Introduction
The human brain is capable of coordinating a wide variety of complex adaptive behaviors. When appropriately wired over development, the same collection of cell bodies and synapses can subtly alter their patterns of activity and communication to coordinate a range of vastly different behaviors. In After Phrenology, Michael Anderson (Reference Anderson2014) presents a unique and elegant thesis that attempts to explain how, over the course of evolutionary and developmental time, the human brain has come to afford such a range of flexible behaviors. Specifically, it is proposed that the mechanisms of natural selection allow existing brain regions to recombine into novel architectures capable of executing new behavioral patterns. Although Anderson's attempt to synthesize a wide variety of literature across multiple fields is commendable, some of the evidence that he uses to defend this “neural reuse” hypothesis may be overstated.
In this commentary, we will provide evidence that suggests potential challenges to the thesis presented in After Phrenology. We will first discuss a major result provided in support of the “neural reuse” hypothesis – namely that meta-analytic evidence from human functional neuroimaging experiments demonstrate a dearth of specificity in the human brain, and suggest that the conclusions of this experiment may reflect limits inherent to the use of meta-analytic methods within cognitive neuroscience, rather than the re-instantiation of existing neural regions for novel ends per se. Second, we will suggest that brain function is most appropriately categorized according to the functional capacity of each brain system, rather than the specific task states that elicit its activity. Indeed, in considering the computational capacities of independent brain regions, we will make the argument that computational specialization is not only abundant in the brain, but also that it would be difficult to imagine a working brain that did not contain such specialization.
The limits of meta-analysis
One of the key lines of evidence presented in support of the “neural reuse” hypothesis comes from meta-analytic studies of functional neuroimaging data, such as BrainMap (Laird et al. Reference Laird, Lancaster and Fox2005) or NeuroSynth (Yarkoni et al. Reference Yarkoni, Poldrack, Nichols, Van Essen and Wager2011). Using the BrainMap database, the author describes an experiment that aimed to determine the amount of functional diversity associated with each region of the brain (Anderson et al. Reference Anderson, Kinnison and Pessoa2013). Specifically, the experiment was designed to determine whether given brain regions (here defined using a voxel-wise searchlight approach) were associated with a range of different tasks in a number of unique task categories. The results of this experiment suggested that there was almost no evidence of functional specialization in any region of the brain. Indeed, the author goes as far as to conclude that all brain regions are involved in multiple functions and then takes this as evidence for his hypothesis that all “individual neural elements are put to use for multiple cognitive and behavioral ends” (sect. 2, para. 1).
However, the underlying assumption is that the tasks used in these databases measure “cognitive and behavioral ends” that directly map onto neural computation, which they do not. Instead, these tasks measure surrogate patterns of brain activity elicited by particular task demands, which themselves may rely on a multitude of fundamental neural computations. For example, a neuroscientific experiment might be designed to interrogate the neural mechanism of language production using alterations in the BOLD response, as is measured by fMRI. Assuming appropriate analysis and power, the experimenter will most likely discover an increase in the BOLD response associated with the language task in the left lateral frontal lobe, a region commonly associated with language production (Price Reference Price2010). However, it would be problematic to compare the resultant statistical map with a similar pattern derived from a different, yet related experiment (e.g., an experiment measuring working memory demands associated with word formation) and, after finding extensive overlap between the two, conclude that this convergence represented the functional diversity of the left lateral frontal lobe. Indeed, it is much more plausible that this region was performing a similar computation in both experiments (e.g., the top-down manipulation of language-related information). Put another way, even if the brain is perfectly modular in the computations it performs, we would expect a mixture at the level of behavior that would mimic flexible recruitment. That is, “absence of evidence is not evidence of absence” when it comes to modularity.
Importantly, few studies in cognitive neuroscience have been designed to accurately describe the taxonomy of neural computations across the brain. Instead, cognitive neuroscientists propose hypotheses at the level of psychological processes and/or behaviors, which may or may not reflect the true dimensions of neural organization. As such, the maps that are created (and hence populate meta-analytic databases) will not necessarily reflect an accurate taxonomy of brain function. In addition, there are likely to be a number of neural computations that are simply not amenable to analysis using fMRI. For example, Dubois and colleagues (Reference Dubois, de Berker and Tsao2015) recently showed that, during a face identification task, the identity of individual faces could be effectively decoded from posterior regions of ventral occipital cortex using both fMRI and direct neuronal recordings, whereas in more anterior regions the information was only able to be decoded from the neuronal signals. Together, this suggests that the data within these meta-analytic repositories are necessarily limited and, hence, conclusions deriving from these studies should be tempered with an appropriate level of caution.
Specialization as a computational entity
Given the nature of the computational demands placed on the brain, it is quite possible that a number of functional units are inherently modular. For example, there are cells within the retina that fire rapidly when exposed to a particular pattern of light, such as a thin, dark horizontal line on a white background,that falls within their receptive field. The information from these cells, when appropriately combined and fed forward to thalamic and occipital cortical regions, underpins the very building blocks of visual perception. However, should the pattern of light exposed to this receptive field alter in some small way (e.g., the light hitting the retina may shift its angle of approach by some small degree; or the contrast-to-noise ratio along the outlines of the image might change), the retinal cells will quickly decrease their firing rate. Although there is no doubt that these same retinal cells will be used in the completion of a wide variety of behavioral tasks and for very different ends (e.g., tracking the flight of a ball that one is attempting to catch versus deciphering the presence or absence of a particular letter that may be associated with a cognitive task), the computations that they perform are specialized. That is, these retinal cells display a specialized computational architecture.
In contrast, other brain regions display a much more flexible computational repertoire. For example, there are neurons within the prefrontal cortex that can switch the target of their receptive fields according to whichever cue is currently of motivational value to an animal (Barbas & Zikopoulos Reference Barbas and Zikopolous2007). Similarly, experiments using direct cell recordings in non-human primates have shown that prefrontal cortical neurons can change their sensitivity to patterns of information based on shifting goals (Mante et al. Reference Mante, Sussillo, Shenoy and Newsome2013) and can also alter the direction of information flow within frontal circuits that control sensorimotor decision making (Siegel et al. Reference Siegel, Buschman and Miller2015). In humans, distinct regions of the prefrontal cortex modulate their functional connectivity as a function of task-set (Sakai & Passingham Reference Sakai and Passingham2003). Importantly, although these neural regions will presumably be among the most commonly reused regions, the computations that they are performing will likely be highly specialized. That is, the neural systems processing the information from whichever object is the current focus of the animal, be it exogenous (such as counting a series of objects in the world) or endogenous (such as mentally counting the number of papers published by a rival colleague), will be computed upon in a similar way by the neurons within the frontoparietal control networks.
Attributing computational specificity to a particular region is not to suggest that the brain is composed of discrete modules, each performing specialized computations. Rather, subsections of the overall brain network display properties (such as specialized architecture, cell types, or connectivity), which afford computational roles to individual brain regions that emerge through the interaction with the rest of the network. For example, there might exist a subcircuit incorporating the motor cortex, putamen, and motor thalamus that is particularly important for the execution of a specific motor pattern (say, moving the right index finger), but the output of this circuitry would be simply unable to function out of context (i.e. in vitro), as its computational role arises directly from its location in space (i.e., within the brain, which is itself within a larger organism) and time (i.e., the manner in which its been developed over learning). As a field, cognitive neuroscience has been attempting to understand these computations by manipulating behavior through the use of neuropsychological tasks; however, to properly understand the functional role of each subnetwork within the brain, it may be more useful to describe the computational role of each circuit in context. If such an undertaking were performed, we imagine that some regions perform a well-defined computational role regardless of context (e.g., retinal cells processing light), whereas others are more context-dependent (like prefrontal cortex), flexibly altering their role based on the activity of the overall brain network.
Refining cognitive ontologies
Many of the issues highlighted in this commentary could be solved through the progressive refinement of a cognitive ontology that more accurately describes the relationships between the many and varied computational capacities instantiated within the brain (Poldrack et al. Reference Poldrack, Kittur, Kalar and Miller2011). Indeed, given our current lack of conceptual clarity as to the appropriate ontological framework in which to interrogate the brain, any attempt to classify behaviors according to our current framework will necessarily be flawed, hence leading to faulty distinctions in the neural architecture of behavior. To proceed, we agree that we should seek to appropriately define the “psychological factors that best capture and account for the differential activity of the brain in various circumstances” (sect. 6, para. 3). However, we differ in our predictions of the likely neurocognitive architectures that will best explain the functional landscape of the brain. Whereas the author would advocate for an architecture in which all “individual neural elements are put to use for multiple cognitive and behavioral ends” (sect. 2, para. 1), we propose that there will likely be a range of fundamental neural computations that, when combined with information regarding the context of the individual neural population within the larger neuronal network (such as the structural connectivity profile or the relative proportion of different neuromodulatory afferents to a region), will effectively explain the functional repertoire of the brain.
That is not to say that these computations will necessarily make a region more “specialized” in a behavioral sense, but rather that the computational abilities of a given region will define its involvement across multiple tasks. For example, regions within the prefrontal cortex that hold information online over time (Curtis Reference Curtis2006; Sakai & Passingham Reference Sakai and Passingham2003) can be utilized to maintain information during multiple different behavioral tasks, such as those invoking language, social, mathematical and logical reasoning capacities. However, in each case, the computation performed by the region is likely to be similar, whereas the behavioral outcome will differ, depending on the context in which the computation is deployed. Within this framework, regions can also develop over evolutionary or developmental time to become computationally specialized (Bassett et al. Reference Bassett, Yang, Wymbs and Grafton2015) and hence become recruited only under the precise situations that require their involvement (such as Broca's area in response to the processing of language-related information; Price Reference Price2010).
Conclusion
Together, our arguments suggest a possible reformulation of the author's main hypothesis, which would place specialization and reuse together along a functional spectrum, in which the computational properties associated with each brain system define its potential involvement in the mechanistic definition of a range of behavioral capacities.
Introduction
The human brain is capable of coordinating a wide variety of complex adaptive behaviors. When appropriately wired over development, the same collection of cell bodies and synapses can subtly alter their patterns of activity and communication to coordinate a range of vastly different behaviors. In After Phrenology, Michael Anderson (Reference Anderson2014) presents a unique and elegant thesis that attempts to explain how, over the course of evolutionary and developmental time, the human brain has come to afford such a range of flexible behaviors. Specifically, it is proposed that the mechanisms of natural selection allow existing brain regions to recombine into novel architectures capable of executing new behavioral patterns. Although Anderson's attempt to synthesize a wide variety of literature across multiple fields is commendable, some of the evidence that he uses to defend this “neural reuse” hypothesis may be overstated.
In this commentary, we will provide evidence that suggests potential challenges to the thesis presented in After Phrenology. We will first discuss a major result provided in support of the “neural reuse” hypothesis – namely that meta-analytic evidence from human functional neuroimaging experiments demonstrate a dearth of specificity in the human brain, and suggest that the conclusions of this experiment may reflect limits inherent to the use of meta-analytic methods within cognitive neuroscience, rather than the re-instantiation of existing neural regions for novel ends per se. Second, we will suggest that brain function is most appropriately categorized according to the functional capacity of each brain system, rather than the specific task states that elicit its activity. Indeed, in considering the computational capacities of independent brain regions, we will make the argument that computational specialization is not only abundant in the brain, but also that it would be difficult to imagine a working brain that did not contain such specialization.
The limits of meta-analysis
One of the key lines of evidence presented in support of the “neural reuse” hypothesis comes from meta-analytic studies of functional neuroimaging data, such as BrainMap (Laird et al. Reference Laird, Lancaster and Fox2005) or NeuroSynth (Yarkoni et al. Reference Yarkoni, Poldrack, Nichols, Van Essen and Wager2011). Using the BrainMap database, the author describes an experiment that aimed to determine the amount of functional diversity associated with each region of the brain (Anderson et al. Reference Anderson, Kinnison and Pessoa2013). Specifically, the experiment was designed to determine whether given brain regions (here defined using a voxel-wise searchlight approach) were associated with a range of different tasks in a number of unique task categories. The results of this experiment suggested that there was almost no evidence of functional specialization in any region of the brain. Indeed, the author goes as far as to conclude that all brain regions are involved in multiple functions and then takes this as evidence for his hypothesis that all “individual neural elements are put to use for multiple cognitive and behavioral ends” (sect. 2, para. 1).
However, the underlying assumption is that the tasks used in these databases measure “cognitive and behavioral ends” that directly map onto neural computation, which they do not. Instead, these tasks measure surrogate patterns of brain activity elicited by particular task demands, which themselves may rely on a multitude of fundamental neural computations. For example, a neuroscientific experiment might be designed to interrogate the neural mechanism of language production using alterations in the BOLD response, as is measured by fMRI. Assuming appropriate analysis and power, the experimenter will most likely discover an increase in the BOLD response associated with the language task in the left lateral frontal lobe, a region commonly associated with language production (Price Reference Price2010). However, it would be problematic to compare the resultant statistical map with a similar pattern derived from a different, yet related experiment (e.g., an experiment measuring working memory demands associated with word formation) and, after finding extensive overlap between the two, conclude that this convergence represented the functional diversity of the left lateral frontal lobe. Indeed, it is much more plausible that this region was performing a similar computation in both experiments (e.g., the top-down manipulation of language-related information). Put another way, even if the brain is perfectly modular in the computations it performs, we would expect a mixture at the level of behavior that would mimic flexible recruitment. That is, “absence of evidence is not evidence of absence” when it comes to modularity.
Importantly, few studies in cognitive neuroscience have been designed to accurately describe the taxonomy of neural computations across the brain. Instead, cognitive neuroscientists propose hypotheses at the level of psychological processes and/or behaviors, which may or may not reflect the true dimensions of neural organization. As such, the maps that are created (and hence populate meta-analytic databases) will not necessarily reflect an accurate taxonomy of brain function. In addition, there are likely to be a number of neural computations that are simply not amenable to analysis using fMRI. For example, Dubois and colleagues (Reference Dubois, de Berker and Tsao2015) recently showed that, during a face identification task, the identity of individual faces could be effectively decoded from posterior regions of ventral occipital cortex using both fMRI and direct neuronal recordings, whereas in more anterior regions the information was only able to be decoded from the neuronal signals. Together, this suggests that the data within these meta-analytic repositories are necessarily limited and, hence, conclusions deriving from these studies should be tempered with an appropriate level of caution.
Specialization as a computational entity
Given the nature of the computational demands placed on the brain, it is quite possible that a number of functional units are inherently modular. For example, there are cells within the retina that fire rapidly when exposed to a particular pattern of light, such as a thin, dark horizontal line on a white background,that falls within their receptive field. The information from these cells, when appropriately combined and fed forward to thalamic and occipital cortical regions, underpins the very building blocks of visual perception. However, should the pattern of light exposed to this receptive field alter in some small way (e.g., the light hitting the retina may shift its angle of approach by some small degree; or the contrast-to-noise ratio along the outlines of the image might change), the retinal cells will quickly decrease their firing rate. Although there is no doubt that these same retinal cells will be used in the completion of a wide variety of behavioral tasks and for very different ends (e.g., tracking the flight of a ball that one is attempting to catch versus deciphering the presence or absence of a particular letter that may be associated with a cognitive task), the computations that they perform are specialized. That is, these retinal cells display a specialized computational architecture.
In contrast, other brain regions display a much more flexible computational repertoire. For example, there are neurons within the prefrontal cortex that can switch the target of their receptive fields according to whichever cue is currently of motivational value to an animal (Barbas & Zikopoulos Reference Barbas and Zikopolous2007). Similarly, experiments using direct cell recordings in non-human primates have shown that prefrontal cortical neurons can change their sensitivity to patterns of information based on shifting goals (Mante et al. Reference Mante, Sussillo, Shenoy and Newsome2013) and can also alter the direction of information flow within frontal circuits that control sensorimotor decision making (Siegel et al. Reference Siegel, Buschman and Miller2015). In humans, distinct regions of the prefrontal cortex modulate their functional connectivity as a function of task-set (Sakai & Passingham Reference Sakai and Passingham2003). Importantly, although these neural regions will presumably be among the most commonly reused regions, the computations that they are performing will likely be highly specialized. That is, the neural systems processing the information from whichever object is the current focus of the animal, be it exogenous (such as counting a series of objects in the world) or endogenous (such as mentally counting the number of papers published by a rival colleague), will be computed upon in a similar way by the neurons within the frontoparietal control networks.
Attributing computational specificity to a particular region is not to suggest that the brain is composed of discrete modules, each performing specialized computations. Rather, subsections of the overall brain network display properties (such as specialized architecture, cell types, or connectivity), which afford computational roles to individual brain regions that emerge through the interaction with the rest of the network. For example, there might exist a subcircuit incorporating the motor cortex, putamen, and motor thalamus that is particularly important for the execution of a specific motor pattern (say, moving the right index finger), but the output of this circuitry would be simply unable to function out of context (i.e. in vitro), as its computational role arises directly from its location in space (i.e., within the brain, which is itself within a larger organism) and time (i.e., the manner in which its been developed over learning). As a field, cognitive neuroscience has been attempting to understand these computations by manipulating behavior through the use of neuropsychological tasks; however, to properly understand the functional role of each subnetwork within the brain, it may be more useful to describe the computational role of each circuit in context. If such an undertaking were performed, we imagine that some regions perform a well-defined computational role regardless of context (e.g., retinal cells processing light), whereas others are more context-dependent (like prefrontal cortex), flexibly altering their role based on the activity of the overall brain network.
Refining cognitive ontologies
Many of the issues highlighted in this commentary could be solved through the progressive refinement of a cognitive ontology that more accurately describes the relationships between the many and varied computational capacities instantiated within the brain (Poldrack et al. Reference Poldrack, Kittur, Kalar and Miller2011). Indeed, given our current lack of conceptual clarity as to the appropriate ontological framework in which to interrogate the brain, any attempt to classify behaviors according to our current framework will necessarily be flawed, hence leading to faulty distinctions in the neural architecture of behavior. To proceed, we agree that we should seek to appropriately define the “psychological factors that best capture and account for the differential activity of the brain in various circumstances” (sect. 6, para. 3). However, we differ in our predictions of the likely neurocognitive architectures that will best explain the functional landscape of the brain. Whereas the author would advocate for an architecture in which all “individual neural elements are put to use for multiple cognitive and behavioral ends” (sect. 2, para. 1), we propose that there will likely be a range of fundamental neural computations that, when combined with information regarding the context of the individual neural population within the larger neuronal network (such as the structural connectivity profile or the relative proportion of different neuromodulatory afferents to a region), will effectively explain the functional repertoire of the brain.
That is not to say that these computations will necessarily make a region more “specialized” in a behavioral sense, but rather that the computational abilities of a given region will define its involvement across multiple tasks. For example, regions within the prefrontal cortex that hold information online over time (Curtis Reference Curtis2006; Sakai & Passingham Reference Sakai and Passingham2003) can be utilized to maintain information during multiple different behavioral tasks, such as those invoking language, social, mathematical and logical reasoning capacities. However, in each case, the computation performed by the region is likely to be similar, whereas the behavioral outcome will differ, depending on the context in which the computation is deployed. Within this framework, regions can also develop over evolutionary or developmental time to become computationally specialized (Bassett et al. Reference Bassett, Yang, Wymbs and Grafton2015) and hence become recruited only under the precise situations that require their involvement (such as Broca's area in response to the processing of language-related information; Price Reference Price2010).
Conclusion
Together, our arguments suggest a possible reformulation of the author's main hypothesis, which would place specialization and reuse together along a functional spectrum, in which the computational properties associated with each brain system define its potential involvement in the mechanistic definition of a range of behavioral capacities.