Cognitive models make assumptions about the types of functions that mediate a cognitive task. In many cases these models do not have any relationship to the functional neuroanatomy of the brain. For example, a well-known cognitive model of reading (Coltheart et al. 2001) includes, among others, modules for visual analysis and for grapheme-phoneme conversion, although no attempt was made to link activity in these modules to fMI or PET data, or even to link them to specific brain regions based on lesion studies. Recently, however, such efforts have been made by imposing additional assumptions relating each cognitive function to specified brain regions. An early attempt at combining a cognitive model with fMRI data can be found in the work of Just et al. (1999), who used a computational (production) model of sentence comprehension called 4CAPS to explain how fMRI activation levels varied as a function of sentence complexity in three brain areas (Broca, Wernicke and dorsolateral prefrontal cortex). In their computational model, Just et al. proposed that resource utilization in a given unit of time in each component of the system corresponds to the amount of activation observed with the neuroimaging measure in the corresponding component during that time interval. Good agreement between the experimental number of activated voxels in Broca and Wernicke's areas and the predictions of their model for three types of sentences of different complexity were found.
A recent study by Anderson et al. (2003) represents another example of this type of modeling. They examined symbol manipulation tasks using a model called ACT-R, which contains a number of buffers. Somewhat different from the assumption used by Just et al. (1999), Anderson et al. (2003) proposed that the fMRI response in a brain area represents the integrated duration of time that a buffer is active. They showed that calculated fMRI activity in one buffer of the model (the imaginal buffer, which tracks changes in problem representation) predicted the fMRI response of a left parietal region, activity in a second buffer (the retrieval buffer) predicted activity in a left prefrontal region, and activity in a third buffer of the model (the manual buffer) was related to fMRI activity in a motor region. In a second study (Anderson et al. 2005), they extended the model to a more complex task, the Tower of Hanoi, and were able to explain latency data in move generation and the fMRI responses in the three aforementioned regions.
One difference between the approaches of Just and Anderson is that Anderson and colleagues (Anderson et al. 2003, 2005) assume that each module of their model corresponds to a given brain region. Just et al. (1999) also ascribe specific cognitive functions to different brain regions, although more than one cognitive function can occur in a region and, conversely, they also assume that a given cognitive specialization may occur in more than one area, albeit with different degrees of efficiency.
Another example of this type of modeling involves an investigation of the role of the anterior cingulate in cognitive control. Brown and Braver (2005) used a cognitive computational model that embodied the hypothesis that the response of the anterior cingulate to a given task condition is proportional to the perceived likelihood of an error in that condition. Simulations performed using the model with a modified stop-signal task resulted in a pattern of behavioral performance that fitted human data. Moreover, the pattern of cin-gulate activity in the model across task conditions (as indexed by the neural firing rate) was qualitatively similar to fMRI activity in the anterior cingulate obtained in an event-related experimental study. They also constructed a second model that viewed the anterior cingulate as detecting conflict between incompatible response properties. They found that they could also fit this model to human behavioral data, but that the pattern of simulated activity of the anterior cingulate across conditions did not match the pattern of the fMRI data. The Brown-Braver study provides an interesting example of how functional brain imaging data can be used in conjunction with cognitive modeling. Namely, different and competing cognitive models may equally well fit the performance data that they typically aim to explain. By comparing the ability of competing models to also match functional neuroimaging data, one is able to select one of the competing models over the other.
The examples we have presented illustrate some of the limitations to this "top-down" approach. One important limitation is that each study we presented employed a different relationship between the activity of a module and fMRI activity. For Just et al. (1999), the fMRI signal was indexed by the rate of resource utilization. This is probably similar to the measure used by Brown and Braver (2005) (i.e., the neural firing rate). For Anderson and colleagues (2003), the fMRI signal is proportional to the integrated duration of time that a buffer is active. Note that although each of these relationships between model activity and fMRI signal is plausible, and all are probably similar, there is no experimental way to verify the relation between a cognitive model component and a neural/hemodynamic variable. As we shall see in the next section, one advantage of a bottom-up approach is that it allows one to relate neural activity to the hemodynamic signal in a testable way. That is, assumptions about the relationship between neural activity and fMRI activity are also expressed in biological, not cognitive or functional, terms.
On the other hand, a big advantage, at least at the present time, for employing cognitive models in conjunction with functional neuroimaging data is that only relatively low-level cognitive functions can be readily addressed in neural terms, since non-human models of high-level human cognitive function (e.g., language) do not exist. The use of cognitive models with fMRI data allows one to deal with very high-level cognitive phenomena (e.g., sentence processing, symbol manipulation), but gives little indication as to how such functions are implemented at a neural level. For the near future, this approach will be useful and will provide interesting insights, and the cognitive modelimaging combination can, when successful, generate a set of target functions at which neurally based large-scale modeling can aim. It is to these bottom-up, biologically based models that we now turn.
Was this article helpful?
Learning About Letting Go, Moving On Can Have Amazing Benefits For Your Life And Success! Don't be held back by the past - face your guilt and fears and move on! Letting go is merely arriving at a decision, no more allowing something from the past tense to influence your life today or to cut down your inner sense of peace and welfare.