Specific frontal neural dynamics contribute to decisions to check

Abstract

Curiosity and information seeking potently shapes our behaviour and are thought to rely on the frontal cortex. Yet, the frontal regions and neural dynamics that control the drive to check for information remain unknown. Here we trained monkeys in a task where they had the opportunity to gain information about the potential delivery of a large bonus reward or continue with a default instructed decision task. Single-unit recordings in behaving monkeys reveal that decisions to check for additional information first engage midcingulate cortex and then lateral prefrontal cortex. The opposite is true for instructed decisions. Importantly, deciding to check engages neurons also involved in performance monitoring. Further, specific midcingulate activity could be discerned several trials before the monkeys actually choose to check the environment. Our data show that deciding to seek information on the current state of the environment is characterized by specific dynamics of neural activity within the prefrontal cortex.

Figure 1. Experimental protocol and behaviour.

(a) In each trial, monkeys are allowed to freely decide either to Work in a main task (where they are asked to perform a Cued decision) or to Check the gauge state that increases with the number of correct Cued decisions. The speed of increase is based on the total number of correct trials required (four possible speed; see Methods). A check leads to the onset of the current gauge stimulus (green disk with circle). A bonus reward is obtained when checking occurs, while the gauge is full. The gauge is reset after delivery of the bonus reward. (b) Monkeys increase check frequency with increasing gauge size (bottom), while keeping stable performance during Cued decisions in the main task (top). Dots are actual data (average over sessions), whereas lines are logistic fits. (c) Logistic regressions for each monkey testing the contribution of gauge size, performance in the previous trial (Previous; Incorrect and Check compared with correct) and speed of gauge increase (Speed; compared with speed 1) to the probability of checking. The estimated coefficients notably show a reduced probability to check (negative value) after incorrect trials in comparison with after-correct trials during the Cued decision. (d) MRI-based reconstructions highlighting the approximate recording locations in LPFC (blue) and MCC (black) from the genu of the arcuate sulcus (ArcGen).

Figure 2. Effects of checking.

(a) Evolution of average reward rate during a block. Solid lines show for each monkey the average amount of reward per minute obtained during a block (condition: ‘With check'). Each block was divided in ten successive time bins covering several trials. The evolution is expressed in per cent of total time in a block. The single point at 100% corresponds to the last decile that includes the bonus reward. Data can be compared with a hypothetical but ideal case where the animal does not check by excluding the time spent in checking the gauge size (dashed lines, condition ‘Without check'). The number in parenthesis is the difference in cumulated reward rates between the two conditions (‘With check' versus ‘Without check'). It reveals an overall positive gain for all monkeys. (b) Effect of observed gauge size on checking. The plots show for each monkey the number of trials (distance) that the animal takes before a check n+1 after having seen a particular gauge size (x axis) at check n. The larger the gauge size seen at check n, the smaller the number of trials (distance to check) before the next check n+1. The effect is independent of the number of trials performed in the main task in the current block (symbols), that is, independent of time. This suggests that monkeys properly used information gathered from the gauge size to regulate checking. Correlations displayed on graph were made on the three sets of data.

Figure 3. Typical examples of single unit activity.

Average firing rates and rasters for six different single unit activities in MCC (top) and LPFC (bottom) discriminating either between negative/positive feedback (left), Check versus Work decisions (middle) and Cued decisions in the main task (right). Black lines on time axis highlight the time period where the difference in firing rate is significant (KW test, P<0.01). Only subsets of trials are displayed (n represents the total number of trials).

Figure 4. Proportions of neurons encoding feedback and decisions.

(a) Venn representations of the overall number of neurons discriminating significantly negative/positive feedback, Check versus Work decisions and/or Cued decisions in the main task during the time epochs indicated in orange on top of graphs in b. Stars indicate significant differences between MCC and LPFC. (b) Time resolved proportion of cells extracted from the sliding glm (see ‘Sliding generalized linear models' in Methods) with a significant discrimination of feedback, Check versus Work and Cued decisions. Data presented during the time period between feedback (FB) and lever onset, and aligned on the end of trial signal (EoT, grey bars on the x axis represent the duration of the EoT) for Feedback and Check versus Work and on target onset for Cued decisions. Dashed grey lines represent the 5% level.

Figure 5. Linear population coding.

Linear decoding (a) and cross-temporal decoding (b) reveal a double dissociation between MCC contributing earlier and more reliably to Feedback processing and Check versus Work decisions, and LPFC contributing earlier and more reliably to Cued decisions during the main task. Arrows in a indicate first significant decoding (bold) for each area (colour). Statistical threshold was set at P<0.05 using permutation testing. Significant decoding is depicted by bold lines in a. Cross-temporal decoding in b were thresholded (grey colourmap) depending on the smallest significant decoding rate at P<0.05. Statistical comparisons of MCC and LPFC population coding were tested using KW tests. Significant differences (at P<0.05, Bonferroni corrected) are displayed in red on the time axis in a and within a red contour in b. The red line is drawn at the level of the structure with better decoding performance.

Figure 6. Bias for Check and dynamic of firing rate.

(a) Both MCC and LPFC show more cases with higher activity in Check compared with Work at decision time. The population histograms show the individual cell Z-values from the sliding glm for all time bins for all cells (see Fig. 4). (b) Average population activity shows for all cases greater dynamical changes for decision to Check (plain lines) than for decisions to Work (dashed). PSTH are displayed for subpopulations of cells that had higher activity in Work than in Check (left), and that had higher activity in Check than in work (right). Data for each decision (Check: lines; Work: dashed lines) and each structure (MCC: black, LPFC: blue). (c) Normalized variance of activity across bins is higher for Check than Work in both structures. Stars indicate significance at P<10⁻³, analysis of variance on log values (Check versus Work decision × Check versus Work preference).

Figure 7. Feedback and Check versus Work specificities at the cell group level.

Estimates (β-coefficients) obtained from the MCC and LPFC population, and time-resolved glm for Check versus Work (green) and Feedback (blue) variables (see ‘Group analyses using glmm' in Methods). Significant effects are indicated by a black triangle (P<0.05). Positive values depict a population activity bias towards Check decisions (green line) or negative feedback (blue line), whereas the opposite represent a bias towards Work decisions or Positive feedback.

Figure 8. Encoding of Gauge size and Distance to check.

(a) Estimates (± CI) for the variable Gauge extracted from the time-resolved group glmm (as in Fig. 7). Significant effect of gauge size is indicated using triangles on data for each structure (P<0.05). The grey bar on the x axis represents the duration of the EoT signal. (b) Average per cent of correct decoding between trials at short distances to Check and trials at long distances using inter-trial firing rates (mean±s.d.). The circle size represents the number of neurons used by the decoder (see Methods). Dotted lines represent the chance level calculated from permutation testing and statistical differences are displayed using asterisks for each structure separately (*P<0.05 and ***P<0.001). (c) Venn representations of the number of neurons contributing to the decoding (including neurons with non-zeros coefficients in >10% of trial permutations; see ‘Population analyses' in Methods) of trials n, n−1 and n−2 against those at longer distance to check (as in b). It is noteworthy that in MCC the subpopulation of neurons contributing to decoding both n−1 and n−2 was larger than the one contributing also to decoding trial n (χ²-tests, P<0.05).