As described in yesterday’s post, many theories have been proposed on the possible functional organization of prefrontal cortex (PFC). Although it’s clear that this region plays a large role in human intelligence, it is unclear exactly “how” it does so. Nonetheless at least some general conclusions on prefrontal computation can be made.
A reasonably uncontroversial view is that prefrontal cortex maintains over time representations that integrate sensori-motor with current goal and context information, and that this active maintenance biases processing elsewhere in the cognitive system in accord with the represented goals.
According to this view, increasingly anterior regions of PFC tend to represent increasingly abstract or meta-level goals, and subsequently bias more posterior PFC regions that are “downstream.” This perspective was elegantly elaborated by Koechlin, Ody, and Kouneiher in their 2003 Science article on a “cascade model” of prefrontal function.
In the article, Koechlin et al. propose a hierarchical model of lateral prefrontal cortex (lPFC) function, in which lateral premotor regions select motor actions in response to incoming stimuli, which is biased by “context” processing in caudal lPFC, which is itself biased by goal-related “episodic” and temporal processing in rostral lPFC. According to this model, each PFC region is segregated on the basis of temporal structure, with more anterior regions representing information that spans longer temporal episodes; a cascade of neural information processing travels from anterior to posterior PFC, becoming more and more specific to the current action, and less specific to the broader context in which the action is taking place.
To evaluate this model of PFC, the authors conducted two experiments which were then subjected to information theoretic analysis. I’ll do my best to explain these complex tasks and analyses, but feel free to skip the blocked text below if you’re not particularly interested in the methodology.
In the first experiment, subjects had to press buttons in response to certain colored squares while ignoring distractor squares. The authors varied this task in three different ways. The first type of variation was stimulus-related: the authors varied the number of colors (1 or 2) requiring a response in each block, and there was always just one distractor color to which no response was required. The second type of variation was related to the context of the experiment: the authors varied the number of possible responses (also 1 vs 2) that could be given in any trial of a certain block – in other words, the same response could be required for each of 2 target squares, or a different response could be required for each. The third type of variation was related to teh larger episode in which the previous manipulations took place. To this end, “Instruction cues” were provided before each block; these just informed the subjects which task they would have to perform on the next block of trials. For four blocks of trials, the Shannon information value (a measure of how much information is conveyed by something; see here for more on this) of the instruction cues was 0 bits, since the same responses were required to the stimuli regardless of which instructional cue was provided; in other words, the instructional cues provided no information because they did not reduce uncertainty as to how to respond whatsoever). In two of the remaining four blocks, the information value of instructional cues was 1 bit, because having the instructional cue reduced uncertainty by approximately 50% (i.e., 50% of the stimuli required the same responses regardless of the instructional cue). In the final two blocks, the information value of instructional cues was 2 bits.
In the second experiment, subjects had to perform one of two tasks (vowel/consonant judgment or lower/upper case judgment) on letter stimuli depending on the color of the letters. As in the previous task, instructional cues were provided to inform subjects on how to respond to each color. Also like the previous task, certain instructional cues had more Shannon information value than others, due to overlap in the type of stimulus-response mappings required for each stimulus type.
Subjects were scanned with a 3T fMRI machine, and this data was analyzed with the following contrasts. Greater activations in with two possible responses relative to one possible response were considered “stimulus effects,” whereas greater activations in blocks with two possible tasks relative to those with one possible task were considered “context effects.” Finally, “episode effects” were defined as the change in activation resulting from increases in Shannon information value.
The authors conducted standard ANOVAs on this data with hemisphere, experiment number, number of possible responses and information value of the cues as factors. Effective connectivity was also determined through structual equation modeling as follows. Activations in one PFC region were correlated with changes in all other PFC regions; these correlations were then examined for variation as a function of number of possible responses, instructional cues, or number of tasks.
The results showed that the most anterior regions seemed responsive to information with the longest temporal impact (or requiring longer maintenance), whereas more posterior regions were also responsive to information with the the most current impact (or requiring shorter maintenance). That is, activity in bilateral premotor regions (middle frontal gyrus) was modulated by the stimulus effect (number of possible responses) whereas activity in caudal PFC (inferior frontal gyrus) was not modulated by this effect. Instead, this more anterior region was modulated by the context effect (number of tasks to be performed). Neither of these factors affected activity in rostral PFC (inferior/middle frontal gyrus) which was instead modulated by the episode effect (amount of information provided by each instructional cue).
This pattern of “temporal nesting” strongely implicates exactly the “cascade model” that the authors propose as an organizing principle of prefrontal cortex funciton. Furthermore, this pattern was additive across regions, such that the most posterior region showed some sensitivity to all three factors, the middle region showed sensitivity only to two of these, and the most anterior region was sensitive to only one.
Effective connectivity analyses showed results compatible with this conclusion as well, such that increases in episodic control were associated with increased correlations in activity between rostral PFC, caudal PFC, and premotor regions. In other words, episodic control involved the full neural cascade from the most frontal to the most posterior regions of PFC. Likewise, changes in context (number of tasks performed) was associated with only some of the PFC cascade, in that caudal PFC activity correlated more tightly with premotor PFC. In contrast, changes in stimulus control (how many stimuli might appear) was not associated with any increase in correlations between these regions.
Based on this compelling evidence, Koechlin et al. conclude that prefrontal computations support both cognitive control and temporal organization of behavior, by way of a nested, hierarchical processing structure. Specifically, posterior (premotor) regions of PFC maintain sensory information, while slightly more frontal regions (caudal PFC) maintain task information and while the most frontal regions (rostral PFC) maintain episodic information.
This account has numerous strengths in comparison to the models of PFC function reviewed yesterday. It is compatible with a wide variety of neuroimaging and behavioral data and coherently integrates many of the features ascribed to prefrontal processing by other theories. Although it’s arguable that the temporal distinctions made by Koechlin et al. are somewhat fuzzy (i.e., are the “instructional cues” for a block really more ‘episodic’ than number of responses required for that block?), the conclusions are bolstered by the “cleanliness” with which this model predicted the results. Such cleanliness is somewhat uncharacteristic of most prefrontal neuroimaging research, in which it is usually difficult to predict which subregions of PFC will be activated by specific task manipulations.
Of course, much work remains to be done. For example, it’s unclear how the “temporal cascade” model of prefrontal cortex might relate to more traditional accounts of prefrontal/executive processing (for example, theories that invoke “directed inhibition”.)
Several more specific questions come to mind. How might the known neurobiology of prefrontal dopamine levels relate to this temporal cascade model? Do patients with prefrontal damage actually manifest the profile of “temporally delimited” deficits that would be predicted by this cascade model? Finally, this theory is limited to lateral regions of PFC; might the same prinicples apply to medial regions of PFC, thought to be involved in emotional processing?
The Anterior Frontier: Prefrontal Cortex
Eyes, Window to the Soul – And to Dopamine Levels?
rTMS of dlPFC Dissociates Maintenance and Manipulation Processes
Functionally Dissociating Right and Left dlPFC
Models of Active Maintenance as Oscillation
Models of Dopamine in Prefrontal Cortex
Localizing Executive Functions in Prefrontal Cortex