There is much to like in Michael Anderson's new book, After Phrenology (Reference Anderson2014). First and foremost, he synthesizes a large and diverse literature to support a sophisticated and evolutionarily grounded approach to brain science. He describes the brain as a system in constant flux, transiently coalescing competing and cooperating neural assemblies in the service of action. The brain, he reasons, evolved to engage successfully with the environment rather than to represent the world accurately. He argues convincingly that this difference is underappreciated by cognitive neuroscience, which instead tends to look for neural activity that encodes objective properties in the environment as if they are context- and perceiver-free.
Anderson's description of the brain as a dynamic information processing device that responds to structured signals from the environment to guide action dovetails with his and other theories of neural reuse (Anderson Reference Anderson2010; Chang et al. Reference Chang, Brent, Adams, Klein, Pearson, Watson and Platt2013; Dehaene & Cohen Reference Dehaene and Cohen2007; Marcus Reference Marcus2006; Parkinson & Wheatley Reference Parkinson and Wheatley2015). The basic idea here is that the brain solves new problems by repurposing preexisting neural architecture that solved structurally similar problems. This evolutionary and action-focused account of the brain is interwoven with the “extended mind” thesis, that intelligence is subserved by iterative interaction with the environment. These ideas form a solid, well-articulated, and reasonable book.
Less convincingly, Anderson argues that cognitive neuroscience is held in thrall to a “massively modular” account of the brain. From Anderson's vantage point, the field's modus operandi is to test for and report the single unified function of each brain area, as if the brain is a “collection of organs” (p. 292). As he points out, a brain composed mostly of special-purpose modules would be inefficient, prohibitively large, and inconsistent with the bulk of currently available evidence. We agree with Anderson that massive modularity is untenable, but so too does all but a vanishing minority of neuroscientists. The near-consensus view among contemporary neuroscientists is that most cortical regions exhibit considerable functional heterogeneity. Hence, the pleas to jettison modularity feel dated.
Anderson correctly asserts that the interconnected complexity and flexibility of neural activity presents immense challenges for brain science. The categories that we use to reflect on our own cognition are unlikely to match the functional organization of the brain, and a given brain region likely serves multiple functional roles, which can differ depending on the current environmental context and the networks in which the brain region is currently participating. However, we are somewhat more optimistic that the field can make significant headway in uncovering the structure of operations underlying human thought and behavior. This optimism is fueled in large part by more than a decade of cognitive neuroscience research that has honed techniques for multivariate pattern analysis of functional neuroimaging data. After Phrenology briefly alludes to this approach (Kriegeskorte et al. Reference Kriegeskorte, Mur and Bandettini2008a), but we believe that it is of particular relevance to Anderson's proposed framework and to his suggested ways forward for the field, and hence, merits further emphasis.
We now know that the brain encodes many kinds of information (e.g., memories, tastes, reach trajectories) in neuronal population codes, rather than in the magnitude of single units' responses (Georgopoulos et al. Reference Georgopoulos, Kettner and Schwartz1988; Lin et al. Reference Lin, Osan and Tsien2006; Pouget et al. Reference Pouget, Dayan and Zemel2000; Simon et al. Reference Simon, de Araujo, Gutierrez and Nicolelis2006). Fortunately, much of the information contained in neuronal population codes can be gleaned from the indirect and spatially coarse measurements of brain activity that are currently available to researchers studying human brain function (e.g., functional magnetic resonance imaging, fMRI; Kriegeskorte et al. Reference Kriegeskorte, Mur, Ruff, Kiani, Bodurka, Esteky, Tanaka and Bandettini2008b).
However, much of the information present in neural population codes is lost when analyzing the magnitude of responses of a given functional unit in isolation. For example, one of the first studies that analyzed fMRI data in terms of population response vectors demonstrated that brain regions that are typically characterized in terms of the categories of stimuli eliciting maximal responses (e.g., the fusiform face area) also differentiate between stimulus categories that evoke submaximal responses, such as chairs, shoes, cats, and houses (Haxby et al. Reference Haxby, Gobbini, Furey, Ishai, Schouten and Pietrini2001). This information was carried in unique and consistent topographic arrangements of responses for each stimulus category, and such information had previously been missed by analyzing data in terms of response magnitude only. Hence, analyzing neural responses only in terms of overall response magnitude can lead to an exaggeratedly modular conceptualization of brain function, as it can underestimate the number of brain regions involved in processing a stimulus category, as well as the number of stimuli a given brain region is involved in processing. Therefore, the brain may be even less modular than the functional fingerprinting method that Anderson employs would suggest, given that this method appears to only take into account response magnitude.
Analyzing data in terms of population response vectors, rather than overall response magnitude, can also provide insight into how information is organized within brain regions. To this end, researchers have applied both supervised and unsupervised machine learning techniques when analyzing the distributed topographies of activity within brain regions in order to characterize the distinctions that those brain regions make about various classes of stimuli. For example, by characterizing the similarity structure of population response vectors to visual stimuli within brain regions of the ventral visual pathway, researchers have gained insight into the distinctions that brain areas within that pathway make about sets of stimuli, and hence, the operations that they may contribute to visual processing (Connolly et al. Reference Connolly, Gobbini, Haxby, Kriegeskorte and Kreiman2012). Studies using unsupervised learning techniques (e.g., clustering, multidimensional scaling) have the notable benefit of being wholly data-driven in the structure that is revealed. Hence, such approaches will be useful in following Anderson's call for a science focused on illuminating the psychological factors that best account for observed differences in neural activity, rather than on attempts to map psychological constructs onto the brain.
Considering the information content of multivoxel response patterns, rather than focusing solely on response magnitude, also allows researchers to probe the functional significance of overlap in the brain regions that are recruited to process diverse domains of contents. Univariate, subtraction-based approaches to analyzing brain activity are relatively well suited to asking questions such as, “Is this brain area involved in both Task X and Task Y?” and “Is this brain area recruited ‘more’ for Task X than Task Y?” In contrast, analyzing neural population responses is an approach that is relatively better suited to uncovering how a brain region might contribute to various tasks. Functional overlap can reflect the recruitment of shared or distinct mechanisms, and these possibilities can sometimes be disentangled by studying neural population response vectors. For example, although univariate, magnitude-based analyses of fMRI data suggested that most voxels in a region of lateral occipital-temporal cortex responded to both static pictures of bodies and movement displays, subsequent analysis of the population responses in that region demonstrated that the response patterns to these two sets of stimuli, though spatially overlapping, were unrelated (Peelen & Downing Reference Peelen and Downing2007). Observations such as these are consistent with Anderson's suggestion that a brain region's computational role can differ depending on the circumstances surrounding its recruitment (e.g., the processing demands of a stimulus, the other brain regions with which it is interacting). Hence, analyzing population response vectors can help determine whether spatially overlapping functional activity for two different tasks reflects the recruitment of the same or dissociable underlying mechanisms.
The analysis of population response vectors can also reveal when functional overlap reflects the recruitment of a common underlying neural operation across multiple experimental contexts. For example, a machine learning classifier trained to distinguish between population response vectors within the superior parietal lobule (SPL) associated with leftward and rightward eye movements can correctly distinguish between population response vectors in the SPL corresponding to mental subtraction and mental arithmetic (Knops et al. Reference Knops, Thirion, Hubbard, Michel and Dehaene2009). This suggests that the SPL contributes shared operations when shifting spatial attention within the external environment (during saccades) and within internal mental representations (when performing arithmetic). This likely reflects a case of neural reuse through which functional architecture with a preexisting role in shifting attention in the external environment is exploited in order to perform analogous operations on mental representations, and hence, to perform the culturally learned skill of arithmetic.
We recently used classification and similarity structure-based analyses of neural population response vectors to test for a common neural encoding of egocentric distance in spatial, temporal, and social frames of reference (Parkinson et al. Reference Parkinson, Liu and Wheatley2014). Cross-domain decoding analyses, as well as analyses of the similarity structure of population responses, revealed that areas such as the right anterior inferior parietal lobule, which has a long-established role in representing physical space, organize information not based on distance category (e.g., “Is this a social or temporal distance?”), but based on distance from the self (i.e., “Is this relatively close to or far away from me?”). These results are consistent with Anderson's suggestion that the brain is largely concerned with computing behaviorally relevant information, such as distance from oneself (and hence, relevance for action). Further, these results are consistent with the possibility that neural mechanisms for encoding information about physical space may have been reused in order to operate on increasingly abstract contents, such as temporal and social relationships (Parkinson & Wheatley Reference Parkinson and Wheatley2013; Reference Parkinson and Wheatley2015).
Understanding how the brain gives rise to a given facet of cognition or behavior requires not only identifying the neural substrates involved, but also uncovering the kinds of operations performed by those substrates. We are optimistic that advances in neuroimaging methods, such as the analysis of neural population response vectors, will be useful in moving toward a deeper and more veridical understanding of functional brain organization and of the flexible and adaptive computations that underlie everyday thought and behavior. In so doing, cognitive neuroscience has the potential to elucidate the particular functional capacities that may be reused over the course of evolution and development to solve novel problems.
You have
Access
There is much to like in Michael Anderson's new book, After Phrenology (Reference Anderson2014). First and foremost, he synthesizes a large and diverse literature to support a sophisticated and evolutionarily grounded approach to brain science. He describes the brain as a system in constant flux, transiently coalescing competing and cooperating neural assemblies in the service of action. The brain, he reasons, evolved to engage successfully with the environment rather than to represent the world accurately. He argues convincingly that this difference is underappreciated by cognitive neuroscience, which instead tends to look for neural activity that encodes objective properties in the environment as if they are context- and perceiver-free.
Anderson's description of the brain as a dynamic information processing device that responds to structured signals from the environment to guide action dovetails with his and other theories of neural reuse (Anderson Reference Anderson2010; Chang et al. Reference Chang, Brent, Adams, Klein, Pearson, Watson and Platt2013; Dehaene & Cohen Reference Dehaene and Cohen2007; Marcus Reference Marcus2006; Parkinson & Wheatley Reference Parkinson and Wheatley2015). The basic idea here is that the brain solves new problems by repurposing preexisting neural architecture that solved structurally similar problems. This evolutionary and action-focused account of the brain is interwoven with the “extended mind” thesis, that intelligence is subserved by iterative interaction with the environment. These ideas form a solid, well-articulated, and reasonable book.
Less convincingly, Anderson argues that cognitive neuroscience is held in thrall to a “massively modular” account of the brain. From Anderson's vantage point, the field's modus operandi is to test for and report the single unified function of each brain area, as if the brain is a “collection of organs” (p. 292). As he points out, a brain composed mostly of special-purpose modules would be inefficient, prohibitively large, and inconsistent with the bulk of currently available evidence. We agree with Anderson that massive modularity is untenable, but so too does all but a vanishing minority of neuroscientists. The near-consensus view among contemporary neuroscientists is that most cortical regions exhibit considerable functional heterogeneity. Hence, the pleas to jettison modularity feel dated.
Anderson correctly asserts that the interconnected complexity and flexibility of neural activity presents immense challenges for brain science. The categories that we use to reflect on our own cognition are unlikely to match the functional organization of the brain, and a given brain region likely serves multiple functional roles, which can differ depending on the current environmental context and the networks in which the brain region is currently participating. However, we are somewhat more optimistic that the field can make significant headway in uncovering the structure of operations underlying human thought and behavior. This optimism is fueled in large part by more than a decade of cognitive neuroscience research that has honed techniques for multivariate pattern analysis of functional neuroimaging data. After Phrenology briefly alludes to this approach (Kriegeskorte et al. Reference Kriegeskorte, Mur and Bandettini2008a), but we believe that it is of particular relevance to Anderson's proposed framework and to his suggested ways forward for the field, and hence, merits further emphasis.
We now know that the brain encodes many kinds of information (e.g., memories, tastes, reach trajectories) in neuronal population codes, rather than in the magnitude of single units' responses (Georgopoulos et al. Reference Georgopoulos, Kettner and Schwartz1988; Lin et al. Reference Lin, Osan and Tsien2006; Pouget et al. Reference Pouget, Dayan and Zemel2000; Simon et al. Reference Simon, de Araujo, Gutierrez and Nicolelis2006). Fortunately, much of the information contained in neuronal population codes can be gleaned from the indirect and spatially coarse measurements of brain activity that are currently available to researchers studying human brain function (e.g., functional magnetic resonance imaging, fMRI; Kriegeskorte et al. Reference Kriegeskorte, Mur, Ruff, Kiani, Bodurka, Esteky, Tanaka and Bandettini2008b).
However, much of the information present in neural population codes is lost when analyzing the magnitude of responses of a given functional unit in isolation. For example, one of the first studies that analyzed fMRI data in terms of population response vectors demonstrated that brain regions that are typically characterized in terms of the categories of stimuli eliciting maximal responses (e.g., the fusiform face area) also differentiate between stimulus categories that evoke submaximal responses, such as chairs, shoes, cats, and houses (Haxby et al. Reference Haxby, Gobbini, Furey, Ishai, Schouten and Pietrini2001). This information was carried in unique and consistent topographic arrangements of responses for each stimulus category, and such information had previously been missed by analyzing data in terms of response magnitude only. Hence, analyzing neural responses only in terms of overall response magnitude can lead to an exaggeratedly modular conceptualization of brain function, as it can underestimate the number of brain regions involved in processing a stimulus category, as well as the number of stimuli a given brain region is involved in processing. Therefore, the brain may be even less modular than the functional fingerprinting method that Anderson employs would suggest, given that this method appears to only take into account response magnitude.
Analyzing data in terms of population response vectors, rather than overall response magnitude, can also provide insight into how information is organized within brain regions. To this end, researchers have applied both supervised and unsupervised machine learning techniques when analyzing the distributed topographies of activity within brain regions in order to characterize the distinctions that those brain regions make about various classes of stimuli. For example, by characterizing the similarity structure of population response vectors to visual stimuli within brain regions of the ventral visual pathway, researchers have gained insight into the distinctions that brain areas within that pathway make about sets of stimuli, and hence, the operations that they may contribute to visual processing (Connolly et al. Reference Connolly, Gobbini, Haxby, Kriegeskorte and Kreiman2012). Studies using unsupervised learning techniques (e.g., clustering, multidimensional scaling) have the notable benefit of being wholly data-driven in the structure that is revealed. Hence, such approaches will be useful in following Anderson's call for a science focused on illuminating the psychological factors that best account for observed differences in neural activity, rather than on attempts to map psychological constructs onto the brain.
Considering the information content of multivoxel response patterns, rather than focusing solely on response magnitude, also allows researchers to probe the functional significance of overlap in the brain regions that are recruited to process diverse domains of contents. Univariate, subtraction-based approaches to analyzing brain activity are relatively well suited to asking questions such as, “Is this brain area involved in both Task X and Task Y?” and “Is this brain area recruited ‘more’ for Task X than Task Y?” In contrast, analyzing neural population responses is an approach that is relatively better suited to uncovering how a brain region might contribute to various tasks. Functional overlap can reflect the recruitment of shared or distinct mechanisms, and these possibilities can sometimes be disentangled by studying neural population response vectors. For example, although univariate, magnitude-based analyses of fMRI data suggested that most voxels in a region of lateral occipital-temporal cortex responded to both static pictures of bodies and movement displays, subsequent analysis of the population responses in that region demonstrated that the response patterns to these two sets of stimuli, though spatially overlapping, were unrelated (Peelen & Downing Reference Peelen and Downing2007). Observations such as these are consistent with Anderson's suggestion that a brain region's computational role can differ depending on the circumstances surrounding its recruitment (e.g., the processing demands of a stimulus, the other brain regions with which it is interacting). Hence, analyzing population response vectors can help determine whether spatially overlapping functional activity for two different tasks reflects the recruitment of the same or dissociable underlying mechanisms.
The analysis of population response vectors can also reveal when functional overlap reflects the recruitment of a common underlying neural operation across multiple experimental contexts. For example, a machine learning classifier trained to distinguish between population response vectors within the superior parietal lobule (SPL) associated with leftward and rightward eye movements can correctly distinguish between population response vectors in the SPL corresponding to mental subtraction and mental arithmetic (Knops et al. Reference Knops, Thirion, Hubbard, Michel and Dehaene2009). This suggests that the SPL contributes shared operations when shifting spatial attention within the external environment (during saccades) and within internal mental representations (when performing arithmetic). This likely reflects a case of neural reuse through which functional architecture with a preexisting role in shifting attention in the external environment is exploited in order to perform analogous operations on mental representations, and hence, to perform the culturally learned skill of arithmetic.
We recently used classification and similarity structure-based analyses of neural population response vectors to test for a common neural encoding of egocentric distance in spatial, temporal, and social frames of reference (Parkinson et al. Reference Parkinson, Liu and Wheatley2014). Cross-domain decoding analyses, as well as analyses of the similarity structure of population responses, revealed that areas such as the right anterior inferior parietal lobule, which has a long-established role in representing physical space, organize information not based on distance category (e.g., “Is this a social or temporal distance?”), but based on distance from the self (i.e., “Is this relatively close to or far away from me?”). These results are consistent with Anderson's suggestion that the brain is largely concerned with computing behaviorally relevant information, such as distance from oneself (and hence, relevance for action). Further, these results are consistent with the possibility that neural mechanisms for encoding information about physical space may have been reused in order to operate on increasingly abstract contents, such as temporal and social relationships (Parkinson & Wheatley Reference Parkinson and Wheatley2013; Reference Parkinson and Wheatley2015).
Understanding how the brain gives rise to a given facet of cognition or behavior requires not only identifying the neural substrates involved, but also uncovering the kinds of operations performed by those substrates. We are optimistic that advances in neuroimaging methods, such as the analysis of neural population response vectors, will be useful in moving toward a deeper and more veridical understanding of functional brain organization and of the flexible and adaptive computations that underlie everyday thought and behavior. In so doing, cognitive neuroscience has the potential to elucidate the particular functional capacities that may be reused over the course of evolution and development to solve novel problems.