How does the brain deal with the need to pursue multiple goals simultaneously, particularly if they are associated with different reward values?
One idea, perhaps far-fetched, is that the brain might divvy up responsibility for tracking these goals & rewards: for example, the left hemisphere might respond to a primary goal, and the right hemisphere to secondary goals. To me, this kind of simple division of labor smells like lots of ridiculous and outdated hemispheric asymmetry theories. That’s why I’m dumbfounded that new evidence provides startling support for it, as reported in a 2010 Science paper by Charron & Koechlin.
Charron & Koechlin gave subjects a relatively simple task (warning: nothing else in this post will be simple). Subjects saw letters from the word “tablet” presented one at a time, in sequence, and had to respond with one button if any sequence of two letters appeared in the same sequence as in the word “tablet”, pressing another button if not. For example, the key-presses for the sequence “TABELET” would be “1112211″ where “2″ is the button press for no, and 1 is the button press for yes. Up to 13 letters were presented in a single sequence, with all letters in either green or red: green indicated a large reward was available for that sequence (1 euro) whereas letters presented in red indicated only a small reward was available (.04 euro). So far, this is really only a single task – do the letters match “tablet”? – with a reward manipulation tacked on top.
For the second task, Charron & Koechlin surrounded some of the letters with triangles; this could mean one of two things. If the triangle was downward-pointing, subjects were required to “start fresh” on that trial (i.e., indicate whether the letter was a “T” and continue from there) until the triangle disappeared, at which point they would “start fresh” again in deciding if the next letter was a “T”. This requires a switch away from letters previously represented in working memory, but subjects are still fundamentally completing the same single task as before.
In contrast, upward-pointing triangles required not only switching but also some amount of dual-tasking (or “branching” in Koechlin’s terms). In that case, subjects still had to “start fresh” by responding to subsequent letters appearing within triangles, but when the triangle disappeared they would have to pick up where they had left off – returning to the previous task as though no triangles had appeared at all. This requires a switch of focus away from letters represented in working memory, but also that those represented letters remain in memory – i.e., branching of goals and memory – so that subjects can accurately return to the previous task when the triangle disappears. Thus, the branching condition requires that information pertaining to two goals be simultaneously maintained.
Surprisingly, subjects performed well at these tasks, with an error rate below 5% even on the “secondary task” trials (i.e., those where triangles were present, regardless of whether those triangles required switching or branching). Moreover, error rates and reaction times decreased for secondary task trials when they were associated with a large reward, again regardless of whether the triangles required switching or branching, demonstrating that the reward manipulation worked.
What’s interesting is where switching and branching differ: first of all, reaction time and error rates were higher overall for the branching condition than the switching condition, consistent with the idea that branching is harder. But when the rewards for the primary task – i.e., before the triangles – were high, reaction times and error rates on the subsequent secondary trials increased in the branching condition, but decreased in the switching condition. Thus, subjects seemed to sacrifice performance of the secondary task to faithfully maintain the preceding stimuli without triangles, so as to receive that larger reward when they had to return to that previous task.
Of course, the meat of the paper is how switching and branching neurally differ. First, the authors saw a lot that’s consistent with previous work [see yesterday's post for some background]:
1) putatively “motivational” areas like dorsal anterior cingulate (dACC) and pre-supplementary motor area (pre-SMA) were more strongly activated when secondary rewards were high, along with numerous others areas (including lateral prefrontal cortex);
2) only on branching trials was a more anterior and lateral region recruited – frontopolar PFC – generally consistent with the idea that some “decision tree” for the task required processing at a higher node (e.g., if triangle appears again -> respond conditional on previous letter; if triangle disappears -> respond conditional on some more distantly-preceeding letter). This can be contrasted with the decision tree required by the switching condition, in which higher-level processing is irrelevant: (e.g., if triangle appears again or disappears -> respond conditional on previous letter).
But they also found a real gem:
The hemispheres divvy up branched goals: On “branching trials,” the reward for the primary task was tracked by the left dACC, whereas the reward for the secondary task was tracked by the right dACC. Frontopolar PFC didn’t show this same hemispheric dichotomy, and was instead more strongly recruited bilaterally only when both primary and secondary rewards were higher.
And the clincher:
With only two hemispheres, triple-task branching should be impossible.If the two hemispheres divide under true dual-task conditions – i.e., secondary trials in the branching condition – then it shouldn’t be possible to load-balance goals under triple-task conditions, because there’s only two hemispheres! To test this, Charron & Koechlin gave a separate set of subjects a triple-task condition and found patterns consistent with the idea that triple-task conditions effectively reduce to a dual-task scenario (as though only two goals can be simultaneously tracked) where the pending third task essentially gets dropped from the queue. Apparently, we need a third hemisphere.
In summary, Charron & Koechlin find that when only a single task needs to be performed, even if that’s conditional on a higher-order node in a decision tree, hemispheric division of labor is not observed. But when responses are conditional on that higher-order node and might require consideration of preceding information, medial regions of the left hemisphere represent that “pending” goal, and medial regions of the right hemisphere represent the concurrent goal.
Extraordinary claims require extraordinary evidence, and they have provided it, but as usual it’s worth considering alternatives and the limits to the current work.
A) Because dual-tasking required keeping the preceding task in a pending state, it might be assumed that what drives frontopolar activation is the increased temporal delay required for dual-tasking. However, it seems equally likely that frontopolar activation is driven by the fact that the dual-tasking required consideration of a higher-order conditional in the logical structure of the task (only in dual-tasking was information actually conveyed by the disappearance/repetition of triangles). Thus, as is typical for this literature, we have an interpretational issue in dissociating temporal factors from those involving “logical structure” (for lack of a better term.)
B) The behavioral experiment showing difficulty with triple tasking used a 3-back task to determine whether the increased amount of memory required for successful triple tasking might be enough to explain triple-tasking difficulty. The authors found that subjects do have the working memory to maintain 3 distinct items and compare the current letter with the one occurring 3-letters ago, in terms of its order in the word “tablet” (error rate was only ~10% on this 3-back task). Their conclusion is that working memory can’t explain the disproportionate number of errors subjects make in returning to a stimulus that had been followed by two subgoals, as observed in triple-tasking. But working memory capacity is not clearly best measured in terms of how often an item can be remembered over 2 intermediate items, as opposed to resistance to interference from more intermediate items (as in triple-tasking), or a larger amount of time over which items can be successfully maintained (as in triple tasking), or in terms of its resilience to unexpected shifts in attention (as in triple tasking). All these factors should affect working memory capacity, and all are going to be more pronounced in triple-tasking than the 3-back task. So it’s not clear at all that subjects can’t do triple-tasking simply because they have only 2 hemispheres, as opposed to limited working memory functioning.
C) If one takes this skepticism to an extreme – that hemispheric division of labor is only epiphenomenally related to the number of tasks – then what is driving these hemispheric discrepancies? One idea is that there’s a endogenous/exogenous dichotomy operating in the anterior cingulate, such that left dACC could be resolving conflict arising from the processing of current stimuli in the environment that are worth less than those being maintained, whereas right dACC is resolving conflict arising from the processing of maintained stimuli that really aren’t worth all that much relative to processing the current environmental stimulus. This hemispheric division of labor seems more likely to me than load-balancing of rewards associated with various goals: for example, we know from much other work that the right hemisphere is often implicated in orienting behaviors, whereas the left hemisphere is sometimes more strongly implicated in maintenance operations.
D) Frontopolar PFC may be modulated by simultaneous increase in the reward value of both primary and secondary goals precisely because this is a case where left and right dorsal ACC cannot resolve the conflict arising from attempts to prioritize either goal. In this case, a superordinate area must be brought online in order to mediate between these options. This speculation could apply more broadly to this literature, because it raises the possibility that conflict goal-based conflict, and not the need to traverse some kind of logical decision tree using conditionals, is the operative principle behind recruitment of increasingly anterior regions of the prefrontal cortex. Dissociating these possibilities is an important direction for future work.
One cannot simply dismiss the Charron & Koechlin proposal of goal load-balancing in the prefrontal cortex, given that all of their data is consistent with that proposal. But I suspect the “ultimate truth” about hemispheric division of labor is likely to be more complicated.