The Adaptable Mind: What Neuroplasticity and Neural Reuse Tell Us about Language and Cognition - John Zerilli 2021
Further Evidence of Neural Reuse
Neural Reuse and Recycling
3.3.1 Computational modeling
A number of large-scale computational models of the brain are currently being developed in the hope of understanding the activity of a million neurons or more. At their most sophisticated, they leave behind the biologically unrealistic neural network models of the past and enter the domain of real brain simulation, neurorobotics, and neuromorphic computing. As the computational analogues of real neural networks, they are beginning to offer fresh insights into the brain’s dynamic response properties. The primary advantage of brain simulation is that, “unlike the empirical brain, the model’s internal workings are completely known and the model’s structure can be modified in order to explore how its activity changes” (Sporns 2015, p. 97). One example of brain simulation that is especially relevant in the present context is SPAUN (Semantic Pointer Architecture Unified Network). Spaun has a single eye through which it receives digital images as input, and a moving arm through which it provides behavioral output (Eliasmith 2015). What is interesting is that its 2.5 million “neurons” are organized to simulate about twenty of the roughly one thousand functionally distinct areas that neuroscientists typically identify in the brain (e.g., separate neurons for frontal cortex, basal ganglia, occipital cortex, etc.). One feature of Spaun that supports the theory of reuse comes as a result of this unique “modular” architecture:
One key contribution of Spaun relative to many competing architectures is that Spaun can perform a variety of different behaviours, much like an actual brain. For example, Spaun can use its visual system to recognize numbers that it then organizes into a list and then stores in working memory. It can then later recall this list and draw the numbers, in order, using its arm. Furthermore, Spaun can use this same visual system to parse more complex input. . . . To do so, it uses the same memory system, but in a slightly different way. As well, it uses other brain areas that it didn’t use in the list recall task. That is, Spaun can deploy the same brain areas in different ways depending on what task it needs to perform. (Eliasmith 2015, p. 132, my emphasis)
Spaun’s differentiated circuits manifest the very same interactive dynamics that reuse posits of real brains: “different, specialized brain areas are coordinated in a task-specific—that is, flexible—way to meet a challenge presented by the environment” (Eliasmith 2015, p. 132). This behavioral flexibility marks a distinctive sense in which neural reuse is a form of plasticity: the ability to switch effortlessly from task to task (reading an email, playing a piano, chasing a dog) using the same brain areas in different ways and with little or no delay in shuffling between them. This kind of plasticity serves to set biological intelligence apart from most contemporary artificial intelligence, and indeed explains why Spaun is “atypical of the field” overall (Eliasmith 2015, p. 134). Most machines are good at doing one specific thing (playing chess, solving mathematical equations, driving a car, etc.). Spaun is unique both in the variety of tasks it can perform and its capacity to learn new behaviors independently “while preserving abilities it already has” (Eliasmith 2015, p. 134). Spaun may be one of the first tentative steps towards showing that a domain-general learning system can work.5
3.3.2 Biobehavioral evidence
Casasanto and Dijkstra (2010) report an interesting association between autobiographical memory and motor control. The task involved shifting marbles upward or downward from one container to another while relating memories having either positive or negative valence. Subjects were asked to retell, for example, a negative memory, followed by another negative memory, then perhaps a positive memory, while simultaneously moving marbles from one container to another in a given direction. It was found that subjects retrieved more memories, and moved the marbles more quickly, when the direction of movement aligned with the valance of the memory; i.e., when the upward movement coincided with positive memories, and the downward movement with negative memories. Even when subjects were not asked to relate memories that were specifically positive or negative, but just to relate memories as they came, they were more likely to retrieve memories whose valence correlated with the direction of movement. The directedness of the movements involved suggests an important association between memory, movement, and spatial orientation likely to be reflected in shared neural circuitry.
The reuse of spatio-visual circuits for numerical cognition is illustrated by the spatial-numerical association of response codes (SNARC) effect. Here are just two examples of the SNARC effect (Dehaene et al. 1993): (i) when asked whether a number is even or odd, subjects respond more quickly with large numbers displayed to their right, or small numbers to their left; (ii) when presented with a line of neutral symbols (e.g., XXXXX) subjects fare better at correctly indicating the midpoint than when presented with small numbers (e.g., 22222), in which case there is a bias to the left, or large numbers (e.g., 99999), where the bias is to the right. It appears that in these cases a mental number line running from left to right is being navigated with the help of spatial orientation circuits (Hubbard et al. 2005).
3.3.3 Final thoughts
Despite the extensive and compelling nature of the evidence supporting reuse, not to mention again the powerful evolutionary considerations in its favor, the case has not yet managed to convince everyone. Neural activation and imaging evidence on its own is ambiguous, the skeptics point out, being consistent with multiple neighboring sets of neurons that only appear to be reused as a result of the coarse spatial resolution of contemporary imaging technologies (Anderson 2010, pp. 298—299; 2014, p. 29). Furthermore,
. . . because neural activation may spread around the brain network, this can lead to false positives: regions that are activated only as a side effect of their connectivity and not because they are making a functional contribution to the task under investigation. (Anderson 2014, p. 29)
This “spreading activation” is what Colin Klein (2010, p. 280) has dubbed a “potential confound,” and such worries cannot be lightly dismissed. On the contrary, misgivings about the use of neuroimaging evidence are precisely why converging biobehavioral evidence (of the kind just cited) will be critical in a debate like this. The more biobehavioral evidence of functional and semantic inheritance between task domains, the greater our confidence that the very same neural structures are involved (Anderson 2014, p. 30). The limitations of neuroimaging technology can thus be overcome by adopting supplementary research paradigms. An interference paradigm, for example, asks participants to process two stimuli at the same time. If the processing of these stimuli draws on shared neural resources, one would expect this to be reflected in performance: perhaps a slower reaction time as compared to performance on similar tasks that do not make processing demands on the same neural elements. Thus, on the assumption that the fusiform face area would respond to objects of expertise as well as faces, Gauthier et al. (2003) predicted that face processing in car experts would be impeded by the presentation of cars at the same time—and this is just what they found. Here we have evidence of reuse coming from a research paradigm outside neuroimaging, and none the worse for that. Later on I cite yet a further type of evidence, this time from single-neuron studies, demonstrating that, while concerns over poor spatial resolution and spreading activation may be legitimate, they can hardly be decisive (see the discussion of “strong context effects” in § 5.1). The simple fact of the matter is that evidence in support of reuse comes from many quarters, including from disciplines—such as neurology and neuropsychology—in which distributed parallel processing models had been proposed well before the advent of neuroimaging technology (Bressler 1995, p. 289). Still, even assuming that the general thrust of this hypothesis is correct (as I for one do), it is not immediately obvious that anatomical modularity is dead, for perhaps it is only in respect of its functional scope that it stands in need of revision. Moreover, as I suggested earlier in this chapter, outstanding questions concerning the existence of a dedicated language module remain as acute as ever, and these are tied in part to an extensive dissociation literature as well as to the concerns over spatial resolution and neuroimaging I just raised. I turn now to consider these issues, and begin by inquiring into just what the implications of reuse and neuroplasticity might be for the modularity of mind.