We recorded from 91 vPFC neurons while the monkeys participated in the same-different task (). For 53 of these 91 neurons, we collected blocks of data in which both bad
were the reference stimulus. In the other 38 neurons, we only collected blocks of data in which either bad
(22 neurons) or dad
was the reference stimulus (16 neurons). 67 of the 91 neurons were classified as “auditory” [16
]; these neurons had reliably different firing rates during the 500-msec period that began with test-stimulus onset than during the 500-msec period that occurred prior to test-stimulus onset (t-test
, p < 0.05).
Which aspects of the task could this vPFC neuron be coding? Since the stimulus-presentation dynamics in our same-different task are similar to that used in oddball tasks and stimulus-specific adaptation [2
], stronger “pop-out” vPFC responses might reflect the automatic detection [2
] of uncommon test stimuli. Therefore, stronger responses could reflect test stimuli that are acoustically distinct from the reference stimulus (i.e., the 0% – 80% morphs). However, since this neuron responds weakly to several of the test stimuli (see ), its response pattern does not reflect the presence of acoustically distinct test stimuli.
Relationship between the monkeys’ choices and neural activity
Does vPFC activity reflect what the monkeys should
choose or does it reflect the monkeys’ actual
choices? To gain insight into whether neural activity reflects what the monkey should choose, a neurometric analysis [20
] (see Supplemental Material
) was conducted. This analysis tests whether an ideal observer can use vPFC activity to predict the differences between test stimuli and whether this activity co-varies with the monkeys’ behavioral reports. However, as seen in the Supplemental Material
, the results of the neurometric analysis indicate that vPFC activity is not a good predictor of the test stimulus and, hence, is not a good predictor of what the monkey should choose.
To test the hypothesis that vPFC activity reflects the monkeys’ actual choices, we calculated the choice probability (CP) [22
]. On a neuron-by-neuron-basis and using both
successful and error trials, we first formed two distributions. One distribution contained the test-stimulus-period firing rates from trials when the monkey reported that the reference and test stimuli were the same. The second distribution contained the firing rates when the monkey reported that the stimuli were different. From these two distributions, a receiver-operating-characteristic curve was generated; the area under this curve is a neuron’s CP [23
]. The CP values from different neurons and from different variations of the analysis were grouped together to form different population distributions of the CP values; to minimize the differences between different neurons’ firing rates, firing rates were normalized with a z-score.
If vPFC activity reflects the automatic detection of acoustically-uncommon test stimuli, neural activity should not be modulated by the monkeys’ choices. Under this hypothesis, the CP should equal 0.5. On the other hand, if vPFC activity reflects the monkeys’ choices, the CP should be > 0.5 or < 0.5 if vPFC activity, on average, increases or decreases, respectively, when the monkeys report that the reference and test stimuli are different.
We first calculated the “grand” CP [24
]. In this analysis, the “same” and “different” distributions were formed using the data generated from all of the potential test-stimulus morph values (i.e., 0% – 100%). The data in represent the grand-CP values generated when the reference stimulus was bad
, whereas the data in represent the grand-CP values generated when the reference stimulus was dad
. The mean grand-CP values from both distributions were reliably greater than 0.5 (t
-test; p < 0.05). This result is consistent with the hypothesis that vPFC activity during the same-different task reflects the monkeys’ choices.
Figure 3 Choice probability. The distribution of grand choice-probability values for each neuron in our population of vPFC neurons when (A) the reference stimulus is the prototype word bad and when (B) the reference stimulus is dad. Panel C shows the distribution (more ...)
This population-level result was also seen at the single-neuron level. When the reference stimulus was bad, 33 of the 56 neurons had grand-CP values reliably larger than 0.5 (permutation test, p <0.05); this proportion of neurons is reliably greater than chance (binomial test, p < 0.05). Similarly, when the reference stimulus was dad, a significant proportion of vPFC neurons (n = 26/55; binomial test, p < 0.05) had grand-CP values that were reliably larger than 0.5 (permutation test, p <0.05). We did not find a reliable population (p > 0.05) of vPFC neurons with significant CP values < 0.5.
Next, we considered whether the results of the grand-CP analysis might be biased by particular test-stimulus morph values. The CP values might have been biased toward large values during those trials when the monkeys’ choices were “easy” (i.e., those trials when the reference and test stimuli were very different or identical) and there were few error trials. In contrast, the CP values might have been biased toward values equaling 0.5 during those trials when the monkeys choices were “hard” (i.e., those trials when the reference and test stimuli were similar) and there were relatively more error trials.
To eliminate this possibility, we calculated the CP values from the neural data generated during easy trials (0%, 20%, 80%, and 100% morphs) and during hard trials (40% – 60% morphs). The population distributions of these easy- () and hard-CP () values were both reliably greater than 0.5 (t-test; p < 0.05). Thus, vPFC neurons code the monkeys’ choices for both easy and hard morph values.
Finally, we examined, at the population level, the grand-CP time course. shows this analysis when the data were aligned relative to test-stimulus onset. As expected, when the reference stimuli were presented (i.e., time < 0), the mean CP value was not reliably different than 0.5. However, following test-stimulus onset, the CP increased and became reliably > 0.5. The average CP value remained > 0.5 following test-stimulus offset for another ~250 msec before returning to a value of 0.5.
Figure 4 Time course of choice probability. A, B: The CP values were calculated from non-overlapping 200-msec epochs of neural activity. The data in A are aligned relative to the onset of the test stimulus, whereas the data in B are aligned relative to onset of (more ...)
To test how the grand CP is modulated before the monkeys report their choices, we realigned the data relative to the onset of the two LEDs. As seen in , the CP preceding LED onset remained elevated. Additionally, there was a slight increase in the CP following LED onset that correlates with the monkeys’ saccade to one of the two LEDs. This CP increase was not wholly related to any potential spatial tuning of neural activity during the saccade period: saccade-related activity in our population of vPFC neurons was not, in general, spatially tuned (data not shown).
Control analyses for task-dependent activity
If vPFC activity reflects the automatic detection of acoustically uncommon stimuli, we would expect that vPFC activity would habituate with repeated presentations of the reference stimuli as seen in stimulus-specific adaptation studies [19
]. We found that, on average, vPFC activity was not modulated by the number of reference stimuli and hence, did not habituate (bad: F
(6,383) = 1.51, p > 0.05; dad: F
(6,376) = 1.25, p > 0.05). We also tested whether a vPFC neuron’s response to the test stimulus was dependent on the number of reference stimuli. As the number of reference stimuli increases, the probability that the next stimulus is a test stimulus also increases. To test this possibility, we sorted the average test-stimulus firing rates as a function of the number of reference stimuli that preceded test-stimulus onset (see Supplemental Figure 3
). We did not find a main effect for the number of reference stimuli (bad: F
(2,1105) = 1.68, p > 0.05; dad: F
(2,1082) = 0.43, p > 0.05), but there was a main effect for the morph percentage on the test-stimulus firing rates (bad: F
(6,1105) = 29.83, p < 0.05; dad: F
(1,1082) = 8.89, p < 0.05); this latter result indicates that test-stimulus firing rates were modulated by the morph percentage as seen in . Finally, we asked whether the context in which the prototype stimuli were presented (i.e., as a reference or test stimulus) modulated vPFC activity. To test this issue, we calculated, on a neuron-by-neuron basis, an index that quantified how similarly a neuron responded to a prototype when it was the reference stimulus versus when it was the test stimulus (top rows of Supplemental Fig. 4A and 4B
). Since these index-value distributions were not reliably different than zero (t
-test; p > 0.05), vPFC neurons, on average, responded comparably to a prototype when presented as reference or test stimulus. In contrast, when the reference and test stimuli were different prototypes (see the bottom rows of Supplemental Fig. 4A and 4B
), the average index value was reliably different than zero (p < 0.05) indicating that vPFC neurons responded differently to the two prototypes.