A Neural FPGA? Dynamic Task-Relevant Code in Prefrontal Cortex

How does the brain exert flexible control over behavior? One idea is that high-level areas of the brain self-organize representations that lead to reward in a certain task, in a sense by "programming" or "executing" a pattern of activity that controls activity in more posterior and domain-specific regions (i.e., sensory or motor cortex). This portrays prefrontal cortex as a kind of field-programmable gate array, which can be dynamically reconfigured on the basis of dopaminergic reward signals, so as to perform different computations at will.

Concrete evidence for this view is provided by Hasegawa et al. in their recent article on the activity recorded from 310 neurons in the dorsolateral prefrontal cortices of 2 rhesus monkeys while they performed a surprisingly difficult task.

In this task, the authors trained the animals view a centrally-displayed small red or green cue square, followed by a white square at one of 6 locations, followed by a delay, finally followed by two probe squares. One of these two probe squares appeared at the same spot as the white square had, while the other appeared at one of the unused 5 locations.

The monkeys were trained to look at the probe square in the same location previously occupied by the white square, but only if the initial cue square was green; if the initial cue square had been red, they were to look at the probe square that did not occupy the same location as the previous white square.

The recordings of individual neurons showed surprising diversity:

- The majority of neurons were "spatially-tuned"; in other words, they were more responsive to squares in certain locations than in others

- Some neurons were responsive only for the red or green cue squares (which determined the nature of the task); others were responsive only for the two probe squares (which determined the location of the response); some were responsive to both. This shows that many neurons were tuned to the type of goals and behavioral performance required on each trial.

- Among those that fired when the animal had to choose which probe square to look at, 10% of the neurons fired only in response to those probe squares that the monkey was supposed to look at (the "look" neurons) whereas another 10% fired only in response to those probe squares that the monkey was supposed to not look at (the "don't look" neurons). The authors suggest that these "don't look" neurons reflect active suppression of a saccade.

- A further 10% of both the "look" and "don't look" neurons only began to show this selectivity as the task wore on; in other words, they appeared to be changing their function as the task progressed, perhaps reflecting the on-the-fly "dynamic coding" of task rules. Yet other neurons were "purely visual" in the sense that they fired whenever a probe square appeared, regardless of its relevance to the task

- These three types of selectivity were also present among those neurons that fired during the delay period (when no visual stimuli were presented to the animals). 53% of 128 neurons were "look" memory neurons, 19% were "don't look" memory neurons, and a remaining 28% were "pure memory" neurons.

- Some neurons that were initially sensitive only to the presence of visual stimuli became also sensitive during the delay period, as the task progressed. Neurons that became "look neurons" tended to show a burst of activity to the stimuli in the match trials, which decayed over the delay before building up again into a presaccadic "burst" of activity. "Don't Look" neurons tended to have no activity to the visual stimulus but rather slowly increase their activity in the delay period until a saccade was made (incidentally, these same neurons gradually increased their activity in the match task as well).

- Neuronal activity during the delay was generally stronger when the monkeys subsequently made a correct response, suggesting these representations are important for task performance and goal memory. "Don't look" activity predicted performance on tasks beginning with green squares, whereas "look"' activity predicted performance on tasks beginning with red squares.

The authors argue that this cannot reflect mere attention or arousal effects, since the effect of task on neuronal firing was much greater than saccade latency or velocity. They also argue that the "look" and "don't look" neurons were not merely color-sensitive since they didn't respond to white stimuli and since many developed their selectivity only late in the task (color-sensitive neurons typically do not do so).

In their discussion, Hasegawa et al. review previous single-unit recording evidence for suppression of all saccades through tonic neuronal firing, but they make a more specific argument here: the "don't look" neurons appear to represent the active inhibition of viewing that particular square.

This claim is actually of substantial theoretical importance - and it can be rephrased as a programming question: for a program to accomplish a certain action, does it need explicit instructions to cease performing other actions? Obviously, there are many problems with the brain-computer metaphor, but here it conveys the simple fact that programming what not to do is grossly inefficient, and perhaps completely unnecessary.

On the other hand, a program to accomplish this task would have to store the location of the "don't look" square in memory, so it could later compare that with the locations of the two probes squares and make the correct saccade. This is the alternative possibility that I advocate, on the basis of a few additional problems with the idea that "don't look" neurons represent inhibition of specific saccades:

- Maybe "don't look" neurons are not actively inhibiting anything but are merely "descriptive" of the current task goal. In other words, they may merely respond to the conjunction of a green cue square with a following white square in a certain location. In fact, such a representation is required, even if inhibition is not: the monkeys need to remember where the previous item was so they can later compare the locations of the probe stimuli to that memorized location.

- These neurons may actually send additional activity to the neurons spatially selective for other locations, rather than actively and directly inhibiting a specific saccade. Accordingly, "don't look" neurons did not show large spikes in activity to the cue square - instead, their activity built more slowly over the delay period than "look" neurons

- If these neurons actually actively inhibit a certain area, sufficient stimulation of spatially-tuned "don't look" neurons should make saccades to those locations impossible, even during tasks where the monkey's goal was to look in that direction. This was not performed.

In summary, Hasegawa et al. provide fascinating evidence for the dynamic coding model of prefrontal cortex function, in which representations may become self-organized so as to accomplish certain task goals. These "programs" may or may not include instructions to inhibit specific actions - the evidence seems somewhat inconclusive on that point. Nonetheless, it is a powerful demonstration of the computational flexibility of the prefrontal cortex even in monkeys.