sensory effects of movement are possible at layer 2 only because the motor system has learned that movements of a given type are correlated with certain sensory effects. Hurley followed Cecilia Heyes in thinking of this learning as arising in development through associations that wire sensory neurons (in superior temporal sulcus, for example) together with motor neurons (in premotor and parietal cortices; see Heyes [Reference Heyes2010] for more on a recent presentation of this hypothesis). Crucially, Hurley is assuming that the sensory inputs from one's own movement and from the movement of others are similar enough for sensory neurons to respond to both without distinguishing them. Thus, sensorimotor associations can underwrite an “inference” from the sensory effects of observed behaviour to the motor processes that tend to cause behaviour. In this way, sensorimotor associations can be used both to control the sensory effects of movement and to simulate the movements that have similar sensory effects when carried out by others. For SCM then, it is associative learning that delivers the kind of semantic inheritance required for bootstrapping.
I finish by drawing a tentative moral for MRH. The functional inheritance that underpins neural reuse might bear cognitive fruit only when it is accompanied by semantic inheritance. Reuse of functional mechanisms in SCM is understood as simulation that establishes a space of shared meaning. Semantic inheritance, as appealed to in concept empiricism and conceptual metaphor theories, is also naturally understood as a form of simulation which opens up a space of shared meaning. While neural reuse could well turn out to be a “fundamental organisational principle” of the brain, the pressing question that remains is how neural reuse could deliver a shared space of meaning of a kind that supports bootstrapping.
You have
Access
Evolution has devoted by far and away the largest part of its history to building organisms that can move around in a dynamic environment, sensing their environments in ways conducive to their own survival and reproduction (Brooks Reference Brooks1991). The challenge to cognitive scientists is to explain how the strategies organisms use to solve these basic problems of perception and action scale up to the strategies humans use in solving abstract higher-level problems. I call this the “bootstrapping challenge.” Embodied cognitive science offers a programmatic response to the bootstrapping challenge that attempts to show how high-level problem solving might have been built upon the foundation of a substrate of perception and sensorimotor control. The ascent from sensing and moving to thinking, planning, and language understanding is an incremental and gradual one, and a key strategy may have been the redeployment of sensorimotor capacities to perform high-level cognitive tasks.
Anderson has done the embodied cognition community the enormous service of framing a global hypothesis about how these incremental changes might have taken place in our brains over the course of evolution. The central claim of his massive redeployment hypothesis (MRH) is that more recent cognitive functions such as those involved in abstract problem solving might have their origin in the reuse of evolutionarily older neural circuits that served biologically basic functions. In this commentary, I want to take up Anderson's claim that the principle guiding reuse is “functional inheritance” and not “semantic inheritance.” By “semantic inheritance,” I mean the kind of relation that concept empiricists and conceptual metaphor theories take to hold between concepts and sensorimotor representations. What connects both theories is the use of our experience and competence in one domain to guide our thinking in a distinct domain. Anderson describes very many instances of neural reuse that do not obviously involve the sensorimotor system, and hence do not involve semantic inheritance. He takes this to show that semantic inheritance may be a “side effect” (see sect. 4.6) of neural reuse. I will argue that it is only when reuse is accompanied by semantic inheritance that you find any bootstrapping from low-level cognitive functions to high-level cognitive functions. This follows from an argument Anderson himself makes against Susan Hurley's (2008) shared circuits model (SCM). Therefore, the question of what kinds of reuse support semantic inheritance (a question Anderson himself raises in sect. 7) becomes a particularly pressing issue for the embodied cognition research programme. I will finish up by suggesting that neural reuse and semantic inheritance may actually be much more closely tied than Anderson suggests.
We can see how semantic inheritance is required for bootstrapping by considering Anderson's discussion of Susan Hurley's (2008) shared circuits model (SCM). The model is complex, and I shall restrict my discussion to layer 3 of SCM, which describes how functional mechanisms used to predict sensory feedback in the control of motor behaviour might be reused to “simulate” the motor processes that stand behind the observed behaviour of another. This simulation is hypothesised to take the form of “mirroring” that can underwrite the copying of instrumental behaviour either in the form of priming, emulation, or imitation. Anderson worries that the inputs and outputs required for mirroring are “impoverished” and “abstract” when compared to those inherited from layer 2. When I perform an action myself, for instance, the action is represented from my own point of view. Anderson supposes that when I observe another's action, I must represent the other's action from a third-person point of view. Hence, the progression from layer 2 to layer 3 would seem to require a translation of a first-person representation of action into a third-person representation. Without some explanation of how this translation gets effected, we will not have shown how high-level cognitive abilities like imitative learning can have their basis in the reuse of low-level sensorimotor representation.
This problem Anderson has identified for SCM would however seem to apply equally to MRH. What allows the control mechanisms found at layer 2 to be reused at layer 3 are the functional properties of those control mechanisms. According to MRH, it is a neural region's functional properties that allow a region used in one domain to get reused in a distinct domain. The inheritance of functional properties falls some way short of guaranteeing semantic inheritance. Functional inheritance doesn't on its own explain the abstraction and informational impoverishment you find as you move from lower-level sensorimotor behaviour to higher-level cognition. If this is right, it seems to follow that neural reuse won't suffice for bootstrapping.
Hurley's SCM may however have resources for responding to this problem that are different from those outlined by Anderson in his target article. What is missing from Anderson's framing of the problem is any mention of the sensorimotor associations that drive the predictions at layers 2 and 3 of SCM. Predictions of the sensory effects of movement are possible at layer 2 only because the motor system has learned that movements of a given type are correlated with certain sensory effects. Hurley followed Cecilia Heyes in thinking of this learning as arising in development through associations that wire sensory neurons (in superior temporal sulcus, for example) together with motor neurons (in premotor and parietal cortices; see Heyes [Reference Heyes2010] for more on a recent presentation of this hypothesis). Crucially, Hurley is assuming that the sensory inputs from one's own movement and from the movement of others are similar enough for sensory neurons to respond to both without distinguishing them. Thus, sensorimotor associations can underwrite an “inference” from the sensory effects of observed behaviour to the motor processes that tend to cause behaviour. In this way, sensorimotor associations can be used both to control the sensory effects of movement and to simulate the movements that have similar sensory effects when carried out by others. For SCM then, it is associative learning that delivers the kind of semantic inheritance required for bootstrapping.
I finish by drawing a tentative moral for MRH. The functional inheritance that underpins neural reuse might bear cognitive fruit only when it is accompanied by semantic inheritance. Reuse of functional mechanisms in SCM is understood as simulation that establishes a space of shared meaning. Semantic inheritance, as appealed to in concept empiricism and conceptual metaphor theories, is also naturally understood as a form of simulation which opens up a space of shared meaning. While neural reuse could well turn out to be a “fundamental organisational principle” of the brain, the pressing question that remains is how neural reuse could deliver a shared space of meaning of a kind that supports bootstrapping.