Developing Intelligence

When I started this blog back in ’06, new hypotheses were appearing on a possible functional architecture of the lateral prefrontal cortex – a recently-evolved brain area implicated in high-level cognitive functions like planning, analogical reasoning, and cognitive control. Since then, these hypotheses have been refined, and the results replicated numerous times.

Today, it’s essentially incontrovertible that the prefrontal cortex is parcellated into a functional hierarchy in which more anterior areas influence processing in more posterior areas according to the more abstract information represented anteriorly. (Though definitions of “abstract” vary somewhat). Regardless, carefully designed tasks involving multiple levels of complexity – much like a classic decision tree – show that the stimuli pertaining to higher-level nodes in these decision trees reliably activate more anterior regions of the lateral prefrontal cortex.

But as a whole, higher cognitive functions require not only the selective processing of information at a particular level in these decision trees. They also require that the selection process itself be guided by goals.

Systems-level cognitive modeling, typically using neural network models, often consolidates this representation of goals and and the selection processes into single, unitary entity: goal relevant information is selected because the goals themselves are activating goal-relevant information; that additional activation helps that relevant information to “beat out” competing goal-irrelevant information for processing. In this way, selection unfolds naturally from goal representations… at least in the large scale functioning of these models.

But at a lower level, this “unitary process” of goal-based selection could be seen to rely on distinct mechanisms: activation of goals (free energy), and competitive inhibition of goal-irrelevant information (entropy). While a single simulated area in a neural network model can accomplish both, this is not necessarily the way the brain is divvied up (as modeling-skeptics are found to point out). Indeed, a new Nature Neuroscience article from Kouneiher, Charron & Koechlin suggests that the mechanisms supporting this “unitary process” could actually involve a fairly widely distributed brain network, with profound implications for current work on cognitive modeling.


To demonstrate this, Kouneiher et al gave subjects a classic “decision tree” task: subjects saw letters: variously-colored upper/lower-case vowels/consonants. The task involved four levels in a decision tree:

  • level 0 – “default” trials where a single response is required (if the letter is black)
  • level 1 – “baseline” trials where one of two buttons must be pressed depending on letter color
  • level 2 – “context” trials where one of two buttons must be pressed depending on letter color and whether the letter vowel/consonant vs. upper/lower case
  • level 3 – “episodic” trials where one of two buttons must be pressed depending on everything in level 2, but also a preceding instruction cue.

In case you feel like reading the nitty gritty, here it is: just skip the blocked text to get the big picture.

Because level 0 trials were interspersed throughout all other levels, which are each administered in “blocks,” the authors were able to analyze neural activity that is sustained throughout blocks of levels 1, 2 and 3 or is more transiently related to trials that require processing at level 1, 2 or 3.

Brain activity and behavior in this task unfold in apparent lockstep: left inferior frontal gyrus (lIFG; BA44) shows a progressive increase in activation from level 1 to level 2 and then to level 3 that is transient in nature, whereas an area just anterior to it – the left middle frontal gyrus (lMFG BA46/9) – shows a sustained increase across all trials in level 3, relative to levels 1&2. Reaction times show the same pattern as the more posterior area lIFG, consistent with the idea that lIFG is closer on the processing chain to the production of responses than lMFG. So far, all of these findings lie on the surface (lateral) of the brain. They show that the more anterior area (MFG) is representing things in a more abstract and sustained fashion at the highest node of the decision tree, whereas the more posterior area is participating in more transient processing at lower levels of the decision tree.

Additional activations were observed deeper in the brain, along the medial surface – but only when the authors manipulated motivation. In particular, “bonus” trials occurred on 50% of trials – in these trials, subjects had the opportunity to win 1.05 or 3 euros (low vs. high incentives, respectively) for each correct trial, relative to the 1 euro they would have won otherwise.

High-incentive “bonus” trials lead to a transient increase of activation in a relatively posterior area called the pre-supplementary motor area (pre-SMA), as well as the bilateral IFG; more anterior regions, including the dorsal anterior cingulate (dACC) and bilateral MFG showed a sustained increase in activation on blocks that contained high-incentive trials. Again, reaction times proceeded in lockstep: longer reaction times were observed precisely in the same cases that the more posterior IFG region showed more activity. So, bringing motivation into the picture activates the same lateral areas but a corresponding network of more medial areas as well, and these effects are additive, not interactive, with the “control effects” (i.e., levels 1-3) described above.

This picture was confirmed with analyses of how neural activity reliably covaries across areas as a function of both motivation and control (“effective connectivity” analysis, enabled by structural equation modeling). In particular, during level 3 blocks, the anterior lMFG region involved in sustained level 3 processing showed more sustained effective connectivity with the posterior lIFG region involved in transient level 1 & 2 processing. Likewise, the pre-SMA area involved in transient motivation effects showed more transient effective connectivity to the posterior right IFG area when incentives were high. Finally, dACC was more effectively connected with the bilateral MFG, and the left side of the latter with the lIFG, in a sustained fashion, in high incentive blocks.

The results seem to support the following rather clear picture:

1) lateral PFC regions are the interface between control and motivation: posterior areas are modulated by both control and motivation in a transient fashion, whereas more anterior ones are modulated by both control and motivation in a sustained fashion
2) medial PFC regions are involved in “energization” according to motivation: they are only more strongly recruited when incentives are high relative to low. Nonetheless, they follow the same anterior/posterior gradient, with more anterior areas showing a sustained effect of motivation and more posterior ones showing a transient effect of motivation
3) the effective connectivity of these regions is consistent with the idea that connectivity exists along two dimensions simultaneously: anterior regions bias more posterior regions, and medial regions bias the lateral ones.
4) posterior lateral regions seem most tightly coupled to reaction times, as though these areas are carrying out selection of goal-relevant information that is relatively closer than other areas to the generation of responses, along the processing hierarchy.

The authors suggest that the medial/lateral dichotomy of might be understood with respect to the concepts of free energy and entropy: hemodynamic activity in lateral cortex reflect both the overall level of excitation in the system (free energy) but also the extent to which that excitation is concentrated on particular options (entropy). Both types of constructs require blood flow, as picked up via fMRI, but might affect blood flow relatively independently. Generally speaking, this message is compatible with new optogenetic work, with revised neuropsychological taxonomies of the prefrontal cortex, and with systems-level theories of how medial and lateral PFC may interact.

On the other hand, there are important ways in which this work points the way towards important follow-up studies:

A) Modeling implications. Existing computational models might need to incorporate explicit representations of reward magnitude in order to capture medial/lateral interactions, if that is truly driving the medial effects; on the other hand, medial areas might be responding to the conflict that arises from having more activation in lateral areas, which would be less focused on the appropriate representations given the “glut” of activation. This kind of dynamic would be fully consistent with existing models of medial circuitry, as well as the “expected reward”-based modulation of subcortical areas that are known to be interconnected with lateral prefrontal cortex. Although this proposal may seem to conflict with the directionality of connectivity described in the paper, directionality in the kind of structural equation modeling done here is basically invertible (other methods, such as Granger causality, can enable stronger inferences about directionality).
B) Transient vs Sustained effects. Level 3 & levels <3 are confounded in the sense that stimuli pertaining to level 3 were not presented simultaneously with the stimulus (they were preceding instructional cues), so the sustained vs. transient dichotomy observed in anterior vs posterior lateral regions could have been an effect of the design, and not the way these areas intrinsically function;
C) Discreteness vs. gradedness in function & localization. The distribution of function across these areas is very likely more graded, because the areas that “light up” are separated by rather large swaths of cortex. What do those areas do? It’s possible that the “decision tree” structure the authors used is not lighting up intermediate areas of cortex because the brain operates as a more graded or probabilistic device, but appears to correspond to discrete nodes in a decision tree only when that is explicitly designed into the subject’s task. Even then, the parcellation of function might not be so clear: some of the higher-order contrasts (e.g., using ROI as a factor) were not reported, so the apparently “neat” parcellation of function is in many cases relying on null effects, which is risky business in the statistically low-powered endeavor that is fMRI.

All that said, it should be clear that this is an excellent paper, and breaks much new ground in extending our current understanding of anterior/posterior gradients in lateral prefrontal cortex to the medial surface.