A mechanistic account of value computation in the human brain

Abstract

To make decisions based on the value of different options, we often have to combine different sources of probabilistic evidence. For example, when shopping for strawberries on a fruit stand, one uses their color and size to infer-with some uncertainty-which strawberries taste best. Despite much progress in understanding the neural underpinnings of value-based decision making in humans, it remains unclear how the brain represents different sources of probabilistic evidence and how they are used to compute value signals needed to drive the decision. Here, we use a visual probabilistic categorization task to show that regions in ventral temporal cortex encode probabilistic evidence for different decision alternatives, while ventromedial prefrontal cortex integrates information from these regions into a value signal using a difference-based comparator operation.

Fig. 1.

Experimental task and behavioral performance. (A) Rapid-event-related fMRI design. Stimuli were presented for 1.25 s, and subjects responded with a button press after a forced delay (2–4 s). The response cue, which lasted for 750 ms, indicated the mapping between choice (face and house) and response hand (left and right). Subjects were instructed to respond after the response cue was extinguished and during a second delay period (1.5–3.5 s). Feedback (reward or not) was provided for 750 ms after the second delay. Intertrial interval (1.5–3.5 s) followed feedback. (B) The 10 images used in the task with their associated weights. On each trial the sum of the weights of the four presented images was used to manipulate reward probability. (C) Proportion of face choices as a function of logLR averaged across subjects. Trials were grouped into seven logLR bins. The sigmoidal curve is a logistic fit through the data. Individual responses depicted in Fig. S1B. (D) Average subjective weights plotted against the assigned weights. Individual estimates depicted in Fig. S1C. Error bars in C and D represent standard errors across subjects. For the most part, standard errors are smaller than the data points.

Fig. 2.

Representation of probabilistic evidence in ventral temporal cortex. (A) A region in the left PFG [(x −40, y −70, z −20), Z = 4.13, peak Montreal Neurological Institute (MNI)] correlated positively with logLR (+logLR selective voxels), whereas a region in the left PHG [(x −20, y −74, z −8), Z = 4.28, peak MNI] correlated negatively with logLR (–logLR selective voxels). Activity in corresponding voxels in the right hemisphere showed a similar pattern but ultimately failed to survive our stringent significance tests. For visualization purposes images are thresholded at Z > 2.6 and Z < −2.6, respectively (uncorrected). Images are radiological convention. (B) Event-related BOLD signal averages (Eq. 10) for five different logLR levels, from each of the two regions shown in A. Traces are aligned to the onset of visual stimulation at 0 s. The statistical contrast used to identify the regions (logLR) predetermined the shape of these plots, which are shown for illustrative purposes. Error bars represent SE across subjects.

Fig. 3.

Separate representation of WOE_F and WOE_H. Mean parameter estimates (βs) for the WOE_F and WOE_H regressors in voxels correlating positively with WOE_F (Left, WOE_F selective voxels) and WOE_H (Right, WOE_H selective voxels). These results demonstrate that activity in WOE_F and WOE_H selective voxels encoded primarily the face and house weight of evidence, respectively. Error bars represent SE across subjects.

Fig. 4.

Value signal computation in vmPFC. A region of the vmPFC covaried both with |logLR| [orange (x −6, y 50, z −2), Z = 4.07, peak MNI] and |PFG(t) − PHG(t)| [green (x −2, y 52, z −2), Z = 4.45, peak MNI], providing strong evidence that this region is involved in computing a value signal by combining the weight of evidence for face (F) and house (H) by using a difference-based comparator operation. For visualization purposes images are thresholded at Z > 2.6 (uncorrected). Images are radiological convention.