Multiple brain networks contribute to the acquisition of bias in perceptual decision-making

Abstract

Bias occurs in perceptual decisions when the reward associated with a particular response dominates the sensory evidence in support of a choice. However, it remains unclear how this bias is acquired and once acquired, how it influences perceptual decision processes in the brain. We addressed these questions using model-based neuroimaging in a motion discrimination paradigm where contextual cues suggested which one of two options would receive higher rewards on each trial. We found that participants gradually learned to choose the higher-rewarded option in each context when making a perceptual decision. The amount of bias on each trial was fit well by a reinforcement-learning model that estimated the subjective value of each option within the current context. The brain mechanisms underlying this bias acquisition process were similar to those observed in reward-based decision tasks: prediction errors correlated with the fMRI signals in ventral striatum, dlPFC, and parietal cortex, whereas the amount of acquired bias correlated with activity in ventromedial prefrontal (vmPFC), dorsolateral frontal (dlPFC), and parietal cortices. Moreover, psychophysiological interaction analysis revealed that as bias increased, functional connectivity increased within multiple brain networks (dlPFC-vmPFC-visual, vmPFC-motor, and parietal-anterior-cingulate), suggesting that multiple mechanisms contribute to bias in perceptual decisions through integration of value processing with action, sensory, and control systems. These provide a novel link between the neural mechanisms underlying perceptual and economic decision-making.

Keywords: decision-making; fMRI; motion discrimination; reinforcement learning; reward.

Figure 1

Experimental paradigm. Each trial is composed of a context, motion stimulus (illustrated as white dots), and reward points. The number of reward points that one can earn depends on the context and the choice. P, probability; Δt, duration; AVG, average; s, second.

Figure 2

Behavioral results. (A) Psychometric function. The motion strength was plotted against the probability of choosing “up” in each context across the five runs (dots). This change is modeled with logit function in which its intercept reflects the value difference between the two motion directions that has been learned up to the end of each run (solid lines). The dashed lines are the indecision points. Motion strength: the percentage of coherent moving-dots; ±: upward/downward motion. Error bars: ±1 s.e.m. (B) The bias acquisition process. The trial numbers are plotted against the indecision points estimated by the reinforcement-learning model using individuals' data. Solid lines: group mean. Shaded areas: ±1 s.e.m. Dashed lines: the end of each run. Colors: corresponding to the reward context as illustrated in (A). (C) The autocorrelation functions. The correlation estimates using the residuals from the context-based learning model is plotted against each lag. Solid lines: group average. Dashed lines: 95% confidence interval of the autocorrelation estimated from a random series with the same number of trials (a total of 280 trials).

Figure 3

The acquired bias in the bran. The maps show the brain areas whose activation positively correlates with the amount of acquired bias on each trial. No brain areas negatively correlate with this signal after the whole-brain correction of multiple comparisons. All maps are presented at p < 0.05 whole-brain corrected using cluster-based Gaussian random field and overlaying on the mean anatomical images from the group of participants. R, right hemisphere; L, left hemisphere; Z, the MNI coordinate of the axial slice.

Figure 4

The reward prediction errors. The brain maps show the regions whose activation positively correlates with the two types of reward prediction error signals in perceptual decisions after adjusting one type over the other. All maps are presented at p < 0.05 whole-brain corrected using cluster-based Gaussian random field and overlaying on the mean anatomical images from the group of participants. Red-Yellow, context-based RPE; Blue-Light-blue, context-free RPE; R, right hemisphere; L, left hemisphere; Z, the MNI coordinate of the axial slice.

Figure 5

Three functional connectivity patterns underlying the growth of bias. The left panel shows the seed regions that were used in the psychophysiological interaction analyses. These seed regions were selected according to the literature and centered at the MNI coordinates (vmPFC: [−6, 39, −8]; left-frontal: [−45, 21, 0]; left-parietal: [−36, −39, 45]) with the radius of 10 mm. The brain maps on the right panel show the areas that positive correlate with the interaction between each of the seed regions and the amount of acquired bias on each trial. The statistical maps are corrected for multiple comparisons at the whole-brain level using cluster-based Gaussian random field correction at P < 0.05 and overlaying on the mean anatomical images from the group of participants. R, right hemisphere; L, left hemisphere; X, Y, Z, the MNI coordinate of the brain slice.