Primate prefrontal neurons signal economic risk derived from the statistics of recent reward experience

  1. Fabian Grabenhorst  Is a corresponding author
  2. Ken-Ichiro Tsutsui
  3. Shunsuke Kobayashi
  4. Wolfram Schultz
  1. University of Cambridge, United Kingdom
7 figures, 1 table and 1 additional file

Figures

Risk, choice task and basic behavior.

(A) Relationship between risk measured as reward variance and reward probability. (B) Choice task. The animal made a saccade-choice between two visual stimuli (fractals, ‘objects’) associated with …

https://doi.org/10.7554/eLife.44838.002
Neuronal coding of objective risk for objects and actions.

(A) Activity from a single DLPFC neuron in fixation period related to the true risk associated with object B, derived from reward probability. Top: peri-event time histogram of impulse rate, aligned …

https://doi.org/10.7554/eLife.44838.004
Figure 3 with 1 supplement
Deriving subjective risk from the variance of reward history.

(A) Subjective weights of influence of recent rewards on object choice, as derived from logistic regression. Filled symbols indicate significance (p<0.005, t-test; pooled across animals). (B) …

https://doi.org/10.7554/eLife.44838.006
Figure 3—figure supplement 1
Influences on saccadic reaction times.

Standardized regression coefficients from multiple regression of saccadic reaction times on different task-related variables. Data pooled across animals; reaction times were z-standardized within …

https://doi.org/10.7554/eLife.44838.007
Figure 4 with 4 supplements
Subjective object-risk coding.

(A) Activity from a single DLPFC neuron coding subjective risk associated with object A before choice (fixation period). Object risk was derived from the variance of recently experienced rewards …

https://doi.org/10.7554/eLife.44838.011
Figure 4—figure supplement 1
Anatomical location of recording sites.

Anterior-posterior position was defined with respect to inter-aural line. Orange crosses indicate locations for all recorded neurons. PS, approximate position of principal sulcus. Lower right …

https://doi.org/10.7554/eLife.44838.012
Figure 4—figure supplement 2
Reward-history control.

(A) Activity of a single DLPFC neuron reflecting the non-linear interaction between rewards in the previous two consecutive trials. The neuron showed stronger activity in the fixation period on the …

https://doi.org/10.7554/eLife.44838.014
Figure 4—figure supplement 3
Control analyses for neuronal object-risk coding.

(A) Results from supplementary analyses in which neuronal activity was regressed on object-risk measures derived using different exponential weighting functions for past rewards. Numbers next to the …

https://doi.org/10.7554/eLife.44838.016
Figure 4—figure supplement 4
Numbers of neurons (and percentages of recorded neurons) encoding risk and value for alternative risk definitions.

The alternative definitions were used for calculating the object risk regressors in our main neuronal regression model (Equation 10).

https://doi.org/10.7554/eLife.44838.017
Subjective action-risk coding.

(A) Activity from a single DLPFC neuron coding subjective risk associated with rightward saccades (action R) during fixation. Action risk was derived from the variance of recently experienced …

https://doi.org/10.7554/eLife.44838.020
Population decoding of risk from unselected neurons.

(A) Leave-one-out cross-validated decoding accuracy (% correct classification) of a linear support-vector-machine classifier decoding object risk and action risk in pre-cue period. Decoding …

https://doi.org/10.7554/eLife.44838.022
Figure 7 with 2 supplements
Prefrontal neurons dynamically code risk with other behaviorally important variables.

(A) Neuronal reward history-to-risk transition. A single DLPFC neuron with fixation-period activity that initially reflected whether reward was received from a particular choice on the last trial …

https://doi.org/10.7554/eLife.44838.024
Figure 7—figure supplement 1
Coding of risk and value jointly with spatial variables.

(A) Numbers of neurons with joint and separate coding of object risk and left-right cue position. Numbers derived from sliding window analyses (Equation 11). (B) Numbers of neurons with joint and …

https://doi.org/10.7554/eLife.44838.025
Figure 7—figure supplement 2
Utility control.

(A) A single DLPFC neuron encoding the utility associated with object A at the time of choice (cue period). Utility was defined as the weighted linear combination of object value (based on reward …

https://doi.org/10.7554/eLife.44838.027

Tables

Table 1
Comparison of different models fitted to the animals’ choices.

Best fitting model indicated in bold.

https://doi.org/10.7554/eLife.44838.010
ModelDescriptionBoth animalsAnimal AAnimal B
AICBICAICBICAICBIC
(1)Value from reward history12.24822.24901.50771.50847.35717.3636
(2)Value from reward history and risk22.24772.24921.50771.50927.35227.3653
(3)Value from choice history32.16142.16221.49001.49076.50436.5109
(4)Value from choice history and risk2.03852.04001.40231.40377.35287.3660
(5)Value from reward and choice history42.00892.00971.39141.39226.08806.0945
(6)Value from reward and choice history and risk2.00732.00881.38991.39146.07476.0878
(7)Objective reward probabilities52.12132.12201.46151.46226.49726.5037
(8)Objective reward probabilities and objective risk62.12102.12251.46161.46316.49826.5114
(9)Reinforcement learning (RL) model72.07632.07791.43761.43916.21616.2293
(10)RL learning, stack parameter (Huh et al., 2009)82.08102.08261.43741.43896.31986.3330
(11)RL, reversal-learning variant92.26142.26301.53301.53447.28087.2939
  1. 1:Value defined according to Equation 6; 2: Risk defined according to Equation 8; 3: Value defined as sum of weighted choice history derived from Equation 5; 4: Value defined according to Equation 7; 5: Objective reward probabilities defined according to Equation 1; 6: Objective reward risk defined according to Equation 2; 7: Standard Rescorla-Wagner RL model updating value of chosen option based on last outcome; 8: Modified RL model incorporating choice-dependency; 9: Modified RL model updating value of chosen and unchosen option based on last outcome.

Additional files

Download links