How can we enhance perception, learning, memory, and cognitive control? Any answer to this question will require a better understanding of the way they are best enhanced: through cognitive change in early development.
But we can’t stop there. We also want to know more about the neural substrates that enable and reflect these cognitive transformations across development. Some information is provided by developmental neuroimaging, but even that’s not enough, because the real question we have can only be answered via mechanisms (“how”/”why”) – quite different than the “what” “where” and “roughly when” questions addressed by neuroimaging. For “how/why,” we ultimately need a mathematical way of describing cognitive changes and how they unfold in tandem with changes in neural information processing. This, Russ Poldrack argues in his 2010 HBM paper, can come only from a computational integration of cognitive and neural development: something called “computational developmental cognitive neuroscience.”
Here I’ll outline Poldrack’s argument for computational developmental cognitive neuroscience. I’ll elaborate his examples with ways in which computational explorations of cognitive neuroscience already provide strong clues about the elusive “why”/”how” questions we seek to answer.
Poldrack begins with a daunting list of challenges:
1. Neurobiology that could allow strong inferences unfolds too early to underlie what we’re interested in. Poldrack notes that while we know a LOT about developmental neurobiology (timeline of neural migration, synaptogenesis, myelination, etc), most of it pertains to the very earliest years of life (<1.5 years-of-age). At that point, the cognitive changes of most interest have not yet occurred. On the other hand, those neurobiological changes known to occur after 2 years-of-age are few and poorly understood: first, changes in cortical thickness >2 years have largely unknown effects on the hemodynamic response as well as cognition; second, changes in neural coding have barely been identified, much less specifically linked to differences in cognitive processing.
2. In other cases, inference is severely underconstrained. Even when performance differences across ages have been equated, changes in hemodynamic activity could reflect age-related differences in neural information processing OR the same neural processes, just differences in efficiency (e.g., less efficiency in younger subjects leading to greater activation; or less efficiency leading to the same activation in one area but requiring more support from other areas). Moreover, as Poldrack covers later in his article, the correlational nature of neuroimaging means that we can’t tell whether any hemodynamic differences causally underlie or merely reflect differences in neural information processing.
3. Common inferences are descriptive, not explanatory, and often logically incoherent. Developmental decreases in activation are commonly interpreted to reflect increased efficiency, and increased efficiency is interpreted to be reflected in decreases in activation. This circularity obscures a deeper problem: taken to its extreme, the position holds that the most efficient processing involves no activation at all. Disregarding these logical issues, “efficiency” is descriptive rather than explanatory, since we don’t know whether it might arise from developmental changes in neurovascular coupling (for example) or from increasing sparsity of representations (for example).
4. A Distributed to Local Shift. Building on the idea that neural coding becomes increasingly sparse or efficient, some have observed a shift in recruitment from relatively large swaths of cortex to more spatially-delimited foci of activation as children mature. This finding contrasts with work demonstrating that cortical patterning is completed before 5-year-of-age, and that reported changes in cortical patterning after that point take the form of reduced spatial focalization! In addition, Poldrack argues the observed “distributed-to-local shift” is in empirical peril, because maturation is associated with reduced short-range connections and increased long-range connections during this time, which might predict the opposite developmental shift in hemodynamics. In addition, registration errors or large individual differences could smear out activation maps at the group level, giving rise to an apparent “distributed to local” shift that actually has nothing to do with the underlying computations.
So, what’s the way forward? Poldrack suggests computational models allow us to get a handle on the above inferential difficulties. Why computational models?
- The core problem of 1-4 above is how to relate hemodynamic differences to differences in neural computation
- Existing reinforcement learning models have revolutionized the adult literature by dissociating the information processing reflected by hemodynaimc patterns in e.g., ventromedial striatum vs. prefrontal cortex, and various subregions of the anterior cingulate.
- Existing models in the developmental literature have not yet addressed developmental neuroimaging, which provides the best neurobiological data we have during the maturation window of interest
This seems reasonable, but unfortunately the paper ends before saying exactly how computational models will solve the problems highlighted earlier. So let me elaborate:
1) Are there delayed effects of early neurobiological change on later cognitive development? In a computational model I’m about to submit to JoCN, we find that the absolute earliest differences in striatal function reliably predict the absolute latest differences in complex cognitive performance, on an n-back task. If that pattern is taken seriously, it means that well-established neurobiological changes in 1-year-olds may in fact make very specific predictions for cognitive performance occurring in a 5-year-old. Computational models can evaluate this idea in a fairly rigorous way, providing specific and transparent predictions for developmental neuroimaging in the age ranges of interest, at least when a model’s learning curves are taken seriously.
2) Eliminating the performance burden. Performance differences between ages could be an extraordinarily useful thing, if you have a computational model that explains how and why performance differs across ages. If made sufficiently biological, such a model can then turn “the performance burden” into “the performance bonus.” For example, we need not preselect subjects of similar abilities despite different ages (which is weird anyway, since we’re interested in age only insofar as it pertains to ability!), but can account for both features of the data simultaneously in a principled way. Another example: we need not adopt the problematic practice of giving subjects of different ages different tasks, in an attempt to equate performance, if performance can be explicitly and mechanistically accounted-for.
3) Neuroimaging can allow inferences about neural information processing, if neural information processing is the basis of our predictions. Differences in neural efficiency can be simulated by increasing the sparsity of representations using lateral competitive inhibition in a neural network model, and tested with multivariate methods in neuroimaging. Differences in neurovascular coupling might be simulated by changing the relative weight of excitatory and inhibitory neurotransmission, in conjunction with net input to each cell type, in a proxy-hemodynamic measure from the model. These two types of manipulations may make different predictions about the neuroimaging results that should be observed, using two techniques which sometimes lead to divergent predictions (a problem that can be addressed with a biological-modeling method, too). This eliminates logical circularity and points the way towards a mechanistic theory, regardless of which hypothesis is supported.
4) Reconciling maturational change in functional connectivity with the putative distributed to local shift. Decreased short-range and increased long-range functional connectivity may in fact not predict a shift from local-to-distributed processing with age. One could imagine that functional connectivity among disparate regions would become stronger at precisely the time that the intrinsic connectivity of each region becomes more focal. Because each region in this functional network recruits an increasingly more focal pattern (owing to decreased short-range connectivity), the univariate hemodynamic pattern in each area should be of increased focalization, while the correlations in activity among areas should be increasingly coherent. This hypothesis awaits direct computational assessment, but I’m planning to try it.
I think Poldrack has nicely made the case for why computational models should have a prominent role in understanding the neurobiology of cognitive development. In fact, I agree that they must be crucial if we are to ever conclusively link developmental change in information processing with that in hemodynamic measures. Do you?